Rank reduction of LSTM models for online vibration signal compensation on edge computing devices

Paper i proceeding, 2025

There is increasing demand for the deployment of machine learning models on lightweight microcontrollers and internet of things devices in situations where off-device processing is impractical. Deploying models on such devices can be challenging due to performance and memory constraints, as the memory footprint and the number of operations necessary for inference scale rapidly with model size. This study proposes a technique that employs singular value decomposition to reduce the number of weights in long short-term memory models while having a minimal impact on their accuracy, leading to improvements in both memory footprint and performance. Furthermore, the rank reduction process results in dense matrices, which are more computationally efficient when compared to sparse matrices. The proposed technique is particularly advantageous for deployment on microprocessors like the Teensy 4.0’s Arm Cortex-M7 CPU, which, despite having a floating-point unit, lacks significant parallelization features, making it an ideal candidate for demonstrating the benefits of rank reduction in real-world applications. We validate the method by training a long short-term memory model to perform non-linear signal compensation on micro-electromechanical systems accelerometer data, extending low-frequency resolution without requiring hardware upgrades. Training data was collected using a reference accelerometer and a function-generator-driven electromagnetic shaker. Performance evaluation shows that the rank-reduced long short-term memory model achieves a 16.5% reduction in parameter count (from 10,451 to 8,861 parameters) while maintaining nearly identical signal fidelity. The reduced model achieved a signal-to-noise ratio of 14.8498 dB compared to 14.0765 dB for the original. This is a key step towards the first known deployment of a deep-learning-based signal compensation technique on a drone-deployable structural health monitoring sensor package, showcasing the potential for scalable, efficient long short-term memory model deployment on edge devices in infrastructure monitoring applications.