Aiming at the problems of low efficiency, high labor cost, slow response to dynamic regulations and insufficient use of unstructured data in the traditional mode of information compliance audit of power industry expansion, this paper proposes an automatic comparative analysis method based on deep residual network. This method realizes intelligent audit through four technical modules: unified representation of multimodal data, semantic extraction of long text compliance, dynamic compliance rule engine and closed-loop mechanism of man-machine collaboration. In terms of technical realization, a Hybrid-ResNet dual-channel depth residual network is constructed, and the text branch uses depth-separable convolution residual block and hierarchical gradient clipping technology to process long texts of more than 4,000 words to solve the problem of gradient disappearance. The image branch is based on the ResNet-101 handwriting enhancement module (HEM) to improve the accuracy of handwritten content recognition. The cross modal attention mechanism integrates text and image features to form a unified representation. Design a Recursive Residual Semantic Parser (RRSP) to enhance deep compliant semantic extraction of long texts through recursive mechanisms, residual connections, and domain knowledge weight matrices. Establish a dynamic compliance rule engine, integrate 12 laws and regulations to form mandatory, early warning and explanatory rules, and realize automatic updating of rules combined with Bi-GRU semantic analysis. Conflict detection and incremental updating of knowledge map ensure the timeliness of rules. The closed-loop mechanism of man-machine cooperation takes 95% confidence as the threshold, which realizes efficient cooperation between automatic review and manual review, and manual correction feedback is used for online fine-tuning of the model. The experimental results show that the recognition accuracy of key terms is 95.7%, the recognition accuracy of handwritten terms is 92.7%, the detection rate of hidden violations is 96.3%, and the processing time of a single contract is only 11.8s, which is significantly better than manual review and BERT step-by-step processing method. The dynamic rule engine shortens the response time of policy update from tens of days to several hours, effectively promoting the transformation of compliance audit in power industry from manual comparison to risk judgment.
Zhang et al. (Sun,) studied this question.