基于改进多任务学习网络的零样本故障诊断

Zero-Shot Fault Diagnosis Based on Improved Multi-task Learning Network

多任务学习网络结构和参数冗余、网络规模过大,导致网络实时性差的问题;无法获取元件的部分或者全部故障类型样本,导致零样本问题。针对上述问题,提出一种基于元学习优化的轻量化多任务学习网络。为了提高实时性,利用MobileNetV3轻量化网络构建具有多个子任务诊断网络的轻量化多任务学习网络模型;研究了跨元件零样本问题,利用模型无关(MAML)元学习方法,对轻量化多任务学习网络的训练方式进行优化,构建基于元学习优化的轻量化多任务学习网络;最后,从不同微调步数和测试任务数角度,测试了所提网络的诊断性能。通过齿轮和轴承多元件的实测故障分析可知,所提方法可以实时高精度地解决多任务故障诊断问题和跨元件零样本问题。

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.Ai-ming 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.