In industrial plants, a multi-agent diagnostic system using machine learning with high discrimination performance to prevent the recurrence of accidents are desired because the damages to the company, industrial world, and humans are serious if an accident occurs. However, machine learning needs to be optimised according to the target, e.g. a plant operating condition diagnosis. Thus, we propose a method that selects process signals to optimise the performance of a diagnostic machine for a plant using factor analysis and decision-tree analysis. A feature of our optimisation method is that it accounts for combinations of process signals. Further, an advantage of our method is that the time to create an optimised diagnostic machine is short. The diagnostic machines need to be updated in a limited period, e.g. plant equipment repair, which changes regular process values. Another advantage is that our method can be applied to various learning machines to improve performance. This advantage allows the designer of the diagnostic system to use the best machine-learning method on each diagnostic machine. This paper reports our methodology, our proposed method, and the experimental results where a diagnostic machine was improved by 5.3% to 98.8% from 93.5% abnormality detection accuracy when our method which implemented a support vector machine was applied to the diagnostic machine to detect the blockage of a pipe in the desulfurisation process in a chemical plant simulator.