Abstract: The intelligentization of mechanical equipment in fully mechanized mining face is an attractive field, and abnormal detection of mechanical equipment, as an indispensable part, has broad application scenarios. However, the following problems are faced in practical applications. One is that unexpected anomalies encountered in the actual environment bring challenges to data labeling; the other is the imbalance of data categories caused by too little abnormal information. In this paper, a shearer anomaly detection method based on self-encoding twin neural network is proposed, and a weak label data set is constructed by combining shearer working condition information to solve the first problem. For the second problem, build the LCAS (LSTM-CNN Auto-Encoder Siamese) neural network, reduce the impact of class imbalance on training through the twin neural network, and combine the LSTM and CNN auto-encoder neural network for feature extraction to improve model accuracy when diagnosing time series data. The test results of the model show that the model can effectively extract high-quality features, has good robustness against unbalanced data, and the model has a certain generalization ability, which is effective and practical.