In this paper, we propose a framework for integrating fault diagnosis and incremental knowledge acquisition in connectionist expert systems. A new case solved by the Diagnostic Function is formulated as a new example for the Learning Function to learn incrementally. The Diagnostic Function is composed of a neural networks-based Example Module and a symbolic-based Rule Module. While the Example Module is always first invoked to provide the short-cut solution, the Rule Module provides extensive coverage of cases to handle odd cases when Example Module fails. Two applications based on the proposed framework will also be briefly mentioned.