To address the limitations of conventional deep-learning-based fault diagnosis methods in complex pattern recognition tasks, particularly insufficient feature extraction and suboptimal parameter configuration, this paper proposes a novel IDCS-SSA-BiTCN framework. The proposed method integrates adaptive signal processing with deep learning in a unified diagnosis architecture. In the signal enhancement stage, an improved singular spectrum analysis (SSA) method is employed, in which the embedding dimension is optimized using the IDCS algorithm and the decomposed components are reconstructed according to a correlation-coefficient matrix, thereby achieving adaptive feature enhancement and noise suppression. In the modeling stage, a bidirectional temporal convolutional network (BiTCN) is constructed, and its key hyperparameters are globally optimized by IDCS to improve fault recognition accuracy and generalization capability across different fault categories. Furthermore, to enhance the stability and convergence efficiency of high-dimensional optimization, the original differential creative search (DCS) algorithm is improved by incorporating a nonlinear parameter control strategy and a Lévy flight mechanism. Experiments conducted on the rolling bearing datasets from Xi'an Jiaotong University and Southeast University demonstrate that the proposed method achieves accuracies of 99.20% and 98.75%, respectively, indicating excellent fault-classification performance.