Radar Operation Mode Recognition via Multifeature Residual-and-Shrinkage ConvNet

Radar operation mode recognition holds an increasingly critical place in electronic countermeasure as well as in remote sensing. However, the overlapped waveform parameters pose huge challenges to performing the radar operation mode recognition task in severe electromagnetic environments, particularly with large measurement errors or small sample lengths. By analyzing the timing patterns of a single radar pulse parameter and the correlation characteristics of multiple radar pulse parameters, this article first provides a revolutionary representation of the radar operating state by integrating interpulse parameter characteristics. Subsequently, a multifeature fused stream-level recognition framework with an attention mechanism, named residual-and-shrinkage ConvNet, is proposed to identify typical radar operating states. This tailored-made deep learning framework can effectively extract the timing and correlative features, which are conducive to pattern classification. The results of numerical experiments suggest that the proposed approach affords superior performance for the operation mode recognition task, even when the measurement error is large and the sample length is small, signifying the proposed method is strongly robust and time-efficient.


I. INTRODUCTION
T HE modern battlefield environment has become increasingly complicated [1]. To deal with the diversity and complexity of aerial missions, the airborne radar needs to develop appropriate operating modes [2], [3], [4], [5], including but not limited to single target tracking (STT), multiple target tracking (MTT), track while scan (TWS), track and search (TAS), range while search (RWS), velocity search (VS), and so on [6]. Due to the different signal patterns, each mode can be further subdivided into different submodes, which are applied to various tactical strategies and marked with different threat degrees [7]. In an air combat, determining the operating mode of the adversary's radar is a critical segment for threat level assessment and a prerequisite Manuscript  for precise jamming [8]. Meanwhile, operation mode recognition is actually the state classification problem, which is a core technology for target status monitoring in microwave remote sensing applications. Consequently, accurate identification of radar operating modes is considered significant in the field of electronic countermeasures (ECM) and remote sensing [9], [10]. The parameters that are commonly employed to distinguish radar operating states consist of pulse repetition interval (PRI), radio frequency (RF), pulsewidth (PW), etc. [11]. However, random combinations of the above parameters are exploited by advanced radars for search and tracking, resulting in a single operation mode with multiple different combinations of these parameters, or overlapping parameter intervals from one mode to another. This poses new challenges for radar operation mode recognition.
Wang and Guan employed Dempster-Shafer (D-S) evidence theory based on the regular pulse parameter characteristics for the identification of STT, TWS, and other regular operation modes of the airborne fire control radar (AFCR). But the scheme is less applicable to the phased array system [12], [13]. Yang et al. [14] intensively considered the pulse amplitude (PA) distribution characteristics of the intercepted radar signals and designed a rearranged differential algorithm to identify tracking signals. This method can accurately determine the radar operation mode with a high threat level, such as STT and TAS, but it does not work well in the identification between the two modes with both search attributes, like TWS and RWS. A single feature cannot completely portray all the characteristics of the operating mode, so Wang et al. [15] built a clustering recognition structure using multidimensional features from a semisupervised perspective. Although such a clustering analysis method is capable of real-time processing, the classification boundary is weakly defined. Up to now, machine learning algorithms have already been broadly adopted in the field of pattern recognition for radar [16], [17], [18]. Some scholars studied the application of support vector machine (SVM) in radar operation mode recognition [19], [20]. And more lately, Hui et al. [21] successfully applied deep learning technology to this field, enabling the recognition of several operating modes of airborne phased array radar equipped with dissimilar parameters. Unluckily, the network used was rather simple and it was difficult to classify more operation modes with various waveform parameter combinations. Most of the algorithms mentioned above are built on a single feature parameter or a single time-of-moment, which fails to reflect the This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ temporal association of the pulse sequence. Although these algorithms have improved the classification accuracy rate somehow, they are restricted in cases of high signal similarity and severe parameter overlap.
To achieve a favorable result for radar operation mode classification, it is necessary to acquire the variation regularities of signal parameters from the radar pulse sequences. Jia et al. [22] analyzed the feasibility of airborne radar operating mode identification based on coherent processing interval (CPI) features, but no specific recognition scheme was put forward. By applying the syntax rule to shape the multifunctional radar signal, Haykin and Visnevski effectively divided the radar pulse sequence into the "word-phrase" hierarchy [23]. As a result, different operation modes correspond to different radar phrases, which efficiently represent the temporal variation of the radar signal. In spite of that, a mathematical representation of radar operation mode can be made using radar words over a period of time with the help of syntactic rules. However, the current algorithms neglect to derive such long-time information, which in turn leads to a failure to extract the temporal regularity of the parameters.
In light of the above analysis, a great deal of research has focused on the construction of classifiers, but undermining the fundamental nature of raw radar signal, that is the sequential knowledge of radar pulse train. In addition, the problem of identifying operation modes that overlap in parameter intervals requires further exploration. Indeed, little research has been conducted under conditions of noise perturbation and missing sample length.
To enable the aforementioned problems to be addressed, a stream-level multifeature fused residual-and-shrinkage Con-vNet (RSCN) is put forward to classify radar operating states whose operating parameters overlap with each other. Specifically, on the one hand, we develop a comprehensive representation of radar operation modes by highlighting the temporal patterns between pulse sequences as well as integrating the correlation characteristics of signal parameters. The effective extraction of timing patterns contributes to improving the accuracy of operating mode recognition when the sample length is short, thus raising the real-time performance of the algorithm. On the other hand, we establish an innovative deep learning framework with the stream-level structure, which can extract more in-detail and sophisticated classification information in an adaptive manner. In particular, we insert variable numbers of residual-and-shrinkage (RS) modules into the convolutional network (ConvNet) backbone. The RS module has the ability to optionally concentrate on the input data, adaptively learn favorable features and automatically reduce those features associated with noise. Through the above innovative work, the accurate identification of multiple operation modes is ultimately realized, even under different measurement errors and different sampling lengths.
The rest of this article is organized as follows. Section II briefly analyzes some typical operation modes of airborne radar and summaries some available features for operation mode recognition. Section III proposes the recognition scheme based on the RSCN. Section IV provides simulation experiments and analysis. Finally, Section V concludes this article.

