Highly reliable mechanical systems can lead to significant losses in the event of failure, and the lack of comprehensive failure data presents challenges for developing techniques such as critical part identification and failure prediction. In light of this, this paper proposes a meta‐action‐based fault prediction method that effectively addresses the issue of limited fault data. Initially, the mechanical system is decomposed utilizing the “Function‐Motion‐Action” (FMA) methodology to derive individual meta‐action units (MAUs). Subsequently, the limited sample of fault data from the mechanical system is combined with processed expert knowledge to construct the corresponding fault propagation‐directed graph. Furthermore, the key MAUs are determined by applying the Decision–Making Trial and Evaluation Laboratory (DEMATEL) method. Last, the degradation data of the key MAUs is acquired by monitoring them, and a non‐homogeneous discrete grey model (DNGM) integrated with an improved BP neural network is proposed to facilitate the fault prediction of MAUs. Using an industrial robot as a case study, the prediction results demonstrate the superiority of the method proposed in this paper over a single gray model and neural network, thereby providing a reliable prediction approach for anticipating the future trends of data‐deficient mechanical systems.