Scholarlink

To address the real‑time classification requirements of aircraft engine vibration, airborne communication, and radar echo signals under resource‑constrained onboard hardware, this paper proposes an MLP classification framework that combines fast Fourier transform (FFT) feature extraction with a lightweight bottleneck fully connected (Bottleneck FC) structure. First, each raw time‑domain signal is processed by a 256‑point FFT to extract amplitude‑spectrum features, which are then standardized. Next, a bottleneck module is inserted at the MLP input to reduce the 256‑dimensional feature vector to 64 dimensions before reconstructing it, significantly lowering both parameter count and computation. Finally, a two‑layer fully connected network followed by a Softmax layer performs classification. Experiments show that adding only the Bottleneck FC increases accuracy from 97.65% (baseline MLP) to 98.82%, holds inference latency at 314.7 ms per 1000 samples, and enables real‑time deployment on an FPGA with 15,000 LUTs. The method delivers high precision with ultra‑low latency and hardware friendliness, offering a practical solution for intelligent signal monitoring in avionics systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FFT–MLP for Real‑Time Aircraft Signal Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shaozhen Pan,
  • Baoji He,
  • Yue Cheng

摘要

To address the real‑time classification requirements of aircraft engine vibration, airborne communication, and radar echo signals under resource‑constrained onboard hardware, this paper proposes an MLP classification framework that combines fast Fourier transform (FFT) feature extraction with a lightweight bottleneck fully connected (Bottleneck FC) structure. First, each raw time‑domain signal is processed by a 256‑point FFT to extract amplitude‑spectrum features, which are then standardized. Next, a bottleneck module is inserted at the MLP input to reduce the 256‑dimensional feature vector to 64 dimensions before reconstructing it, significantly lowering both parameter count and computation. Finally, a two‑layer fully connected network followed by a Softmax layer performs classification. Experiments show that adding only the Bottleneck FC increases accuracy from 97.65% (baseline MLP) to 98.82%, holds inference latency at 314.7 ms per 1000 samples, and enables real‑time deployment on an FPGA with 15,000 LUTs. The method delivers high precision with ultra‑low latency and hardware friendliness, offering a practical solution for intelligent signal monitoring in avionics systems.