II. TYPICAL OPERATION MODE ANALYSIS
The state characteristics are the key elements to distinguish the radar operating states. In order to fully explore the characteristics of radar electromagnetic data for rapid and accurate recognition of radar state, this article first analyzes the functional characteristics and data characteristics of six common operating states, namely STT, MTT, TWS, TAS, RWS, and VS, and summarizes the key features conducive to radar state recognition [24], [25], [26], [27].
A. Typical Operation Mode 1) Single Target Tracking: STT refers to the operating state in which the radar locks a single target for continuous tracking. This state has high measurement accuracy and data rate, the radar antenna always keeps pointing at the target, and the target is tracked and filtered in real time, so that accurate tracking data of the target can be obtained.
2) Multiple Target Tracking: MTT devotes all radar resources to a continuous high-precision tracking task for multiple targets, without searching the rest of the airspace, so the tracking data rate is high.
3) Track While Scan: TWS can not only complete the search task by transmitting the search beam to the specified airspace, but also can correlate the information obtained from a single airspace search to maintain the low data rate tracking of the specific target.
4) Track and Search: TAS utilizes the principle of phased array antenna beam agility and time splitting, so that tracking and searching alternate at different data rates, with higher tracking data rates. 5) Range While Search: RWS provides a rough indication of the target's location, azimuth and elevation, but does not correlate point traces. This mode allows for multitarget detection, but the measurement accuracy is low and cannot be used directly to control missile attacks, therefore the threat level is low.
6) Velocity Search: VS is mainly used for the detection of head-on (down-looking) targets at large distances and can provide the target's parameters, like velocity, azimuth, and elevation. It has a low data rate and a low threat level. High pulse repetition frequency (HPRF) waveform is a typical characteristic of this mode.

