A chemical reactor can drift toward poor quality or danger through changes that are hard to see in time. This study found that a computer could sort several kinds of reactor trouble with remarkable accuracy—but the real plant test still lies ahead.

Batch reactors are commonly used in pharmaceutical and chemical industries to make products in separate batches. But faults and unusual behavior can lead to lower yield, changes in product quality, or hazardous conditions. Early detection and classification of these faults are essential for timely intervention and effective process control.

The reactor data was used for this purpose. The difficulty is that the warning signs are not always simple. The process contains strong relationships among its measurements, complex data, and patterns that are not straight lines. So the study introduced a way to categorize internal and external issues, including changes in reactor temperature, coolant temperature, and jacket temperature, using data from experiments.

The system began with four measurements that were originally free of errors, along with five categories representing different fault conditions: no error, or an error in the flow control, coolant temperature, reactor temperature, or jacket temperature. The data was then used to illustrate and describe how those error categories were distributed.

A straight decision boundary separates the two groups, while the dashed margins mark the widest safe gap between them. The circled points are the support vectors—the cases closest to that boundary—which determine where the separation is placed. Because the data were nonlinear, the study used a nonlinear multikernel SVM approach to compare the measurements rather than relying on straight-line relationships.

It also optimized the radial bias function kernel and key hyperparameters through a grid search, identifying a combination that significantly improved model performance. The resulting approach improved accuracy on nonlinear data, which the study described as different from other available methods.

The nonlinear approach captured the intricacies of the data and produced a model capable of accurately classifying faults. Searching for the best settings was key to finding the strongest version of the model. The chosen settings produced an accuracy of 98.33 percent.

The study inferred in the data that multiclass nonlinear kernels were well suited to this model and its several fault categories. Across each target class, the proposed model achieved performance of at least ninety-six percent on the defined parameters reported by the study.

Finally, the selected radial basis function kernel performed better on the defined parameters than the other kernel methods tested in the reported comparison. Integration with the pilot-plant batch reactor has yet to be tested, including its use with the closed-loop experimental setup.

Live experiments still need to verify whether the model can find other randomly generated faults during real-time experimentation. The model has not yet been tested on the pilot-plant reactor together with the plant’s control systems. The time-based data also has not yet been tested with a deep-learning model.

The developed SVM Python code is still being downloaded onto the Jetson Orin eight-gigabyte board for closed-loop experimental validation and online fault classification. Integration of the proposed SVM model on the pilot-plant batch-reactor setup has yet to be tested experimentally, so live validation remains incomplete.

The main finding is that a nonlinear computer model classified reactor faults accurately, reaching 98.33 percent in the reported evaluation. That could support faster intervention, but it is not yet proof of live, plant-floor performance.