This paper studies the fault detection of discrete-time stochastic systems with linear-time temporal logic (LTL) as correctness requirement-A fault is a violation of LTL specification. The temporal logic allows system correctness properties to be specified compactly and in a user-friendly manner (being close to natural-languages), and supports automatic translation into other formal models such as automata. We introduce the notion of input-output stochastic hybrid automaton (I/O-SHA) and show that the refinement of a continuous physical system (modeled as stochastic difference equations) against a certain class of LTL correctness requirement can be modeled as an I/O-SHA. The refinement preserves the behaviors of the physical system and also captures requirement-violation as a reachability property. Probability distribution over the discrete locations of hybrid system is estimated recursively by computing the distributions for continuous variables for each discrete location. This is then used to compute the likelihood of fault, a statistic that we employ for the purpose of fault detection. The performance of the detection scheme is measured in terms of false alarm (FA) and missed detection (MD) rates, and the condition for the existence of a detector to achieve any desired rates of FA and MD is captured in form of Stochastic-Diagnosability, a notion that we introduce in this paper for stochastic hybrid systems. The proposed method of fault detection is illustrated by a practical example.Note to Practitioners-Many cyberphysical systems, such as building automation systems, automotive vehicles and smart power grids, can be modeled as stochastic systems with mixed continuous and discrete dynamics subject to disturbance and noise, whose behaviors are monitored and controlled by networked (digital) control systems. This paper investigates fault detection in the form of temporal logic specification violation in model-based approach, by transforming it into a state estimation problem for stochastic system models. We provide an algorithm for online fault detection and introduce the notion of Stochastic-Diagnosability for the existence of a detector with any desired accuracy of detection as measured by false alarms and missed detections. The work is illustrated by a room heating system example.