The assurance of the diagnosis obtained from a nuclear power plant (NPP) fault-diagnostic advisor based on artificial neural networks (ANNs) is essential for the practical implementation of the advisor to transient detection and identification. The objectives of this study are to develop a validation and verification technique suitable for ANNs and apply it to the fault-diagnostic advisor. The validation and verification is realized by estimating error bounds on the advisor's diagnoses. The two different partition criteria are developed to create computationally effective partitions for generating the error information associated with the advisor performance. The bootstap partition criterion (BPC) and the modified bootstap partition criterion (MBPC) can alleviate the computational requirements significantly. In addition, a new error-bound prediction scheme called error estimation by series association (EESA) is constructed not only to infer error-bounds but also to alleviate the training complexity of an error predictor neural network. The EESA scheme is applied to validate the outputs of the ANNs modeled for a simple nonlinear mapping and more complicated NPP fault-diagnostic problems. Two independent sets of data simulated at San Onofre Nuclear Generating Station, a pressurized water reactor, and Duane Arnold Energy Center, a boiling water reactor, are used to design the fault-diagnostic advisor systems and to perform the reliability assessment of the advisor systems. The results of this research show that the fault-diagnostic systems developed using ANNs with EESA are effective at producing proper diagnoses with predicted error even when degraded by noise. In general, EESA can also be used to verify an ANN system by indicating that the ANN system requires training on more data in order to increase generalization. The EESA scheme developed in this study can be implemented to any ANN system regardless of ANN learning paradigm.