Determining the integrity of single-source condition-based maintenance data

Abstract

Data plays a vital role in modern society. With an influx of sensors, data generation is rapidly increasing. Having access to these vast amounts of data enables data-driven decision-making. However, reliable data are needed to make informed decisions, as unreliable data can lead to negative consequences. Reducing costs has become crucial with the mining industry under enormous pressure to remain profitable. By implementing preventative condition-based maintenance, unnecessary maintenance, downtime and expenses can be avoided. However, reliable data are needed to be effective. Due to financial pressures, often only single-source data are available for planning purposes. Data can be analysed using data quality dimensions from intrinsic and contextual perspectives to evaluate the reliability. The literature identified a need to use the combination of intrinsic and contextual methods to estimate the integrity of single-source condition-based maintenance data. To address the identified need, this study proposes a novel method to estimate the integrity of single-source condition-based maintenance data using a combination of intrinsic and contextual methods. This method was developed by using data quality dimensions and existing methods from literature. To supplement the proposed novel method, software system design to estimate the integrity of single-source condition-based maintenance data using a combination of intrinsic and contextual methods is proposed as an additional novel contribution. The software system design makes use of the proposed novel estimation method to evaluate the running status, electrical current drawn, temperature and vibration data streams of pumps, compressors, fridge plants and fans. A software system was created using the novel contributions. The system was verified using three datasets; two clean datasets and one erroneous dataset, correctly identifying roughly 94.5% of data points. The system was implemented on twenty case study components and identified 22% of data points as unreliable. From the data points identified as unreliable, 69% were identified using the contextual methods, highlighting the need for the combination of intrinsic and contextual methods. A few

common data-related problems were identified, with uncalibrated sensors being the most commonly occurring issue. Some system limitations were discussed, such as its susceptibility to user errors and the system accuracy being influenced by the data resolution. Future work to improve both the method and the software system is proposed, including an event-based implementation, information display system and notification system. Ultimately, all the identified study objectives were met, delivering the two proposed novel contributions and addressing the identified need for the study.