Linear wear monitoring for ore grinding mills
Grinding mills for ore processing use rubber liners to protect against impacts, but wear leads to high costs. The project integrates external acceleration sensors, wireless data transmission and machine learning to monitor liner conditions.
Factsheet
- Schools involved School of Engineering and Computer Science
- Institute(s) Institute for Intelligent Industrial Systems (I3S)
- Research unit(s) I3S / Prozessoptimierung in der Fertigung
- Funding organisation Innosuisse
- Duration 01.12.2014 - 31.03.2017
- Head of project Prof. Dr. Axel Fuerst
- Partner ABB Schweiz AG - Low Voltage Power
Situation
The world’s largest grinding mills for minerals ore processing have diameters of up to 12m. To protect the expensive mill drum against impact of rocks and/or steel balls, hard rubber liners are used. Due to the immense throughput, any downtime of a grinding mill is very costly. Additionally, the liners itself are expensive parts. It is therefore beneficial to change liners as a late as possible. However, to achieve this, it is essential to have good information about the condition of the liners.
Course of action
An acceleration sensor is mounted on the mill drum to collect real-time data, which is wirelessly transmitted to a central storage unit. The time-series data is processed and analysed using machine learning algorithms to classify the liner's condition, enabling early detection of wear and minimising downtime.
Result
Finite-element-method (FEM) simulations and analysis demonstrated that the acceleration signal is highly sensitive to variations in liner thickness, providing a solid theoretical foundation for monitoring wear. To validate this, a physical scale model was constructed to test the full pipeline of the approach. Experiments were conducted with liners in different wear states, and measurements were collected using the acceleration sensor. The resulting data underwent preparation and analysis, where time-series classification was applied. The proof of concept (PoC) confirmed that the approach is feasible and effective, showing promising results for real-world applications. The field test results revealed the effectiveness of the approach in detecting linear wear on the liners. A neural network classifier with seven target classes was employed, using raw data sliced into drum revolutions. Additional process data, including throughput and operating temperatures, was incorporated to enhance classification accuracy. Key features such as signal statistics, FFT (Fast Fourier Transform), and wavelet analysis were extracted from the sensor data. The classifier achieved approximately 83% accuracy on test data. Notably, the shoulder and toe angles were identified through signal entropy analysis, emphasizing the importance of maintaining the correct rotation speed for accurate monitoring.
