The state of health (SOH) of lithium batteries can characterize their aging status, and accurately estimating SOH is crucial for reliable operation. A GA-SA-BP neural network algorithm is proposed to improve the accuracy of SOH estimation to solve the problem of low convergence efficiency and susceptibility to local optima of BP neural networks optimized by simulated annealing (SA) and genetic algorithms (GA), which cannot reach the global optimum. First, the correlation between various health indicators (HI) and SOH in NASA’s publicly available dataset is analyzed, and the four HI values of lithium battery output voltage, output current, capacity, and equal voltage drop discharge time with higher correlation with SOH is selected as input values for the BP neural network to improve the accuracy of SOH estimation. Secondly, the GA-SA-BP neural network algorithm is proposed to estimate SOH, and the global optimal solution is found by jumping out of the local optimum when trapped in order to further improve the accuracy of SOH estimation. Finally, the results obtained on the NASA lithium battery dataset and lithium battery experimental testing platform indicate that the proposed approach improves the accuracy of SOH estimation compared to traditional BP neural networks, GA-BP neural networks, and SA-BP neural networks, and remains effective even in the absence of some data.