B. Available Feature Summary
Through the above analysis, the key information for radar state recognition includes the following: 1) The PA variation patterns are different. STT amplitude remains basically stable. MTT amplitude will show multiple distributions. TWS amplitude shows "Sinc" shape. TAS amplitude contains both "Sinc" shape and track pulses. 2) The recall period is an important feature that distinguishes track from scan. In track states, the radar beam keeps pointing at the target, hence without a tracking recall period. In composite states, the radar searches the airspace while simultaneously tracking the target with a certain period of revisit.
3) The correlation characteristics (or signal patterns) consisting of PRI, RF, and PW are different. For instance, in the STT mode, the PW is smaller and the PRF is higher (mainly using medium pulse repetition frequency (MPRF) and HPRF waveforms), the PRF does not switch frequently and the signal is consistent without interruption.

III. PROPOSED METHOD
Modern radars usually apply the same group of waveform parameters to different operating states, resulting in overlapping parameter intervals between different operation modes, which results in reduced pattern recognition capability. To raise the recognition accuracy in the case of such overlapping parameter ranges, a multifeature merged RSCN method for radar operation mode recognition is proposed based on the long-term interparameter correlation information. The proposed approach takes into account the feature variations of the pulse sequence and combines multiple characteristics together, thus a comprehensive representation of radar operation mode can be finally obtained. The proposed stream-level deep learning framework is engineered with cognitive capability, more specifically, it can automatically perceive and pick up features that are associated with the task, while granting less attention to the less important features. The above innovative work is beneficial to improve the recognition accuracy rate of radar operation modes.
The main processing steps of the proposed method are illustrated in Fig. 1. First, by analyzing the differences in the waveform of each operation mode, we now choose the distinguished parameters that will help to separate various operating states, regardless of whether they are in the temporal, frequency, or spatial domain. Second, the radar pulse sequences are modified to radar phrases in accordance with radar syntax rules [28]. Later on, the radar phrases corresponding to each parameter are merged together as input to the recognition network RSCN. Ultimately, the RSCN gives an output of operation mode recognition. In particular, by means of the built-in attention mechanism, the proposed method makes it possible to automatically retrieve the critical characteristics, including the temporal information of a single parameter as well as the correlation information between multiparameters.

A. Feature Construction
According to tactical performance and metric demands, the radar selects and designs its own operating parameters, of which pulse description word (PDW) data is a digital description of the radar pulse. Due to the large number of mission requirements in the modern battlefield environment, the parameters of various operation modes usually overlap, so it is not reliable to distinguish between the radar working modes by the instantaneous radar word. For the actual radar system, in order to obtain a  higher coherent processing gain during the echo signal processing, a certain number of filling pulses and coherent pulse trains are transmitted during the beam dwell period to design the waveform of each operation mode. The PDW parameters of these pulses are basically the same as the PDW parameters of the initial pulse, that is, a coherent processing interval (CPI) is formed.
Consequently, we may characterize a radar pulse sequence for a single operation mode with a number of sequential CPIs. This description method can preserve the timing law of the radar signal. Radar words are arranged to form radar phrases according to syntactic rules, so that the modeling of a radar pulse sequence is completed with a "word-phrase" structure, as vividly depicted in Fig. 2, where T1 and T2 schematically show the arrival time of pulses. Namely, each operation mode corresponds to at least one radar phrase (RP) that reflects the timing variation of the radar signal. Since the PDW parameters of one CPI stay comparatively static, all variables in a radar word (RW) are expressed as the average values. For instance, the pulsewidth in a radar word is the average value of the pulsewidth in one CPI. RW is denoted as RW = [P RI, RF, P W, P A] T .
The assumption is that a sequential CPI of operation mode i consists of L groups of CPIs so we denote the RP as RP i = [RW 1 , RW 2 , . . . , RW L ], which would be used to describe the operation mode i. To state the RP as a matrix where P W i1 stands for the mean value of PW in the 1 th CPI of operation mode i. Likewise, P W iL is the mean value of PW in the L th CPI of operation mode i. Equation (1) represents the general signal model since it describes the signal pattern of the operation mode.
In general, a certain number of radar pulses need to be accumulated so as to achieve great detection performance. As a result, a CPI often consists of several hundred pulses, and this value can be adjusted according to radar velocity measurement accuracy in practice. A larger CPI, i.e., a greater number of integrated pulses, means a higher velocity resolution, and vice versa [5]. In addition, the radar often uses the medium pulse repetition frequency (MPRF) waveform to resolve the ranging ambiguity, in the meanwhile, the MPRF is usually used with several PRF values together. Hence, it would take several thousand pulses to record all information about a specific operation mode, which may well not meet the real-time requirement of particular applications. In such cases, we can cut down the dimensionality of the feature matrix by employing certain preprocessing tools.

