Continual Radar Intra-Pulse Modulation Recognition via Multiview Replay and Class-Balanced Classifier Fine-Tuning Authors: Yilin Liu, Xiaofang Wu, Caoyi Mei, Shuo Yu Published: 2026 ABSTRACT Continually deployed radar recognizers must acquire emerging intra-pulse modulation classes without erasing learned signal types under small replay memories. We present MM-CILNet, a radar-specific multiview replay framework that combines complementary I/Q and short-time Fourier transform (STFT) representations, sample-wise fusion, class-balanced memory, and a one-epoch update of three classifier heads after each incremental session. On the 20-class CIL-20 protocol with 12 matched seeds, MM-CILNet improves final-session All and Old accuracy by 2.7 percentage points each over matched Replay (All: p=0.0034, Cohen’s d=1.07; Old: p=0.0103, d=0.89). Under the shared configuration, it achieves the best All and Old accuracy among the listed radar-adapted CIL methods and the lowest average forgetting among the four methods with recoverable session histories. Equal-budget controls identify significant post-session effects on All and Old accuracy and show that trainable scope and sampling determine the retention–plasticity operating point. The Old-accuracy advantage persists at the tested 400- and 800-sample budgets and across five class orders. Under 2.5:1 recent-class-heavy replay, Old accuracy decreases by 1.63–2.01 pp and forgetting increases by 3.66–4.69 pp relative to class-balanced replay. These results establish class-balanced memory as the retention-oriented operating choice and provide an integrated protocol for memory-constrained continual radar recognition. 1. Introduction Radar intra-pulse modulation recognition is a core sensing task in electronic reconnaissance, spectrum monitoring, and cognitive electronic warfare. Modern radar sensors operate in increasingly congested and dynamic electromagnetic environments where new waveform classes appear incrementally after initial deployment [1,2]. A practical recognition system must therefore acquire new modulation types, while retaining previously learned ones, without storing all historical samples or retraining from scratch. This requirement defines a class-incremental learning (CIL) problem under real-world sensing constraints: limited memory, evolving emitter populations, and a wide signal-to-noise-ratio (SNR) range. The CIL problem has been extensively studied in computer vision [3,4,5,6,7,8], where the catastrophic forgetting of old classes is the central challenge. Radar signal recognition adds several practical sensing considerations that distinguish it from vision-domain CIL. First, raw radar signals are time-series I/Q complex samples, not images; their discriminative information resides in instantaneous phase continuity, frequency trajectories, and intra-pulse modulation structure. Second, the CIL-20 protocol evaluates a wide SNR range, from −18 dB to +12 dB, over which the usefulness of different signal representations varies with channel conditions. Third, modern radar waveforms increasingly use compound modulation types that combine multiple coding schemes within a single pulse, creating intra-class variability that challenges fixed fusion strategies. Prior work on radar signal recognition has mainly considered closed-set classification [9,10,11,12], where all classes are known during training. Radar systems can exploit complementary information from either different sensors or different representations of the same measurement. Multimodal ISAC systems, for example, integrate camera and mmWave radar observations for V2I tracking [13]. Multiview modulation recognition follows the latter approach by constructing I/Q and time-frequency inputs from the same received signal [14,15,16]. However, existing multiview studies assume a fixed class set, leaving open whether these complementary representations remain beneficial when a model must learn new classes without forgetting those learned earlier. This paper makes four contributions:We formulate continual radar intra-pulse modulation recognition and present MM-CILNet, a multiview replay framework that integrates I/Q and STFT representations, class-balanced replay, post-session classifier-head fine-tuning, and sample-wise gated fusion.On the 20-class CIL-20 protocol, MM-CILNet improves final-session All and Old accuracy over matched Replay by 2.73 and 2.67 percentage points, respectively. Under the shared configuration, it achieves the best All and Old accuracy among the listed radar-adapted CIL methods and the lowest average forgetting among the four methods with