Abstract:  Combine harvesters as a kind of mechanized equipment will inevitably have mechanical failure, in order to quickly find out the relevant fault entity and solve the mechanical failure, a named entity recognition model RP-TEBC (RoBERTa-wwm-ext+PGD+Transformer-Encoder+BiGRU+CRF) for combine harvester fault field is proposed. RP-TEBC uses the dynamically encoded RoBERTa-wwm-ext pre-trained model as the word embedding layer, uses the adaptive Transformer encoder layer to fuse the Bidirectional Gating Unit (BiGRU) as the context encoder, and finally uses the conditional random field (CRF) as the decoder layer, using the Viterbi algorithm to find the optimal path output. At the same time, the RP-TEBC model generates adversarial samples by adding some perturbations in the word embedding layer. Through continuous training and optimization of the model, the overall robustness and generalization performance of the model can be improved. On the constructed named entity recognition data set in the field of combine harvester faults, experiments have shown that compared with the baseline model, the accuracy, recall rate, and F1 value of this model have increased by 1.79%, 1.01%, and 1.46% respectively.

Key words: combine harvester, fault domain, named entity recognition, knowledge graph, pre?trained model, adversarial sample

摘要： 联合收割机作为一种机械化设备不可避免地会出现机械故障，为快速地找出并解决机械故障，提出一种面向联合收割机故障领域的命名实体识别模型RP-TEBC（RoBERTa-wwm-ext+PGD+Transformer-Encoder+BiGRU+CRF）。RP-TEBC使用动态编码的RoBERTa-wwm-ext预训练模型作为词嵌入层，利用自适应Transformer编码器层融合双向门控单元（BiGRU）作为上下文编码器，利用条件随机场（CRF）作为解码层，使用维特比算法找出最优的路径输出。同时，RP-TEBC模型在词嵌入层中通过添加一些扰动，生成对抗样本，经过对模型不断的训练优化，可以提高模型整体的鲁棒性和泛化性能。结果表明，在构建的联合收割机故障领域命名实体识别数据集上，相比于基线模型，该模型的准确率、召回率、F1值分别提高1.79%、1.01%、1.46%。

关键词: 联合收割机, 故障领域, 命名实体识别, 知识图谱, 预训练模型, 对抗样本