B. Residual-and-Shrinkage ConvNet
To classify these operating states whose parameters overlap with each other, a deep learning framework is employed to capture the long-term sequential patterns of a single parameter as well as the correlation features between multiparameters. With reference to previous experience, the ConvNet could receive a global characterization of the input through learning from numerous training samples, so that the robust high-level characteristics [29] can be extracted from the input data layer by layer. To yield the importance of each feature map automatically, we embed the RS module onto the backbone network ConvNet. The adopted RS module consists of a specialized miniaturized network [30] that removes task-irrelevant characteristics by adaptively adjusting the threshold. Ultimately, by integrating deep semantic characteristics with shallow appearance ones, the accurate and detailed classification of radar operation modes with overlapped parameter values can be accomplished, even in the context of large measurement errors or small sample lengths.
1) ConvNet: As illustrated in Fig. 3, the ConvNet consists of five fundamental parts, including a convolutional block composed of three convolutional layers, a batch normalization (BN),  an activation function named rectified linear unit (ReLU), a global average pooling (GAP) layer, and a multiclassifier with Softmax. In addition, the strides of all the convolution layer are s = 1, the filter numbers in each convolution layer is n 1 , n 2 , n 3 , and the convolution kernel sizes are k 1 × k 2 , k 3 × k 4 , k 5 × k 6 .
2) RS Module: As Fig. 4 depicts, the RS module consists of three fundamental parts, including a residual block, an attention mechanism, and the soft thresholding operator. The residual block consists of BN, ReLU, convolutional layer, and all of them are in pairs, along with an identity shortcut. The structural parameters are set as follows. The filter number of the convolution layers are n 4 , n 5 , the convolution kernel sizes are k 7 × k 8 , k 9 × k 10 , and the strides are s 1 , s 2 .
The problem of gradient vanishing due to the deepened network layers can be well solved by the residual block, breaking the traditional bottleneck of a deeper model but stagnant performance growth. As shown in Fig. 5, the threshold can be adaptively adjusted by the attention mechanism in the RS module. Specifically, the absolute value (ABS) of the feature map X is turned into a one-dimensional (1-D) vector via GAP. The 1-D vector is then fed into a fully connected (FC) network made up of a BN, a ReLU and two FCs for obtaining the coefficient γ. Ultimately, γ is projected by the sigmoid function into the interval from 0 to1 where z stands for the FC network's output. The procedure of obtaining the coefficient γ we called "Scaling" operation. Further, the threshold τ will be obtained by multiplying γ with the mean of |X|, which can be written as where |X| is the input X's ABS, and |X| is |X|'s mean value. The benefit of the adaptive coefficient γ is that it allows the RS module to acquire helpful messages from outside the threshold region efficiently, rather than retaining them completely.
In addition, soft thresholding [30], a popular denoising procedure, transforms near-zero characteristics to zero and retains only those features that are either quite positive or negative. The Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. operation of soft thresholding may be denoted as where x is the input of the soft thresholding and y is its output. We can find from (4) that the gradient of this function can only be 0 or 1, and this may avoid the extinction and explosion of the gradient.

3) RSCN:
We have described each module above, and then, to enable the network to fit the threshold self-adaptively and to extract the highly discriminative characteristics automatically, this article interpolates certain number of RS modules following the second convolutional layer of the ConvNet. In particular, the convolutional layer allows for a significant reduction in the number of trainable parameters, thereby reducing the likelihood of overfitting and enhancing the accuracy rate on the test dataset. In addition, the features can be normalized into a fixed distribution by BN and subsequently tuned to a preferred distribution by continuous training and learning. The overall architecture of RSCN is shown in Fig. 6.

