Intelligent perception and control of fuel injection characteristics is a key frontier technology for achieving engine controllable and efficient combustion. To accurately obtain reliable real-time information on fuel injection characteristics, a fuel injection characteristic perception method that integrated variational Bayesian and bidirectional long short-term memory networks (VB-BiLSTM) was proposed. By the collaborative optimization of variational inference and adaptive moment estimation algorithm, the optimal variational approximation of hyperparameter probability density distribution was determined, and the determination of credible interval for fuel injection rate and the precise perception was achieved. The perception performance was evaluated under different levels of noise interference using different typical deep learning models as controls, and online perception verification of fuel injection characteristics was conducted on a test bench. The research results show that the VB-BiLSTM model has significantly better noise resistance performance compared to other models. The experimental data of fuel injection rate are within the 95% credible interval of perceived injection rate, and the maximum error between the perceived and actual value is less than 6%. The end-to-end perception time is on the order of milliseconds.