recoverable session histories.Through a 12-seed, equal-budget decomposition, we isolate fusion, post-stage parameter scope, and sampling. The factorial analysis identifies a general post-session effect on All and Old accuracy and shows how trainable scope and sample allocation select different retention–plasticity operating points.We evaluate the complete protocol across memory budgets, class orders, replay allocations, distribution shifts, and three related datasets. The Old-accuracy advantage persists at the tested 400- and 800-sample budgets and across all five tested class orders; under 2.5:1 recent-class-heavy replay, the Old-accuracy and forgetting endpoints identify class-balanced memory as the retention-oriented choice. We formulate continual radar intra-pulse modulation recognition and present MM-CILNet, a multiview replay framework that integrates I/Q and STFT representations, class-balanced replay, post-session classifier-head fine-tuning, and sample-wise gated fusion. We formulate continual radar intra-pulse modulation recognition and present MM-CILNet, a multiview replay framework that integrates I/Q and STFT representations, class-balanced replay, post-session classifier-head fine-tuning, and sample-wise gated fusion. On the 20-class CIL-20 protocol, MM-CILNet improves final-session All and Old accuracy over matched Replay by 2.73 and 2.67 percentage points, respectively. Under the shared configuration, it achieves the best All and Old accuracy among the listed radar-adapted CIL methods and the lowest average forgetting among the four methods with recoverable session histories. On the 20-class CIL-20 protocol, MM-CILNet improves final-session All and Old accuracy over matched Replay by 2.73 and 2.67 percentage points, respectively. Under the shared configuration, it achieves the best All and Old accuracy among the listed radar-adapted CIL methods and the lowest average forgetting among the four methods with recoverable session histories. Through a 12-seed, equal-budget decomposition, we isolate fusion, post-stage parameter scope, and sampling. The factorial analysis identifies a general post-session effect on All and Old accuracy and shows how trainable scope and sample allocation select different retention–plasticity operating points. Through a 12-seed, equal-budget decomposition, we isolate fusion, post-stage parameter scope, and sampling. The factorial analysis identifies a general post-session effect on All and Old accuracy and shows how trainable scope and sample allocation select different retention–plasticity operating points. We evaluate the complete protocol across memory budgets, class orders, replay allocations, distribution shifts, and three related datasets. The Old-accuracy advantage persists at the tested 400- and 800-sample budgets and across all five tested class orders; under 2.5:1 recent-class-heavy replay, the Old-accuracy and forgetting endpoints identify class-balanced memory as the retention-oriented choice. We evaluate the complete protocol across memory budgets, class orders, replay allocations, distribution shifts, and three related datasets. The Old-accuracy advantage persists at the tested 400- and 800-sample budgets and across all five tested class orders; under 2.5:1 recent-class-heavy replay, the Old-accuracy and forgetting endpoints identify class-balanced memory as the retention-oriented choice. 2. Related Work 2.1. Radar Signal Recognition Radar signal recognition has evolved from handcrafted feature engineering to deep learning. Early work used time-frequency transforms with expert-defined features [1]. Convolutional networks applied to time-frequency images [9,10] and directly to I/Q sequences [17,18] improved closed-set accuracy. Sequence models have also provided a unified treatment of radar and communications signals together with a public classification dataset [19]. Recent fixed-set studies address limited labels through semi-supervised time-frequency recognition [20] and variable noise through the AIMC-Spec intrapulse benchmark [21]. Adjacent radar-recognition work addresses scarce working-mode sequences through generative augmentation [22] and low-SNR jamming recognition through time-, frequency-, and time-frequency-derived feature fusion [23]. Multiview approaches combining multiple representations [14,15,16] further improved robustness. These studies strengthen recognition under fixed label vocabularies or adjacent radar-signal tasks. MM-CILNet instead targets sequential expansion of intrapulse modulation classes and identifies the components that preserve previously learned classes during that expansion. 