Due to the fast-growing popularity of the Internet of Things (IoT) in the power distribution system, efficient and effective fault detection is another urgent issue of power network maintenance and effective functioning of the power network. High-dimensionality of sensors, inaccurate fault localization and slowness of responses are the usual problems of traditional fault diagnosis approaches. To address these limitations, this paper proposes an Improved Dynamic Interactive Neighborhood Optimization (DINO) Algorithm tailored for fault detection in power distribution networks. Enhanced DINO model adds adaptive exploration of neighborhoods and recalibration of feature weights to make convergence faster and with higher accuracy of detection in noisy and dynamic observations. Signal preprocessing, dynamic thresholding, and multi-metric learning constitute the methodology which deals with the cases of transient and permanent faults. Temporal-spatial fault characteristics are retrieved out of real-time data of smart sensors and intelligent electronic devices (IEDs). These features are then analyzed by the superior DINO algorithm using a loop that carries out a fitness test that adapts on the fly neighborhood size and weight vectors. This hybrid optimization approach is associated with a large decrease in the false-detection level and therefore with an improvement in the model robustness to different load condition and topological changes in the network. A significant set of simulation studies shows that the given model performs better in comparison with the available fault detection models regarding the speed of convergence, accuracy, computational cost, and stability to the variations in the data. Compared with the benchmark methods, i.e., the original DINO and genetic algorithms, and particle swarm optimization, a highest detection accuracy of 96.7 percent, a convergence time reduced by 20 percent, and a response stability improve by 35 percent are observed, the results of comparative evaluations based on simulations and summary tables. The paper also forms its conclusion, which summarizes future work in terms of integrating federated learning and implementing deploys in real-time and deploying cross-domain portability of the proposed approach.