Early Fault Detection for Industrial Machinery

Knowing in advance when an asset is likely to fail is advantageous in order to avoid unplanned downtime and asset damage. Predictive maintenance has been demonstrated to reduce maintenance costs by 25% and reduce asset breakdowns by 70%. In some cases, automotive assembly line manufacturing stops can cost £10,000 per minute, so there is a real incentive to cultivate predictive capabilities to detect developing faults in industrial machinery. The development of AI models for early fault detection has widespread applications across industry, including shipping, food and beverage, automotive, and logistics.

TWI has been working alongside a consortium comprising Vibtek Ltd, CMServices Global Ltd and Brunel University to develop a prescriptive maintenance system as part of the DiMOS Project, which has also gained input from Condor Ferries on the South Coast of England, William Grant and Sons in Scotland, Royal Mail in Sheffield, and Toyota Manufacturing UK in Derby.

The system was designed to perform prescriptive maintenance using artificial intelligence models to allow for early fault detection. TWI led the development of the application of artificial intelligence-based models to identify and locate developing faults within industrial machinery such as motors, fans and pumps at the case study sites. The system has been designed to take into account risk levels and likely failure modes into the prescriptive action.

The aim was to be able to detect and locate a fault at an early stage as it developed using real-time vibration sensor data. An automated process was developed to analyse and interpret new measurements as they were recorded. This process alleviates the time-consuming manual interpretation of the vibration data and allows maintenance to be undertaken in a timely manner to save on unnecessary maintenance costs and prevent damage to assets.

The approach taken was to determine asset condition in relation to normal healthy operation of the asset. Using vibration analysis, one of the main techniques for fault detection in rotating machinery, and a variety of signal processing methods, features were extracted from both the time and frequency domains. Due to the sparse nature of equipment failure, the use of unsupervised novelty detection algorithms to detect developing faults was chosen. These algorithms train the model on healthy machine operating data and detect changes in the measurements which could be indicative of a developing fault. The output of the model was a health indicator giving a qualitative indication of the operating condition of the asset.

The results demonstrated the capability to detect faults at an early stage; up to three days before asset failure. The model was also capable of locating the fault and giving a qualitative indication of the severity of the fault as the condition deteriorates. These results were verified on a controlled experimental set up with real vibration data and real data from industry sites across the UK.

The DiMOS Project has received funding from Innovate UK with Project Reference No. 104505