2.2. Class-Incremental Learning Class-incremental learning addresses catastrophic forgetting when models are sequentially updated with new classes. Foundational regularization methods stabilize parameters important to earlier tasks [24], while output distillation preserves previous predictions without retaining the original training set [25]. Contemporary CIL families include replay-based methods [3,6], bias-correction methods [4,7], distillation-based methods [5], and feature-boosting methods [8]. Replay stores a subset of old-class samples and interleaves them with new-class data during incremental training. Recent replay research improves exemplar selection using distribution and decision-boundary criteria [26], adapts coreset construction to time-varying class frequencies [27], and jointly optimizes exemplars and knowledge transfer through data and feature distillation [28]. Bias correction addresses the classifier tendency to favor recently seen classes by post-hoc adjustment of output logits or classifier weights, whereas distillation penalizes changes in feature representations or output distributions. Our work builds on replay and decomposes the contributions of dual-representation fusion, post-session optimization, parameter scope, and sampling. 2.3. Multiview Learning for Radar Signals Radar signals can be represented in multiple domains. I/Q complex samples preserve instantaneous phase and amplitude, while time-frequency representations (STFT, WVD) expose modulation structure. General gated multimodal units learn intermediate representations by controlling the contribution of modality-specific features [29]. Radar-specific work has shown benefits from fusing multiple representations in closed-set classification [14,15,16,30,31]. Common strategies include early concatenation, intermediate attention, and late averaging. Our experiments extend this comparison to CIL and test representation and post-stage effects in the same matched-seed design. 2.4. Classifier Rebalancing in Class-Incremental Learning Classifier rebalancing has been studied in general CIL. BiC [4] learns a bias-correction layer on class-balanced validation data after each session. WA [7] aligns old and new classifier weight norms to reduce prediction bias toward recently seen classes. LUCIR [32] combines cosine normalization with margin-based constraints for a more balanced feature space. MM-CILNet instead fine-tunes its three existing cosine classifier heads with cross-entropy on updated replay memory while freezing both encoders and the gate. We call this operation class-balanced classifier-head fine-tuning: it rebalances decision boundaries but does not calibrate predictive probabilities. Classifier-level CIL architectures have also combined probability dampening with cascaded gates that rescale task-wise classifier outputs [33]. Those gates modify the incremental classifier, whereas the MM-CILNet sample-wise gate mixes the I/Q and STFT embeddings before its shared classifier. This distinction separates classifier-output intervention from representation fusion. Methodologically, MM-CILNet joins same-pulse multiview learning and class-incremental replay at the point where their complementary strengths matter. Fixed-set multiview methods fuse signal representations without addressing sequential class expansion, whereas generic replay and classifier-rebalancing methods do not exploit the relationship between I/Q and STFT views of one radar pulse. MM-CILNet couples two representation-specific encoders and auxiliary heads with sample-wise fusion, class-balanced replay, and a bounded post-session head update. Its contribution is this radar-specific continual-recognition protocol together with a matched decomposition that identifies how representation, post-stage scope, and sampling govern retention and plasticity. Table 1 makes the method-level distinction explicit. The closest families each address only part of the problem: fixed-set radar multiview methods learn complementary representations without class expansion, and established CIL methods manage class expansion without a same-pulse I/Q–STFT fusion interface. MM-CILNet couples both in one controlled protocol and evaluates the coupling through matched factor-level interventions. 