Edge AI for Fault Diagnosis: Evaluation and Optimization of CNNs on the Jetson Nano Using the MaFaulDa Database

Abstract

The deployment of Convolutional Neural Networks (CNNs) for mechanical fault diagnosis at the edge is often limited by the parameter explosion in fully connected layers when processing high-dimensional spectral features, such as the Fast Fourier Transform (FFT). This paper investigates the feasibility of low-latency inference by comparing a conventional CNN architecture with a proposed Hybrid Dual-Branch architecture that fuses temporal raw signals and spectral descriptors. Experiments were conducted using the MaFaulDa dataset, which includes vibration measurements representing normal operation, imbalance, misalignment, and bearing faults. To rigorously isolate inference efficiency, preprocessing is performed offline. The evaluation considers both predictive performance (Accuracy and F1-score) and hardware constraints on the NVIDIA Jetson Nano platform. Results demonstrate that, by leveraging Global Average Pooling (GAP) alongside a parallel-branch design, the model prevents parameter inflation when fusing domains. This architecture-level optimization enables a highly compressed 608 KB footprint while achieving 93.45% accuracy and sub-millisecond inference latency of 0.434 ms (2,303 inferences per second), demonstrating its viability for memory-bound edge devices.