Abstract: With the wide use of lithium-ion batteries, accurately estimating the state of health (SOH) online has become a significant requirement to ensure the safety and reliable operation of batteries. In this paper, a method based on encoder-decoder framework with attention mechanism was proposed to predict SOH of lithium-ion batteries, which combined CNN and GRU and encoded data into a set of sequences containing intrinsic features; and then with the attention mechanism, the decoder completed the final estimation. This algorithm does not need to establish any battery model or too much prior knowledge. It can get accurate predictions for SOH through the voltage and current in a single cycle. In order to adapt to a variety of situations, this paper designed three inputs modes: fixed-length segment of discharging data, fixed-length segment of charging data, and variable-length segment of charging data. The average error of these modes is less than 1% on the test set, which also confirms that the method proposed in this paper has advantages such as short estimation period, high estimation accuracy, and good adaptability.