An Adaptive Learning Approach to Multivariate Time Forecasting in Industrial Processes

Industrial processes generate a massive amount of monitoring data that can be exploited to uncover hidden time losses in the system, leading to enhanced accuracy of maintenance policies and, consequently, increasing the effectiveness of the equipment. In this work, we propose a method for one-step probabilistic multivariate forecasting of time variables based on a Hidden Markov Model with covariates (IO-HMM). These covariates account for the correlation of the predicted variables with their past values and additional process measurements by means of a discrete model and a continuous model. The probabilities of the former are updated using Bayesian principles, while the parameter estimates for the latter are recursively computed through an adaptive algorithm that also admits a Bayesian interpretation. This approach permits the integration of new samples into the estimation of unknown parameters, computationally improving the efficiency of the process. We evaluate the performance of the method using a real data set obtained from a company of a particular sector; however, it is a versatile technique applicable to any other data set. The results show a consistent improvement over a persistence model, which assumes that future values are the same as current values, and more importantly, over univariate versions of our model.