Automatic identification of microseismic (MS) signals is crucial for early disaster warning in deep underground engineering. However, three major challenges remain for practical deployment, namely limited resources, severe noise interference, and data scarcity. To address these issues, this study proposes the lightweight and robust entropy-regularized unsupervised domain adaptation framework (LRE-UDAF) for cross-domain MS signal classification. The framework comprises a lightweight and robust feature extractor and an unsupervised domain adaptation (UDA) module utilizing a bi-classifier disparity metric and entropy regularization. The feature extractor derives high-level representations from the preprocessed signals, which are subsequently fed into two classifiers to predict class probability. Through three-stage adversarial learning, the feature extractor and classifiers progressively align the distributions of the source and target domains, facilitating knowledge transfer from the labeled source to the unlabeled target domain. Source-domain experiments reveal that the feature extractor achieves high effectiveness, with a classification accuracy of up to 97.7%. Moreover, LRE-UDAF outperforms prevalent industry networks in terms of its lightweight design and robustness. Cross-domain experiments indicate that the proposed UDA method effectively mitigates domain shift with minimal unlabeled signals. Ablation and comparative experiments further validate the design effectiveness of the feature extractor and UDA modules. This framework presents an efficient solution for resource-constrained, noise-prone, and data-scarce environments in deep underground engineering, offering significant promise for practical implementations in early disaster warning.