A Hybrid CNN-LSTM Framework for Intelligent Fault Diagnosis and Predictive Maintenance of Electronic Systems

Electronic systems are fundamental to modern industry sectors such as manufacturing, automotive and energy. With the growing complexity of these systems, methods based on rules and conventions are failing to achieve effective reliability and timely maintenance. This paper proposes an intelligent deep learning-based framework for fault diagnosis and predictive maintenance of electronic systems using Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks for hybrid spatial-temporal feature learning. The fault data were generated using Simulations in SPICE and MATLAB/Simulink, and real-time measurements from embedded platforms such as Arduino, STM32, and Raspberry Pi were used to validate the simulations. Using signals sampled at 10 kHz, the system assesses multiple types of faults, including resistor drift, capacitor degradation, short-circuit and open circuit failures. The CNN-LSTM model, trained with the Adam optimizer at 100 epochs, achieved an accuracy of 97.4%, precision of 96.9%, recall of 96.2%, and final F1-score of 96.5%, performing better than the CNN and LSTM in isolation. The model was deployed as hybrid hardware on a Raspberry Pi 4, with a low inference latency of 18.5 ms, 120 MB memory footprint, and 2.5 W; demonstrating real-time functionality. Finally, it was demonstrated that the model can conserve energy use via model optimization (quantization, pruning) without sacrificing model performance (energy decrease by$\boldsymbol{\sim} \mathbf{1 2 {\%}}$). In order to better describe model predictions, explainable AI (XAI) tools were applied (SHAP, LIME, and Grad-CAM). The highlight of the proposed CNN-LSTM framework is that it provides diagnostic accuracy, energy use, and explainability, making it a well-suited framework for predictive maintenance systems that are energy efficient and scalable in Industry 4.0 and Industrial IoT systems.