针对回转窑火焰图像细节特征难以区分导致的工况识别困难的问题，提出一种基于改进三元组网络的小数据集回转窑火焰图像工况识别方法。该方法在原始三元组网络的基础上优化了距离度量方式并引入类内距离损失。首先，将回转窑火焰图像的 RGB 通道分离后分别进行双边滤波预处理，保留具有火焰边缘信息的特征，并在原始三元组损失中通过引入类内特征距离损失，使其能够在增大类间距离的同时，减小类内距离；然后，通过基于马氏距离的改进三元组网络获取具有细节差异性的火焰图像特征；最后，采用 K 均值聚类算法对有标签的特征向量进行工况识别。实验结果表明，该方法得到的特征具有更强的判别性，可以有效提高回转窑工况识别的分类精度，指导回转窑操作。

To address the problems of working condition recognition caused by the difficulty in identifying the detailed features of rotary kiln flame images, a method based on improved triplet network with small data set is proposed, in which the distance measurement is optimized and an in-class feature distance loss is introduced based on the original triplet network. Firstly, RGB image channels are separated and bilateral filtering is carried out respectively to retain the flame edge information features during preprocessing. The inter-class feature distance is increased while the in-class feature distance is reduced through the in-class feature distance loss. Then, the detailed flame image features are extracted using the improved triplet network based on Mahalanobis distance. Finally, the labeled feature vectors with different working conditions are recognized using K-means algorithm. The experimental results show that the features extracted by the proposed method are easier to be recognized, thus effectively improving the classification accuracy of rotary kiln working condition recognition, and guiding rotary kiln operation.