Extreme learning machine (ELM) has better operation efficiency in fault diagnosis. However, the recognition accuracy of ELM algorithm is actually affected by the activation function. Moreover, most of the testing dataset are coming from high precision and expensive sensors. In this paper, raw data are collected by a low-cost attitude sensor, which is installed on the mobile platform of a delta 3D printer. A doublet activation function is proposed to improve the performance of ELM, named doublet ELM (DELM). The proposed method is evaluated using experimental data collected from the 3D printer, and its advantages are demonstrated by comparing with other activation functions. The experimental results indicate that the proposed method leads to the highest accuracy in different hidden nodes and the testing classification rate achieves 93% and 96% using only 8.33% of the dataset for model training, for R75 and R90 sub-datasets, respectively. Moreover, compared with peer methods, such as random forest, echo state network, and so on, the results show that the present DELM exhibits the best performance in small-sample and improves the accuracy of the 3D printer fault diagnosis.

极限学习机（ELM）在故障诊断中具有更好的运行效率。但是，ELM算法的识别精度实际上受激活函数的影响。而且，大多数测试数据集都来自高精度和昂贵的传感器。在本文中，原始数据由低成本的姿态传感器收集，该传感器安装在增量3D打印机的移动平台上。为了提高ELM的性能，提出了一个双峰激活函数，称为doublet ELM（DELM）。使用从3D打印机收集的实验数据对提出的方法进行了评估，并通过与其他激活功能进行比较证明了其优势。实验结果表明，对于R75和R90子数据集，仅使用8.33％的模型训练数据集，所提出的方法在不同的隐藏节点中的准确性最高，并且测试分类率分别达到93％和96％。此外，与诸如随机森林，回波状态网络等对等方法相比，结果表明，当前的DELM在小样本中表现出最佳性能，并提高了3D打印机故障诊断的准确性。