【摘要】 在工业领域中，六轴机器人的故障监测数据难以收集大量的故障标签数据。传统的智能诊断方法通常依赖于大规模有标签数据的监督学习，但这在实际应用中存在局限。在解决这一问题的同时，针对单一模型特征提取能力不足、分类性能差的问题，结合半监督学习机制与Informer在处理时序数据的优势，提出一种基于半监督学习和概率稀疏注意力的Informer网络架构，实现对少量有标签数据和大量未标签数据的深度学习，以实现对设备故障的精准诊断。对多组真实环境下采集的工业六轴机器人试验数据进行验证，并与CNN、LSTM、GRU 3种深度学习网络对不同故障程度的辨识能力进行比较。结果表明，在无标签数据为100%组的对比实验中所提出方法的故障诊断准确率达到了90%，同时具有更高的分类准确率和更快的收敛速度；在10%标签数据的条件下所提出方法可实现的诊断准确率达到89.7%。更多还原

【Abstract】 In industrial fields, it is difficult to collect large amounts of labeled fault data for 6-axis industrial robots. Traditional intelligent diagnosis methods usually rely on supervised learning with large-scale labeled data, but this has limitations in practical applications. In order to address this problem while tackling the issues of insufficient feature extraction capabilities and poor classification performance of individual models, by combining semi-supervised learning mechanisms with Informer’s advantages in processing time series data, a semi-supervised learning and probabilistic sparse attention-based Informer network architecture model is proposed to achieve deep learning on small amounts of labeled data and large amounts of unlabeled data to realize accurate fault diagnosis. The test data collected in multiple sets of real environments are verified, by comparing with CNN, LSTM and GRU networks on distinguishing different fault severity levels, the proposed method achieved 90% diagnosis accuracy under the 100% unlabeled data setting, with higher classification accuracy and faster convergence;with 10% labeled data, the proposed method attained 89.7% diagnosis accuracy.更多还原