The assessment of state of health (SOH) of batteries is one of the key technologies in battery systems, and its accuracy is crucial for the safe operation of battery systems. The relaxation voltage curve contains rich battery information and has a short relaxation time, making it suitable for evaluating the state of health of batteries under non constant operating conditions. This article uses the relaxation voltage curve to evaluate the state of health of lithium iron phosphate battery modules. Firstly, a relaxation voltage model for lithium iron phosphate (LFP) battery modules based on linear correlation between time constant and relaxation time was established, and particle swarm optimization (PSO) algorithm was used to identify the parameters of the relaxation voltage curve and extract health factors. Secondly, a hybrid model of convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) optimized based on pelican optimization algorithm (POA) was developed to evaluate the SOH of batteries. The research results show that regardless of whether the relaxation voltage curve is obtained by charging and discharging at 1/2 C rate or 1 C rate, the relative error (RE) between the variable time constant voltage value identified by PSO algorithm parameters and the true relaxation voltage value does not exceed ±0.12%, indicating that PSO method has good parameter identification effect on relaxation voltage at different rates. Using the relaxation voltage curve after charging and discharging at a rate of 1/2 C, the relative error of SOH evaluation obtained using the POA-CNN-BiLSTM model in the test set still does not exceed ±1.2% even when the training set is as low as 5%, At a charge discharge rate of 1 C, when the training set was as low as 5%, the relative error of SOH evaluation obtained using the POA-CNN-BiLSTM model in the test set still did not exceed ±1.5%, indicating that the POA-CNN-BiLSTM model has high accuracy in evaluating battery SOH.