Accelerating radar signal modulation recognition based on data flow architecture
[Objective]In the highly specialized domain of radar electronic warfare,where every millisecond of delay can mean the difference between detecting and countering enemy radar systems effectively or missing the opportunity entirely,the ability to swiftly and accurately identify the modulation techniques of adversarial radar signals is critically imperative for securing a tactical advantage.Traditional methodologies often rely on graphics processing units(GPU)for signal recognition and have been proven inadequate in meeting the low-latency requirements of such application scenarios.To address this pivotal challenge,this paper introduces a novel acceleration system for radar signal modulation recognition that leverages a dataflow architecture(DF)to overcome the limitations in low-latency signal processing.This system integrates a binarized convolutional neural network(CNN)structure optimized for deployment on field-programmable gate arrays(FPGA),representing a major shift from conventional approaches.[Methods]Our methodological framework centers on developing a binarized CNN architecture that drastically reduces the computational overhead traditionally associated with signal recognition tasks.By quantizing the weights of the CNN to binary values,the system significantly diminishes resource requirements,enabling a more streamlined and energy-efficient operation.The adoption of a dataflow architecture further enhances this approach by simplifying the multiplication operations within the network to mere additions or subtractions.This design choice ensures efficient data movement through the processing units,minimizing idle times and maximizing system throughput,a critical factor in achieving high-speed data processing and significant reductions in power consumption.Furthermore,this framework facilitates parallel processing,thereby enhancing the system's data throughput efficiency.[Results]Simulation experiments conducted to assess the performance of the proposed acceleration system in recognizing eight different types of radar signal modulations reveal compelling findings.The experiments showcased the system's capabilities when benchmarked against classic neural network models,such as ResNet50 and MobileNetV3-Large.Notably,the binarized CNN maintained comparable recognition accuracy levels at higher signal-to-noise ratios(SNRs),satisfactorily meeting the general SNR requirements.This outcome is particularly noteworthy as it underscores the system's ability to maintain overall recognition accuracy without compromising processing speed or energy efficiency.Specifically,the acceleration system exhibited an extraordinary increase in inference speed,showing a remarkable 44.43 fold improvement over the i7-11800H CPU and a 2.59 fold enhancement compared to the RTX 3050Ti GPU.Moreover,this was achieved while keeping the system's power consumption to a minimal 1.724 W.[Conclusions]The research conclusively shows that the strategic simplification of network weights does not impede the system's ability to recognize radar signals accurately.Rather,this simplification significantly accelerates processing speed,thereby enhancing the system's practical applicability in the dynamic contexts of electronic warfare.This study enriches the dialogue on improving radar signal recognition methods and introduces a viable and efficient solution that is pivotal in enhancing tactical response capabilities in electronic warfare.The implications of this research are manifold,opening avenues for further investigation into optimizing signal processing systems for defense applications.By establishing a solid framework for the rapid and accurate identification of radar signal modulations,this research sets a new benchmark for future advancements in electronic warfare technology.It offers a valuable reference for subsequent studies in this critical field.
modulation type recognitiondeep learningdataflow architecturebinarized neural networkhardware deployment
NETL
NSTL
万方数据
