Neural networks have demonstrated excellent performance in facilitating automatic fault detection and diagnosis in many engineering applications. Their primary advantage over model-based and knowledge-based expert systems is that they require very little development time or expertise. However, neural networks perform only as robustly as the data from which they are trained. Therefore, understanding the content and limitations of process data used to train the network is crucial. This paper presents neural networks as part of a fault recognition framework for diagnosing process inefficiencies. In this framework, incorporating a small amount of process knowledge helps minimize data limitations and maximize the neural network’s performance. Data preprocessing and filtering lead to significant improvement in recognition performance and markedly reduced training time. In addition, the framework allows detection of the "unknown" class. The fault recognition framework will be demonstrated via a simulated continuous stirred tank reactor system which operates under realistic disturbances and noisy measurements.