This paper presents a new integrated methodology for fault detection and diagnosis. The methodology is built using the multivariate exponentially weighted moving average principal component analysis (MEWMA-PCA) and the Bayesian network (BN) model. The fault detection is carried out using the MEWMA-PCA; diagnosis is completed utilizing the BN models. A novel supervisory learning-based methodology has been proposed to develop the BNs from the historical fault symptoms. Although the algorithm has been extensively applied to the Tennessee Eastman (TE) chemical process, monitoring of three specific (difficult to observe) faults, IDV 3, IDV 9, and IDV 15, has been demonstrated in this article. Most of the existing data-based methods have faced the challenge to detect these faults with a good detection rate (DR). Hence, these faults have been reported as either unobservable or strenuous to detect. Overall, fault detection performance of the squared prediction error (SPE) statistics combined with the MEWMA-PCA was found to be better than the T-2-based monitoring model. Although the cumulative sum (CUSUM) PCA-based approaches have demonstrated successful detection and diagnosis to these specific faults, the comparative studies suggest that the proposed methodology can outperform the CUSUM PCA approach.