The power equipment defect text is important data generated by the operation and maintenance of the power grid system. It is characterized by having many professional words from the power field and high complexity. During model training

there are some problems

such as difficulty in semantic understanding

gradient disappearance

negative information loss

and data imbalance

which hurt the text quality and defect analysis effect. To deal with these issues

this paper proposes a ULF-BI-LSTM text quality improvement algorithm integrating UCNN

LeakyRelu activation function

and Focal Loss function. Then

we correctly separate professional vocabulary

delete invalid vocabulary

and normalize the object description by text data preprocessing. Finally

we fill in missed data and correct the wrong data using the ULF-BI-LSTM algorithm. Experimental results show three improvement strategies that effectively improve the accuracy

precision

F1

and recall of the algorithm. The proposed algorithm outperforms the mainstream algorithm TextCNN

SVM

and BI-LSTM

which alleviates the above problems.