To address the issues of artificial intelligence (AI) model bias and insufficient feature learning caused by imbalanced optical network datasets, this paper proposes a fault identification algorithm for optical line terminal (OLT) equipment based on deep cross network (DCN) and multi-task learning (MTL). First, potential faults are assessed using standardized mean Manhattan distance, with high-similarity samples labeled as poor-quality data. Subsequently, a DCN-MTL model is constructed, incorporating high-order feature interactions to enhance learning capability, while utilizing poor-quality detection as an auxiliary task to optimize the training of the primary fault detection task. Experimental results demonstrate that, compared to traditional deep neural networks, the proposed algorithm achieves improvements of 1.15% in accuracy, 11.83% in recall, 6.39% in F1-score, and 5.91% in area under the curve (AUC) under the same data volume, with all metrics exceeding 0.95. This validates the algorithm's strong detection capability in scenarios with scarce fault data.