The parameters redundancy,complex structure and large-scale of the multi-task learning (MTL) network lead to poor real-time performance;unable to obtain some or all kinds of fault samples of the component results in a zero-shot problem.Aiming at the above problems,a lightweight MTL network optimized by meta-learning was proposed.MobileNetV3 lightweight network was used to construct the lightweight MTL network that has multiple sub-task networks for improving the real-time performance.The zero-shot problem based on across-components was studied,model-agnostic meta-learning (MAML) was applied to optimize the training mechanism of the lightweight MTL network,the lightweight MTL network optimized by MAML was constructed.Finally,the fault diagnosis performance of the proposed network was tested from the perspective of different fine-tuning steps and testing tasks.Based on the measured faults of multiple components of the gear and rolling bearing,it can be known that the proposed method can solve the problem of multi-task fault diagnosis and cross-component zero-sample in real-time and high accuracy.