A fault detection and diagnosis technique for multivariate processes using a PLS-decomposition of the measurement space

•An extended PLS strategy is used to decompose the measurement space into four projections.•Four statistics of non-overlapped domains enable an adequate online monitoring of the process.•A robust fault detection and diagnosis technique is able to discriminate many anomalies.•The diagnosing tool is useful for monitoring processes that exhibit collinear measurements.

A new statistical monitoring technique based on partial least squares (PLS) is proposed for fault detection and diagnosis in multivariate processes that exhibit collinear measurements. A typical PLS regression (PLSR) modeling strategy is first extended by adding the projections of the model outputs to the latent space. Then, a PLS-decomposition of the measurements into four terms that belongs to four different subspaces is derived. In order to online monitor the PLS-projections in each subspace, new specific statistics with non-overlapped domains are combined into a single index able to detect process anomalies. To reach a complete diagnosis, a further decomposition of each statistic was defined as a sum of variable contributions. By adequately processing all this information, the technique is able to: i) detect an anomaly through a single combined index, ii) diagnose the anomaly class from the observed pattern of the four component statistics with respect to their respective confidence intervals, and iii) identify the disturbed variables based on the analysis of the main variable contributions to each of the four subspaces. The effectiveness observed in the simulated examples suggests the potential application of this technique to real production systems.