AdaLiteNet: A Channel-Adaptive Lightweight Neural Network for Symbol Detection in Dynamic Wireless Channels

Reliable symbol detection in wireless communications is often challenged by rapidly varying channel conditions, especially on resource-constrained platforms. To address this issue, we propose AdaLiteNet, a lightweight channel-adaptive neural network for symbol detection. AdaLiteNet adopts a symbol-by-symbol detection strategy with Long Short-Term Memory (LSTM) units, reducing size of parameters to 51.3 K while preserving temporal dependencies. A two-stage framework is employed: (1) LiteNet, an offline-trained general detector initialized with joint-channel datasets, and (2) AdaLiteNet, an adaptive detector fine-tuned online when decoding performance drops. In addition, a multi-point collaborative detection algorithm enriches training data by exploiting distributed receivers. Experiments with real-world Automatic Dependent Surveillance-Broadcast (ADS-B) signals demonstrate that AdaLiteNet increases the decoding success rate by about 20 percentage points compared with the pulse-position modulation (PPM) baseline on low-quality datasets. Further tests on simulated Quadrature Phase Shift Keying (QPSK) and 16-Quadrature Amplitude Modulation (16-QAM) signals confirm strong adaptability under dynamic channels. Moreover, the lightweight design enables AdaLiteNet to achieve near real-time inference on resource-constrained platforms, making it a practical solution for deployment on IoT edge devices.