3. Methods 3.1. Problem Formulation Figure 1 provides an overview of the proposed framework. Let Dt={(xiIQ,xiTF,yi)}i=1Nt denote the training set at session t, where xiIQ∈CT is the complex I/Q sequence of length T=1024, xiTF∈RH×W is the STFT magnitude image of size 224×224, and yi∈Ct is the modulation class label. The class set Ct expands as Ct=Ct−1∪Ctnew, with Ct−1∩Ctnew=∅. A fixed-size replay memory M stores a subset of old-class samples, updated to remain class-balanced across all seen classes. The goal is to learn a classifier ft:XIQ×XTF→Ct that performs well on all classes observed up to session t, without access to raw samples from sessions 0 through t−1 beyond those in M. 3.2. Dual-Representation Feature Extraction Figure 2 shows the dual-encoder architecture of MM-CILNet. The I/Q encoder receives the two-channel I/Q sequence and applies four sequential multi-scale one-dimensional blocks with output widths 32, 64, 128, and 256. Within each block, parallel convolution–batch-normalization–ReLU branches with kernel sizes k∈{3,5} are concatenated, normalized and activated, and then downsampled by max pooling. Global average pooling and a linear projection produce a 256-dimensional output, which is ℓ2-normalized to obtain zIQ∈R256. The time-frequency encoder processes the STFT magnitude image with a lightweight squeeze-and-excitation residual backbone. A 7×7 stride-2 convolutional stem with batch normalization and ReLU is followed by 3×3 stride-2 max pooling and four residual stages with channel widths 32, 64, 128, and 256. Each stage contains two residual blocks, each comprising two 3×3 convolutions and a squeeze-and-excitation unit; the first block in stages 2–4 performs stride-2 downsampling. Global average pooling and a linear projection produce a 256-dimensional output, which is ℓ2-normalized to obtain zTF∈R256. 3.3. Sample-Wise Gated Fusion The embeddings zIQ and zTF are fused by a sample-wise gate. Each representation has a cosine auxiliary classifier. Its scalar score cue is the maximum softmax probability, cv=maxksoftmax(ℓv)k, for v∈{IQ,TF}. Following the implementation order, the gate forms g=[zTF;zIQ;cTF;cIQ]∈R514. A two-layer MLP (514–256–2, ReLU between layers) produces two logits, which are softmax-normalized as (αTF,αIQ); output indices 0 and 1 therefore weight the time-frequency and I/Q embeddings, respectively. The fused embedding is z=norm(αTFzTF+αIQzIQ) and is passed to the fused cosine classifier. The concatenated embeddings carry representation-specific signal content and feature geometry, including agreement and conflict between the two views. The two maximum-softmax values summarize how sharply each representation-specific head separates its current class scores and enter the gate only as scalar context. During normal incremental training, both encoders, both auxiliary heads, the fused head, and the gate are optimized jointly with(1)Ltrain=CEfused+0.2CEIQ+CETF, allowing the MLP to learn interactions between feature content and score sharpness. The maximum-softmax values condition the representation mixture; they are not the predictive-entropy variable used in the post hoc stratified analysis. The softmax outputs (αTF,αIQ) are learned sample-wise mixing coefficients. Direct Gate-versus-Mean comparisons in Section 5.3 evaluate whether this adaptive mixture improves classification over uniform averaging under AWGN, fading, and fading-plus-impulsive noise. Fusion is applied after global pooling so that every fusion control combines the same two fixed-dimensional representations. The I/Q branch has a one-dimensional temporal axis, whereas the STFT branch has two-dimensional time–frequency bins; these positions are branch-specific and do not define a common spatial or temporal index over which one shared attention rule can be applied. An encoder-internal attention module would therefore change feature extraction, capacity, and the intervention location instead of isolating the fusion rule. The matched comparator applies one shared 256–128–1 Tanh scorer to the two pooled representation tokens and softmax-normalizes the two scores before weighted summation. Section 5.3 compares this representation-level attention, Gate, and Mean at the common fusion interface. 3.4. Post-Session Classifier-Head Fine-Tuning After each incremental session’s joint training on current-session and replay data, the default MM-CILNet post stage freezes both encoders and the gate and updates three cosine classifier heads: the fused head and the I/Q- and STFT-specific auxiliary heads. It uses one epoch of cross-entropy training at 2×10−4 (0.2 times the main learning rate) on the updated 400-example class-balanced memory. With batch size 16, this is 25 optimization steps and 400 processed examples per incremental session. This operation changes decision boundaries but does not change the inference architecture. Section 5.3 also evaluates head-only sampling from the full current-session-plus-replay pool, full-model class-balanced fine-tuning, and full-model sampling from that same pooled training set with the same example and step budget. 