IV. EXPERIMENTS AND ANALYSIS
This section first gives the implementation details of the experiments. Next, the experiments are divided into two aspects. For one thing, the performance of the proposed algorithm for radar operation mode recognition has been analyzed in Part B. For another, some comparative experiments have been conducted in Part C, specifically, where the influence of sample length and measurement error on the recognition performance have been discussed.

A. Implementation Details
The advanced airborne radar can deal with various missions due to the flexible operation modes and variable operating waveforms [31], [32]. With reference to the waveform patterns of typical operation modes [5], [6], [33], the parameters of STT, MTT, TWS, TAS, RWS, and VS are simulated in this paper. The parameter settings of each operation mode are listed in Table I. From a statistical perspective, the parameter intervals overlap between different operation modes. The typical waveform patterns of each operation mode are shown in Fig. 7. It should be noted    that Fig. 7 only presents some of the waveform parameters, and we also considered many other parameter combinations, where the numerical values and the variation types (e.g., constant, stagger, dwell and switch, etc.) of the parameters are various [6]. The training dataset include 30 400 samples, while the test dataset and validation dataset include 3800 samples.

B. Recognition Performance Analysis
To figure out the contributions of each feature on the recognition performance improvement, this subsection analyzes the impacts of different feature dimensions on the performance of radar operation mode recognition.
Partial features chosen from the four features are selected and combined as input to the neural network. Fig. 8 shows the training curves for different feature inputs. It can be seen that joining PA to training can rapidly improve recognition accuracy, due to the fact that the six operation modes can be easily distinguished by the scanning features. For example, STT continuously tracks a single target, so the amplitude of the received signal approximates a continuous straight line. MTT tracks multiple targets at the same time, so the amplitude shows a distribution of multiple straight lines. The number of targets tracked by the radar can be identified from the PA characteristics. TWS, RWS, and VS all provide a searching function and their PA would contain the characteristic of the scanning envelope, showing a shape of "Sinc." RWS and VS have lower data rates and take longer to scan the same airspace. TAS denotes search and track, so its PA contains both short straight lines and "Sinc." Fig. 9. Influence of different features on average recognition accuracy. ("Single," "Double," and "Triple" mean the average test accuracy of individual feature "PRI or RF or PW," double features "PRI+RF or PRI+PW or RF+PW" and correlative features "PRI+RF+PW," respectively.). Therefore, as long as no large noise or interference is involved in the environment, PA can be reliably used to improve the recognition accuracy.
However, the presence of large amounts of noise and interference in the real electromagnetic environment, as well as the relative location of the receiver, results in the PA not providing useful information, and it is then necessary to combine other features, such as PRI, PW, and RF, to form the other correlated features. Fig. 9 shows the influence of different features on average recognition accuracy. It can be seen that the best results are obtained by adding all features to the training. Specifically, the PA feature indeed has a higher discriminability under ideal conditions. But in actual electromagnetic environments, the detection of PA is greatly affected by factors such as maneuvering and atmospheric disturbances, while the estimations of PRI, RF, and PW are relatively correct. At this time, the recognition performance using only the PA feature significantly decreases, while the recognition performance using correlated features fluctuates less. The reasonable consideration of the real environment is the reason why we introduce these features into the training.
In addition, to investigate the recognition performance of the algorithm when there are estimation errors in a single feature parameter, we considered six different levels of estimation errors, namely 1%, 3%, 5%, 7%, 9%, and 11%. The experimental result is depicted in Fig. 10. It can be seen that the incorrect estimation of a parameter has a small impact on the recognition performance of our algorithm, with only about 1%-2% accuracy loss, which to some extent demonstrates the robustness of the algorithm. This is because the algorithm makes use of the correlative features of multidimensional parameters, so that when the parameter in one dimension is wrong, the correlative features of other parameters can be used to improve the accuracy of the operation mode recognition. Thus, the proposed algorithm is proven to have the ability to deal with such scenarios where some parameters are incorrectly estimated.

