<p>This research proposes a novel hybrid deep learning framework for denoising partial discharge (PD) signals captured through acoustic emission sensors. The model integrates convolutional neural networks (CNNs) and long short-term memory (LSTM) networks to efficiently remove noise from acoustic PD signals, leveraging CNN’s ability to extract spatial features and LSTM’s strength in capturing temporal dependencies. A multi-scale CNN module is incorporated to enhance the robustness of feature extraction by addressing varying signal scales and ensuring adaptability to complex noise patterns. The framework demonstrates superior performance in experiments conducted with measured and simulated PD signals contaminated by various noise levels. These experiments include signal-to-noise ratios (SNRs) ranging from low to high levels, validating the model’s effectiveness across diverse scenarios. Integrating multi-scale convolutional blocks allows the model to capture features across different time scales, significantly improving its ability to extract critical signal characteristics. The proposed method achieves significant SNR improvements, reduced root mean square errors, and enhanced signal clarity, making it an ideal solution for denoising PD signals. Its adaptability to non-stationary noise patterns further underscores its potential for broader applications in signal processing, condition monitoring, and predictive maintenance. By leveraging the strengths of both CNN and LSTM in a multi-scale architecture, this framework sets a new benchmark for acoustic signal denoising in high-voltage systems, bridging the gap between traditional and advanced deep-learning techniques.</p>