<p>This paper proposes a new intelligent system for fault diagnosis by the combination of Support Vector Data Description (SVDD) and a hybrid metaheuristic (PSOGSA) for the reference space optimization. The main purpose of the proposed methodology is to simultaneously perform the fault detection and identification of the impact variables for fault diagnosis in an optimal manner. This procedure is carried out by the development of a SVDD classifier where the corresponding hyperparameters and suitable subset of features (feature selection) are constructed by the binary metaheuristic approach. Moreover, a fast algorithm is implemented to solve the SVDD optimization problem (fSVDD) to obtain a robust and fast method. In order to study the performance of the proposed methodology BPSOGSA-SVDD together with its robust and fast version BPSOGSA-fSVDD, a comparison with respect to the standard BPSO is presented, therefore four different approaches (BPSOGSA-fSVDD, BPSOGSA-SVDD, BPSO-fSVDD, and BPSO-SVDD) are considered and applied to a welding automotive process that involve a total of 18 variables. The experimental results showed that the proposed system using the hybrid metaheuristic BPSOGSA in combination with SVDD and fSVDD reach a lower minimum with respect to the BPSO-based approaches. Furthermore, the system can detect the subset of input variables that impact in a realistic manner on the quality defects in order to achieve an efficient fault diagnosis. The effect of the detected variables was tested by the training of several machine learning algorithms for fault detection such as SVM and KNN. In this context it was found that most of the algorithms do not lost their fault detection capabilities, even in several cases the performance metrics such as accuracy, precision and F1 score were enhanced using the feature subset detected by the proposed system. Finally, a nonparametric statistical analysis is also included in order to compare the considered strategies. The numerical results show that the hybrid BPSOGSA-fSVDD converges to lower values of misclassification than other widely used machine learning algorithms as well as hybrid approaches like Sequential Feature Selection and BPSO-SVDD and this is done with computational times even three times shorter than the classical SVDD approach. Finally, it is concluded that the presented approaches in this work are suitable and promising options to be implemented in other industrial processes.</p>