Abstract: In view of the limitations of the assumptions that all local diagnosers are continuously available, and the observations of events are fixed in advance in the existing methods of fault diagnosis for distributed DES, a novel method of reliable fault diagnosis for distributed stochastic DES (SDES) based on activation strategy of observations is proposed, where each sensor of agents is allowed to be dynamically turned on/off according to the activation strategy of observations. First, a method of reliable codiagnosability of SDES is proposed, ensuring that the performance of fault diagnosis under the activation strategy of observations will not be degraded even when some local diagnostic information is lost. Then, an automaton of reliable codiagnoser is constructed, and the necessary and sufficient condition of reliable codiagnosability of SDES is presented. Moreover, an algorithm for verifying the reliable codiagnosability is proposed to achieve reliable diagnosis of distributed SDES under the activation strategy of observations. Finally, an illustrative example in the automotive braking system is provided. The result shows that the performance of reliable decentralized fault diagnosis can still be achieved even in the case of dynamic changes of event observations and partial loss of local diagnosis information by using the proposed method.