3.5. Training Procedure All models are trained using the Adam optimizer with an initial learning rate of 1×10−3 and weight decay of 1×10−4. The batch size is 16. The base session (S0) trains for 5 epochs; each incremental session (S1–S6) trains for 3 epochs. These hyperparameters are held constant across MM-CILNet and all compared baselines. The main replay memory budget is 400 samples, updated to remain class-balanced over seen classes after each session. Table 2 separates inference overhead from post-stage training overhead. The CPU benchmark used an AMD Ryzen 7 7735HS processor (Advanced Micro Devices, Inc., Santa Clara, CA, USA), one PyTorch 2.6.0 CPU thread, batch size 1, 10 warm-up passes, and 50 timed forward passes; latency is the arithmetic mean per sample. Gated fusion takes 22.2 ms, 3.5% longer than Mean fusion (21.4 ms). The current shared model container serializes 3.502 M parameters because it instantiates all fusion controls, whereas the executed forward paths contain 3.272 M parameters for Gate and 3.140 M for Mean. At K=20, the post-session stage updates only the three classifier heads (15,360 parameters). Removing dormant controls gives 3.272 M parameters for a Gate deployment; a Mean-only fused-prediction deployment can additionally remove the two unused auxiliary heads, giving 3.130 M. Let CIQ and CTF denote the forward costs of the two encoders, d the embedding dimension, h the gate hidden width, K the number of seen classes, and M the post-stage sample count. MM-CILNet inference has complexity O(CIQ+CTF+(2d+2)h+2h+3dK): the first two terms are shared by all dual-representation comparisons, the middle terms are the two-layer gate, and the final term covers the fused and two auxiliary cosine heads. Mean fusion removes the gate term. The gate MLP contains 132,354 parameters, the two-token attention scorer contains 33,025, and the two auxiliary heads contain 10,240 at K=20. For d=h=256, the Gate overhead is small relative to the encoders, which is consistent with the measured 3.5% CPU latency increase. One head-only post stage costs O(M[CIQ+CTF+(2d+2)h+3dK]) in forward computation and O(MdK) in trainable-head backpropagation; it does not backpropagate through the encoders or gate. Full-model post-stage controls have the same forward order but additionally backpropagate through both encoders and the gate. Thus, relative to the common replay-style adaptations, the leading encoder complexity is unchanged, and the additional costs are confined to the lightweight gate and one bounded post-session pass. For a direct comparison with the implemented CIL families, define B=CIQ+CTF, G=(2d+2)h+2h, and H=3dK. Table 3 lists method-specific additions beyond the common current-plus-replay update. All shared-backbone Gate adaptations have the same encoder-dominated inference order O(B+G+H); their differences occur during incremental training or post-session processing. Compared with the distillation families, MM-CILNet avoids a frozen previous-network forward pass; compared with DER-style replay, it avoids storing and matching old logits. Its additional work relative to plain replay is the bounded M=400 head-only pass. 4. Experimental Setup 4.1. CIL-20 Dataset and Protocol We construct CIL-20, a 20-class radar intra-pulse modulation dataset under an 8+2×6 class-incremental protocol. The base session S0 contains 8 classes; each subsequent session S1–S6 introduces 2 new classes, with all 20 classes appearing by S6. After each session, evaluation covers all classes observed so far, matching the CIL condition that task identity is unavailable at test time. CIL-20 includes both basic modulation families and compound modulation types, increasing the difficulty of unified recognition after continual class expansion. The 20 classes span four modulation families: Basic (NS, LFM, Barker), Phase (Frank, P1, P2, P3, P4), Time/Frequency (Costas, T1, T2, T3, T4, NLFM, HFM, SFM), and Compound (BPSK-LFM, Frank-LFM, P4-LFM, SF-PC). Note that Costas codes are frequency-hop waveforms and are classified within the Time/Frequency family, which is consistent with their time-frequency coding structure. Table 4 summarizes the dataset and protocol parameters used in the main experiments. Each clean signal is normalized to unit root-mean-square amplitude, corrupted by complex AWGN at the selected SNR, and normalized again. A single NumPy generator seeded at 42 advances for every generated example. Train, validation, and test files are generated