The invention relates to a fault detection method for a noisy network. The method comprises the following steps of setting a number alpha of network nodes which have faults simultaneously according to the probability p of fault occurrence of N nodes in the network; detecting each path of the nodes, determining suspected fault thresholds of the nodes according to the probability of detection success and the probability of detection failure, comparing the suspected fault threshold of each node with a set target threshold, selecting the nodes as suspected nodes and forming a suspected node set; updating the suspected node set; and repeating the steps for the new suspected node set until the number of the selected fault nodes reaches the set target threshold, namely completing fault diagnosis. According to the method, the number of the fault nodes can be flexibly set according to the actual network condition and the own need, and computation complexity and accuracy in fault diagnosis are timely adjusted.