Congestion detection is one of the key steps to reduce delays and associated costs in traffic management. With the increasing usage of GPS base navigation, promising speed data is now available. This study utilizes such extensive historical probe data to detect spatiotemporal congestion by mining historical speed data. The detected congestion were further classified as Recurrent and Non Recurrent Congestion (RC, NRC). This paper presents a big data driven expert system for identifying both recurrent and non-recurrent congestion and analyzing the delay and cost associated with them. For this purpose, first normal and anomalous days were classified based on travel rate distribution. Later, we utilized Bayesian change point detection to segment speed signal and detect temporal congestion. Finally according to the type of congestion summary statistics and performance measures including (delays, delay cost, and congestion hours) were analyzed. In this study, a statistical big data mining methodology is developed and the robustness of the proposed methodology is tested on probe data for 2016 calendar year, in Des Moines region, Iowa, US. The proposed framework is self adaptive because it does not rely on additional information for detecting spatio-temporal congestion. Therefore, it addresses the limits of prior work in NRC detection. The optimum value for congestion percentage threshold is identified by Elbow cut off method and speed values were temporally denoised