Research on the Method of Audio Data Synthesis for Road Construction Machinery Using Generative Adversarial Networks

This paper tackles the challenge of scarce audio data from construction machinery in developing early warning systems for underground utility strikes during road construction. We introduce an innovative audio data synthesis method leveraging Generative Adversarial Networks (GANs). We gather an initial set of raw audio samples by employing data crawlers. Our approach integrates HIFI-GAN's feature extraction with Bidirectional Long Short- Term Memory networks (BiLSTM) to meticulously fine-tune and train a specialized H-GB model for audio synthesis. The H-GB model achieves a stable loss value of around 0.21, signifying exceptional performance. Our tests confirm the successful generation of high-fidelity audio data, aligning with our objectives. This paper's contribution improves the synthesis quality of construction machinery audio. It bolsters the model's capability to learn from construction scenario audio features, offering a robust solution for enhanced audio data synthesis in construction monitoring applications.