Abstract: Reflective middleware is used to detect parameters and states of Web services which cause services unavailability,and the running state and configuration of services are dynamically adjusted internally to avoid Web services unavailability.The HME(Hierarchical Mixtures of Expert neural network) is deployed to the meta-layer of the reflective middleware to detect and deal with the states and parameters which lead Web services unavailability.An expectation-maximization policy is presented as a learning strategy to train the HME.Experiment and analysis are made,and the results show that HME network as a detector in reflective middleware is highly efficient to detect services unavailability.