The operational condition of railway tracks is critical to ensuring the safety and reliability of railway networks, necessitating real time monitoring and predictive modeling for proactive maintenance. Existing data-driven methods have improved prediction capabilities, but challenges such as inefficient multiscale feature extraction and limited model adaptability persist when processing complex, heterogeneous track data from multisource sensors. To address these limitations, this study proposes a Crested Porcupine Optimizer–Bidirectional Temporal Convolutional Networks–Bidirectional Gated Recurrent Unit (CPO–BiTCN–BiGRU), a novel hybrid model unifying temporal feature extraction, bidirectional sequence modeling, and metaheuristic optimization for railway track condition prediction. The model integrates three tightly coupled components: (1) BiTCN with dilated causal convolutions to capture multiscale temporal dependencies in vibration and displacement signals; (2) BiGRUs to model bidirectional contextual relationships, enhancing dynamic adaptability to operational fluctuations; and (3) a CPO, a metaheuristic algorithm inspired by porcupine defense mechanisms, to autonomously optimize hyperparameters (e.g., learning rate, neuron counts, regularization coefficients), reducing manual tuning efforts and minimizing prediction bias. Experimental validation on two independent data sets—a 103-sample data set and an extended 1,800-sample data set from the China National Railway Group—demonstrates the model’s superior performance. Compared with conventional approaches, the CPO–BiTCN–BiGRU achieves a 63.8% reduction in the root mean square error and a 9% improvement in the certainty correlation coefficient on the multisensor fusion-derived track condition index. Of note, the model maintains robust generalizability across diverse operational scenarios, as validated using stratified five-fold cross-validation. This advancement provides a scalable framework for real time decision-making in railway infrastructure management, bridging the gap between predictive maintenance theory and engineering practice.
Fang et al. (Tue,) studied this question.