Real-time fall detection on microcontroller-based wearable devices requires temporal discrimination under tight memory, latency, and quantization constraints. Convolutional neural network-long short-term memory (CNN-LSTM) models can capture inertial measurement unit (IMU) dynamics, but retaining recurrent inference complicates deployment on microcontroller units (MCUs). This paper presents a Tiny Machine Learning (TinyML) framework that uses cross-architecture knowledge distillation (KD) to transfer supervision from a CNN-LSTM teacher to a compact one-dimensional convolutional neural network (1D-CNN) student. The teacher contains 152,674 trainable parameters, whereas the student contains 20,210 parameters, corresponding to an 86.76% reduction. After training, the student is quantized to INT8 and deployed. On the KFall dataset, under two subject-wise train-test protocols, the distilled student achieved F1-scores of 97.97% and 97.64%, closely matching the teacher (97.97% and 97.71%) and improving over the student trained from scratch (97.19% and 97.10%). After full INT8 quantization, the student retained F1-scores of 97.93% and 97.66%. Deployment with TensorFlow Lite for Microcontrollers (TFLM) on an ESP32-S3 required 32.9 KB Flash and 9.6 KB RAM, with a mean inference latency of 2.51 ms. These results indicate that cross-architecture KD can transfer temporal teacher guidance into a compact, quantization-friendly feed-forward model for real-time fall detection on resource-constrained edge devices.

TinyML,Fall Detection,Knowledge Distillation,Microcontrollers,Edge AI,Cross-Architecture