Monitoring of metallurgical plant performance with Bayesian change point detection algorithms

Abstract

Production data on mineral processing plants often show great variability, owing to measurement errors, changes in plant performance related to changes in ore characteristics or operating conditions, etc. This makes it difficult to assess plant performance where new equipment or reagents are being evaluated, or where the focus is on the detection and identification of process faults that may be occurring, since small systematic changes in the process can easily be obscured by the large natural variation in the data. This paper explores a Bayesian method for the detection of sudden changes in the underlying parameters of time series data representing key performance indicators on mineral processing plants. The problem is cast as a hidden Markov model, where change point locationscorrespond to unobserved states, which grow in number with the number of observations. Rather than optimising a likelihood function of model parameters, the Baum-Welch algorithm was adapted tomaximise a bound on the log marginal likelihood with respect to prior hyperparameters. This empirical Bayesian approach allows scale-invariance, and can be viewed as an expectation maximisationalgorithm for hyperparameter optimisation in conjugate exponential models with latent variables. Judicious selection of change point priors allow for fast recursive computations on a graphical model that aids interpretation of the results by plant operators. The efficacy of the approach is demonstrated on a number of real-world case studies.