in separate loop iterations and have disjoint file identifiers and noise draws. Classes with randomized waveform parameters also receive independent parameter draws across splits; classes defined by a fixed phase-code formula differ across splits primarily through independently generated noise. Thus, the split does not reuse one clean waveform with three noise labels for randomized classes, but it is not a source-separated measured-signal split. Table 5 summarizes the clean-waveform parameters. Frequencies and bandwidths are fractions of the 100 MHz sampling rate fs; Tp=1024/fs is the pulse duration. Compound classes multiply the corresponding independently generated component waveforms before normalization. For the STFT view, the complex signal is Fourier-resampled from 1024 to 256 samples. SciPy’s STFT uses its default Hann window, nperseg=64, noverlap=48, nFFT=64, two-sided output, zero boundary extension, and padding. The magnitude is min–max scaled per sample, bilinearly resized to 224×224, and clipped to [0,1]. Continual class expansion then follows the 8+2×6 protocol. A fixed 400-example replay memory is updated by seeded random selection to remain class-balanced over seen classes. All compared methods use the same class order, memory budget, data split, and preprocessing unless stated otherwise. Beyond the main CIL-20 protocol, we run distribution-shift and external validation experiments. The shift tests include fading and stronger fading-plus-impulsive-noise conditions. Three radar-related datasets provide related-task transfer checks with different signal sources, label spaces, and recognition targets. Table 6 defines the evidential role of each dataset and distinguishes these transfer checks by signal source, label space, and recognition task. 4.2. Compared Methods MM-CILNet is compared with a direct replay baseline and representative CIL-style adaptations. Replay uses the same dual-representation backbone and class-balanced memory update as MM-CILNet but omits the post-session stage. This comparison estimates the complete-protocol effect; the matched controls in Section 5.3 separately identify additional optimization, parameter scope, and sampling. The “-style” label denotes in-house adaptations under the common MM-CILNet training pipeline. The dual-representation backbone, random class-balanced memory update, data split, class order, and seed set remain fixed. iCaRL-style uses replay plus output distillation (weight 1, temperature 2); BiC-style uses the same distillation plus one class-balanced post-session head update in place of the official held-out validation layer; PODNet-style adds prototype regularization (weight 0.1) to distillation; DER-style matches stored replay logits with MSE (weight 1); DER+head-FT-style adds the same post-session head update to DER; and WA-style performs post-hoc old/new classifier-weight norm alignment. The resulting table is a controlled mechanism-family comparison under one signal-processing and training path; performance of the official methods remains defined by their original implementations and protocols. We focus on backbone-agnostic replay, bias-correction, distillation, and weight-alignment families because they can share the signal input and memory protocol. The common base schedule is 5 epochs for S0 and 3 epochs per incremental session. Method-specific loss weights and post-hoc operations are retained as listed above. The reported results characterize these adaptations under the shared configuration; method-specific optimization is reserved for direct reproduction of each official implementation. The ablation study separates representation, fusion, post-stage parameter scope, sampling, and additional optimization. All matched-control conditions use the same seeds, 400 post-stage examples, and 25 post-stage mini-batches whenever the post stage is enabled. Table 7 summarizes the shared conditions and adaptation-specific differences used to ensure a controlled comparison. 4.3. Evaluation Metrics and Statistical Tests The primary metrics are All accuracy, Old accuracy, New accuracy, average forgetting, macro-F1, and the old/new bias gap. All accuracy is computed over all classes observed up to the current session. Old accuracy is computed over classes introduced in previous sessions, and New accuracy over classes introduced in the current session. Together, these metrics distinguish unified recognition, old-class retention, and new-class learning. Let aj,t denote the test accuracy on the classes introduced in session j after training through session t. At final session T, forgetting for session j