C. Comparative Experiments on Recognition Performance
To verify the feasibility and robustness of our proposed method, different methods of operation mode recognition are simultaneously implemented and compared with each other, including the ConvNet [29], the squeeze-and-excitation convolutional network (SECN) [34], and our proposed method. The test dataset is randomly divided into four equal groups, with 950 samples in each group. Fig. 11 shows the training course of each method. The results imply that our proposed method converges faster and ends up with the highest training or validation accuracy and the lowest loss. Finally, the recognition results of the four test groups are illustrated in Fig. 12, and the average test accuracy is shown in Table II. The results verify that better classification performance is acquired by our proposed RSCN over ConvNet or SECN. The reason is that the attention mechanism of the RS module is able to automatically assign more weight to critical characteristics and remove the information related to noise. On the other hand, some slight misclassifications are made by the squeeze-and-excitation (SE) block because those task-irrelevant characteristics are not entirely suppressed.
As can be seen from Table II, the recognition accuracy grows in proportion to the number of RS modules. The reason for this is that RS modules remove noise-related characteristics. In the meanwhile, the cascade structure of RS modules has not only increased the network layer, but has also further refined the taskrelevant characteristics, which in turn reduces the noise-relevant characteristics. In this task, the peak recognition performance is reached using two RS modules. Keep increasing the quantity of the RS module would need more training time due to the deepened layers and parameters of the network. Therefore, our methods in subsequent experiments are set up with two RS modules.
In the following, to explore the robustness of the proposed algorithm, this section will analyze the influences of sample length and measurement error on the performance of radar operation mode recognition.
1) Influence of Sample Length: Different application scenarios have different requirements for recognition accuracy and timeliness. To investigate the effect of sample length on the recognition accuracy of the algorithm, the simulation results are obtained by setting the sample length to 100, 500, 1000, 1500, and 2000, respectively, as shown in Fig. 13. It can be observed that as the sampling length becomes small, the recognition performance of all methods decreases. The recognition performance of the ConvNet deteriorates rapidly at smaller sample lengths, while the recognition performance of the SECN shows a relatively gentle change. However, the recognition performance of the proposed RSCN algorithm outperforms the comparative methods at any sample length. For our proposed algorithm, when the sampling length is less than 1000, the recognition accuracy of the radar operating state shows an increasing trend. After the sampling length is larger than 1500, the recognition accuracy is stable at 100%.
Moreover, the confusion matrices in Fig. 14 show that some samples of TWS and TAS are incorrectly identified when the sample length is less than 1000. Regardless, the above experimental results demonstrate the proposed RSCN performs well in the case of short time windows and small sampling lengths. We conducted 10 000 Monte Carlo simulations for each length to test the average time for identifying the radar operating mode, and the results are recorded in Table III. According to Fig. 13, the recognition performance of our proposed algorithm is almost the same at lengths of 1500 and 2000. Thus, when the number of received pulses reaches 1500, the operation mode recognition can start, in which case a high recognition accuracy can be achieved with less time. Furthermore, the testing accuracy can reach above 98% at the length of 500 while the time is    and can meet the requirements of rapid recognition of operating states.
2) Influence of Measurement Error: In addition, in real reconnaissance environments, the measurement of radar signal parameters suffers from certain random errors due to noise and other reasons. To verify the robustness of the proposed RSCN, the Gaussian disturbances are added to the PRI, RF, PW, and PA sequences to simulate the measurement tolerances. The relative error is set to 1%, 3%, 5%, 7%, 9%, and 11%, respectively. The sample length is 2000. Finally, the recognition results are obtained as shown in Fig. 15. It can be seen that the performance of our algorithm is more robust to noise and jamming than comparative methods. This is because the comparative methods mainly learn the amplitude envelope information of the received pulses. However, the amplitude envelope is greatly affected by  noise and jamming, which usually makes it unable to provide valuable classification information. In contrast, our proposed algorithm simultaneously extracts both the amplitude envelope information and the correlative features of all parameters to improve the recognition accuracy in complex electromagnetic environments.
The confusion matrices in Fig. 16 show that TWS and TAS are easily confused in the presence of measurement errors. Because the noisy clusters added to the PA are more similar to the tracking pulses of TAS, leading to the noise is incorrectly identified as the tracking pulses thus causing confusion between TWS and TAS. In addition, the recognition accuracy of the operation mode decreases as the measurement error increases. Nevertheless, when the measurement error is up to 11%, the recognition accuracy can still reach above 97.8%, which indicates that our proposed RSCN is highly robust for use in airborne radar operation recognition.
All of the experimental results show that our proposed RSCN carries out excellent identification of radar operation modes in the presence of overlapped parameter intervals and complicated electromagnetic environments.

