During operation and maintenance process of port belt conveyor, a great deal of fault texts that contain important entity information have been accumulated. However, the fault texts have a few characteristics such as text entity boundary ambiguity, amount of jargons, and being scarce of entity annotation. As a result, the traditional methods exist the problems of inaccurate entity feature information extraction and low recognition efficiency. In this work we firstly construct more than one thousand belt conveyor fault text dataset. Secondly, the paper proposes the method of entity definition of belt conveyor fault domain. In addition, we present the details of the ALBERT-BiLSTM-SAM-CRF. In this model, A Lite Bidirectional Encoder Representations from Transformers (ALBERT) is applied to convert text into word vectors. Secondly, the bidirectional long short-term memory (BiLSTM) encodes the trained word vectors bidirectionally to acquire long sequence semantic features. Then the entity boundary features are captured by Self-Attention Mechanism (SAM). Finally, prediction labels are then generated based on the annotation rules of the Conditional Random Field (CRF) and the sequence decoding ability. The experiment results on the self-constructed Chinese text dataset of belt conveyor faults indicate that our model has a excellent performance in the field of belt conveyor entity recognition. The precision, recall and F1 reaching 81.53%, 81.91% and 81.72%, respectively.