Remaining useful life (RUL) prediction is crucial for equipment health assessment and maintenance planning for industrial systems. However, accurate RUL prediction faces significant challenges due to the complexity of industrial processes, which typically involve multiple interacting components, variable operating conditions, diverse fault modes, and significant noise interference. To overcome these limitations, this study proposes a RUL prediction method for extracting implicit degradation information under variable operating conditions. The methodology integrates multi-scale temporal convolutional networks (TCN) with long short-term memory (LSTM) networks to automatically capture degradation features. Additionally, handcrafted features are incorporated to enhance prediction accuracy. A Kolmogorov-Arnold network (KAN) serves as the regression to provide interpretability to the RUL prediction while simultaneously reduce computational costs. The proposed approach is validated using a prognostic benchmarking dataset about ion mill etching (IME) process. Experimental results verify that the proposed model outperforms the state-of-the-art methods in terms of prediction accuracy and robustness