Inferential Monitoring of Quality and Process Faults with Supervised Learning
Principal component analysis and unsupervised machine learning models have been applied to process monitoring to analyze variations in process variables and capture the normal region of variability. In practice, however, anomalies detected in process variables alone do not always result in an anomaly in product quality due to corrective actions by human operators and feedback controllers. Concurrent monitoring schemes have been actively studied recently to achieve simultaneous process and quality-relevant monitoring.
In this talk we introduce a new framework of inferential monitoring and diagnosis (IMD) to include quality-relevant monitoring and quality-irrelevant monitoring schemes using supervised learning models such as partial least squares, canonical correlation analysis (CCA), and other machine learning methods. It is related to and complementary to the scheme of inferential sensors or soft sensors. Monitoring indices based on supervised learning models are given to perform inferential monitoring.
Additionally, contribution plots, reconstruction-based contributions, and generalized reconstruction-based contribution methods are adopted for inferential diagnosis. Finally, industrial case studies and the Tennessee Eastman process are used to illustrate the inferential monitoring and diagnosis of quality-relevant and quality-irrelevant faults.
Recent Advances in Multivariate Image Analysis
in the Mineral Processing Industries
Recent developments in deep learning has generated a new impetus in the development of multivariate image analytical methods that are applied in the development of smart sensors, process monitoring and advanced control of process systems. This includes advances in froth image analysis, ore sorting, monitoring of particulate feeds on conveyor belts, state identification of hydrocyclone operations, but are also opening up potentially new avenues in the development of autonomous inspection systems.
Moreover, these approaches are not confined to the processing of optical images, but can also be applied to signal processing in general. In this overview, the latest developments, challenges and future trends in these methods will be considered, with an emphasis on the use of deep learning and emerging applications in the mineral processing industries.
Process Analytics: Transforming Mineral Process Plant Data Into Actionable Insights
The advent of digital revolution has now enabled us with numerous tools that could be leveraged to transform our operational data into actionable insights.
Key opportunities with digitization include better visualization, transparency, integrated planning and execution for value-chain optimization, that results in smarter production, intelligent response to changes in ore, process and equipment conditions, reduce energy and waste along with prevention of asset breakdown, safety and environmental issues.
It is important to realize that these digital tools have limited value, from a metallurgical operational context, if we cannot bring-in the appropriate domain expertise along with getting the basics right. The focus is to ensure that there is adequate depth of different disciplines built into our platforms along with breadth to integrate other disciplines such as geology and mining for an effective Mine-to-Mill integration.
A metal recovery improvement strategy based on optimal Gaudin particle size distribution, air hold up manipulation, integrated grinding-flotation performance monitoring and guidance is discussed. Simultaneous identification of process and equipment constraints enable finding the best overall conditions for metal recovery. Several industrial success stories are presented.
This paper discusses the methodology, findings, and challenges in the ongoing journey of implementing a Metallurgy Analytics platform that evolves from being retrospectively descriptive to anticipatively prescriptive.