V. DISCUSSION
This section first analyzes the performance and influence factors of the proposed algorithm. Next, the cases of recognition errors and the corresponding reasons are analyzed. Finally, the limitations of the proposed algorithm are listed and future work is outlined.
This article proposes a method for recognizing the operation modes of airborne radar based on multifeature residual-andshrinkage ConvNet. The experimental results verify the high recognition performance of the proposed method. It should be noted that the recognition performance of this method is influenced by the sample length and measurement error. First, the sample length refers to the number of received pulses, which directly determines the accuracy of the obtained information about the operation mode recognition of airborne radar. Therefore, a shorter sample length could lead to incomplete changes in the PDW (PRI, RF, PW, PA, etc.) parameter variation, while a longer sample length would increase the processing time. Therefore, choosing an appropriate sampling length is crucial, and this length is usually related to the antenna scan period of the radar. The accuracy of PDW has a direct effect on the performance of radar operation recognition. However, in real scenarios, factors such as noise and interference cause measurement errors in radar parameters, especially in PA. Therefore, measurement error is also an important aspect affecting recognition performance. As the measurement error increases, the performance of operation mode recognition decreases. However, the proposed algorithm not only extracts amplitude envelope information, but also extracts correlative features of the parameters. Thus, the proposed algorithm performs well even when measurement errors are large.
In addition, under certain special circumstances, some operation modes may be easily confused, leading to a decrease in recognition accuracy. Specifically, on the one hand, when the measurement error is large, TAS and TWS are usually confused, and more often, TWS is recognized as TAS due to the parameter jitter caused by the measurement error, resulting in the pulse characteristics of TWS being similar to those of TAS. On the other hand, when the sample length is extremely short, TAS may be confused with TWS and MTT, and more often, TAS is misidentified as TWS, mainly due to the loss of tracking pulses. TAS is also identified as MTT because the partial loss of TAS results in data appearing as a "short straight-line" feature, which is similar to that of MTT.
This article considers the problem of operation mode recognition under parameter interval overlap and parameter measurement error, but ignores the existence of value mutation and data missing in the pulse reception process. Value mutation is mainly caused by pulse leakage, pulse splitting, and inaccurate pulse width measurement. For example, pulse splitting can cause some PW values to decrease. Data missing refers to missing data at certain points or time segments, which can lead to the loss of all features under that point or segment. The main reasons for this include barrage jamming or reconnaissance equipment failure. Data missing may cause significant changes in features, even leading to confusion with other operating states, resulting in inaccurate identification of the true operating state. Compared with value mutation, data missing has a greater impact on operating state recognition. Therefore, in the future, we will explore radar operation mode recognition methods under value mutation and data missing, and consider the influence of different missing lengths, different missing locations, and mutations of different sizes on recognition performance.

VI. CONCLUSION
Accurate identification of the operating state of the adversary's radar is important for threat warning and electronic countermeasure decisions. This article systematically analyzes and summarizes several typical operation modes of airborne radar, focusing on the characteristics of signal parameters of each mode. It is pointed out that the essential difference among the operation modes is to achieve different tactical purposes, which is directly reflected in the signal pattern and beam scanning. To acquire as many interpulse parameter characteristics and scanning patterns as possible, the multifeature fused RSCN method is proposed to identify multiple operation modes with messy and weak radar signals. This method combines interpulse feature regularities to develop a comprehensive temporal representation of radar operating states. The stream-level identification framework allows for the extraction of more high-level characteristics, while potentially removing noise-relevant characteristics, thanks to the embedded RS module. The simulation results indicate that our proposed RSCN has outstanding performance in terms of overlapping parameters, which can make a paradigm for the reconnaissance of ECM and remote sensing systems.