AbstractThe existing machine learning (ML) models for vibration-based damage assessments often use highly compressed features or are restricted to preprocessed signals with fixed durations and sampling rates. Additionally, the learning capacities of ML models and computational resources are limited, which restricts using raw signals as direct input. This paper studied Mel filter banks (MFBs) for seismic signal processing, inspired by speech recognition technology. It is argued that the same filter designs in audio engineering may not be appropriate for seismic records, and therefore, a customized filter bank formulation was developed. Hybrid deep learning models for rapid assessments (HyDRA) were introduced as multibranch neural network architectures that enable end-to-end training for different types of processed vibration data structures. Moreover, the performance-based earthquake engineering (PBEE) equation was adjusted to integrate ML model uncertainties for probabilistic assessments. The proposed concepts were validated in a case study based on a data set of 32,400 nonlinear time-history analyses of a highway bridge in California. Insights and guidelines are provided for optimum filter design based on 5,184 experiments. Several HyDRA architectures were compared with benchmark models. The optimized MFB feature type outperformed features obtained from continuous wavelet transform and a stacked vector of conventional earthquake engineering indexes. A Bayesian variant of HyDRA was investigated to showcase its integration in the modified PBEE equation. Adopting custom filter banks with the HyDRA architecture enables effective feature extraction from raw vibration records by diversifying feature space and preserving information in the time and frequency domains.