A batch reactor can fail through several intertwined temperature and actuator faults, yet this study reports a multikernel SVM reaching 98.33% accuracy. The catch is that real-time pilot-plant validation is still unfinished.

Fault identification matters in process industries because it supports effective monitoring of process and plant safety. Minimizing process downtime is critical for enhancing product quality and minimizing production costs, while real-time categorization of issues is essential for monitoring processes.

The challenge is that robust correlations, complex data, and lack of linearity make fault categorization difficult. The proposed technique uses multikernel support vector machines to categorize internal and external issues involving reactor temperature, coolant temperature, and jacket temperature, using data obtained from empirical research.

Batch-reactor processes are commonly used in pharmaceutical and chemical industries to produce pharmaceuticals, chemicals, and specialty products. These processes are susceptible to faults and anomalies that can cause decreased yield, product-quality deviations, or hazardous conditions.

Early detection and classification of these faults are essential for timely intervention and effective process control. For that purpose, the study used data from the batch reactor shown in Figure 1. Figure one shows the pilot-plant batch reactor setup used in the Machine Learning for Advanced Process Control lab at MIT, Manipal.

The photograph identifies the batch reactor, coolant-flow station, control equipment, and the Jetson Orin eight-gigabyte computing board. This hardware matters because the study uses data from the reactor to support real-time fault identification and classification with machine-learning methods.

The data collected from the batch reactor is nonlinear in nature. Because the batch reactor is a closed-loop experiment where events are tightly controlled and interconnected, methods designed primarily for open-loop experiments may have limited applicability. The study identifies multikernel SVM as the most robust machine-learning method for nonlinear data classification and uses it for fault classification in intricate industrial processes.

SVMs can be used for both linear and nonlinear classifications of data into various categories. The paper separates the discussion into binary classification using a linear SVM and nonlinear SVM methods extended to multiple classes through different strategies.

The SVM kernel trick is designed to deal with nonlinear data. It automatically converts the data into higher dimensions, where a linear hyperplane can separate the classes more effectively. Typical kernel functions include the linear kernel, the polynomial kernel, and the radial basis function, or RBF.

Figure 5 shows two classes of points plotted against the independent feature vectors a1 and a2. The solid line is the SVM decision boundary, while the dashed lines mark the margins; circled points are the support vectors that define this separation.

The visual matters because it illustrates how the classifier assigns samples using the sign of the decision function, with the margin structure determining the fitted boundary. The batch-reactor data was gathered from different sensors, mainly actuator data, coolant temperature, reactor temperature, and jacket temperature.

The recorded data was written to a CSV file, and the data gathered from various sensors was nonlinear in nature. For this data, a linear function cannot accurately model the relationship between variables, and nonlinear patterns may follow curves, circles, parabolas, or other shapes.

Linear models may therefore produce poor predictive performance, while SVMs with a nonlinear kernel can effectively capture nonlinear patterns in pilot-plant batch-reactor data. The preprocessing data set contained four features, originally error-free, and a target variable with five classes representing different fault categories.

The five classes were class zero for no error, class one for an error in Fc, class two for an error in Tc, class three for an error in Tr, and class four for an error in Tj. Figures seven through ten were used to illustrate and characterize the data-set structure and provide insight into the distribution of error classes.

Figure seven compares the Fc signal with an introduced error, shown by the red dashed line, against the error-free Fc signal in blue across the data rows. The plot shows several localized departures between the two traces, while the underlying signal also changes across different temperature levels and curved transitions.

This visualization matters because it illustrates the error class used in the SVM preprocessing stage, where Fc-error samples are distinguished from no-error data. After splitting the data into training and testing sets, GridSearchCV was used to find the best hyperparameters for the SVM model.

A parameter grid explored combinations of hyperparameter values, and GridSearchCV performed cross-validated model training on each combination. The proposed model tuned C, the regularization parameter; gamma, the kernel coefficient for RBF; the kernel type; and the decision-function shape.

The goal was to find the combination that maximizes accuracy in fault classification. Figure ten shows temperature data with introduced errors, plotted against the original temperature series. Figure eleven lists SVM hyperparameter combinations and their corresponding scores, with the displayed top entries using C equal to zero point five, an rbf kernel, and gamma equal to one hundred.

The authors report a score of ninety-eight point three three four four zero three for these configurations, illustrating how the tuned model was evaluated on unseen data. The nonlinear kernel, specifically the RBF, captured the intricacies of the nonlinear data and led to a model capable of accurately classifying faults.

GridSearchCV hyperparameter tuning was key to finding the best set of hyperparameters. The selected values were C equal to zero point five, gamma equal to one hundred, kernel equal to RBF, and decision-function shape equal to one-versus-one.

Those values produced an accuracy of 98.33 percent. Figure twelve shows the multikernel SVM’s confusion matrix using a Gaussian kernel, with true classes on the vertical axis and predicted classes across the horizontal axis. The diagonal cells contain the correctly classified cases, including one point nine times ten to the power of four for No Error, while off-diagonal cells show errors such as ninety-six Tj Error cases predicted as Tr Error.

This makes the model’s fault differentiation and remaining confusions visible at a glance. The selected hyperparameter values produced an accuracy of 98.33 percent, and the data supported the conclusion that multiclass nonlinear kernels are well suited for this model.

Random faults were introduced at time-scale interval points, with the error labeled in the predictive features. The model's performance was evaluated with precision, recall, and F1-score over each target class, and performance was at least 96 percent over each target class.

Performance was also evaluated over each kernel using accuracy, precision, recall, and F1-score, and the proposed RBF kernel performed better than the other kernel methods. Table two reports the SVM model’s precision, recall, and F-one score for each target class: no error, F-sub-c error, T-sub-c error, T-sub-r error, and T-sub-j error.

The listed values range from zero point ninety-six to one point zero zero, showing how classification performance is documented separately across the fault labels. This matters because the authors use these class-level metrics to evaluate the model beyond its reported ninety-eight point three three percent overall accuracy.

Table three compares model performance across four kernels using accuracy, precision, recall, and F1-score. The linear, polynomial, sigmoid, and Gaussian, or RBF, kernels are reported with values for each metric; the Gaussian RBF row shows ninety-eight point ninety accuracy, ninety-seven point seventy precision, ninety-seven point sixty-four recall, and ninety-seven point sixty-five F1-score.

This visual matters because it supports the authors’ choice of a nonlinear RBF kernel for the model. Integration of the proposed SVM model with the pilot-plant batch-reactor setup, together with closed-loop RNN-NMPC and DNN-NMPC for trajectory tracking, has yet to be tested.

Because the data is in a time-series format, the work has yet to be tested with a deep-learning model and have its performance verified. Verification of finding other randomly generated faults is still to be considered during real-time experimentation. Downloading the developed SVM Python code into the Jetson Orin eight-gigabyte board is in progress for closed-loop experimental validation of online fault classification.

The proposed radial basis function SVM classified the nonlinear fault data with 98.33% accuracy, while performance stayed at least 96% across target classes. But closed-loop pilot-plant testing remains future work.