Innovative predictive maintenance for mining grinding mills: from LSTM-based vibration forecasting to pixel-based MFCC image and CNN

This article presents an innovative predictive maintenance for grinding mills, aiming to enhance operational efficiency and minimize downtime. Leveraging advancements in data analytics and IoT sensor technologies, the approach integrates vibration signal forecasting, LSTM-based fast Fourier transform (FFT) analysis, and convolutional neural networks (CNNs) to detect faults early on. The method involves creating LSTM models to forecast vibration signals based on historical data and using FFT analysis to identify fault frequencies associated with the grinding process. Additionally, techniques such as Mel-frequency cepstral coefficients (MFCCs), short-time Fourier transform (STFT), and continuous wavelet transform (CWT) are employed for spectrogram extraction, providing valuable insights into machinery conditions. Validation on real-world datasets with 99.95% of accuracy with 1 in AUC-ROC, showcases the robust predictive performance of the model and has reached 99.96% of accuracy for 16 classes with an AUC-ROC of 1 using CWRU dataset, surpassing existing approaches and demonstrating its potential for proactive maintenance across various industries beyond mining.