Real-Time Anomaly Detection in Grinding Wheels Using a Multimodal Deep Learning Framework

In the manufacturing process involving grinding wheels, challenges arise in fine-tuning grinding machines, typically addressed by craftsmen through subjective observations of sparks and sounds. This paper introduces a novel mechanism comprising two pivotal phases aimed at optimizing grinding wheel production line efficiency and accuracy. Firstly, an AutoEncoder is employed for spectrogram denoising, effectively isolating grinding sounds from environmental noise. Convolutional Neural Networks (CNNs) in the Encoder extract features across time and frequency domains, while deconvolution in the Decoder gradually restores features. ReLU activation ensures computational efficiency and effectively handles nonlinear features. Secondly, an AI-based assessment determines parameter adjustments using a combination of 3DCNN and CNN. By integrating classification results from both networks, features from video and audio data are identified, thereby enhancing classification effectiveness. Anomalies during grinding operations are detected through combined outputs, indicating the need for parameter adjustments