In industrial equipment fault diagnosis, few-shot learning offers promise for scenarios with limited data, but traditional methods often overlook causal relationships, undermining prediction reliability. To address this, we propose the causal few-shot relation measurement network (CFRNet), which integrates causal reasoning into few-shot learning. The CFRNet architecture includes a feature extractor, a causal intervention module, a causal decomposition module, and a relationship measurement module, allowing the separation of causal and non-causal elements in data. By using a causal decomposition loss function, CFRNet reconstructs causal structures and represents them as independent features. The feature extractor and relationship measurement module work together to establish a learnable metric space, improving generalization and classification accuracy. Experimental results on public and proprietary datasets demonstrate CFRNet’s superior performance and greater interpretability, setting a new benchmark for causal reasoning in mechanical fault diagnosis.