Accurate state of health (SOH) is of great significance for safe operation of Li-ion battery storage systems. Aiming at the shortcomings of the current SOH estimation methods in terms of poor applicability, large computational load and low accuracy, a SOH estimation method for lithium batteries based on improved domain adaptive transfer learning is proposed. First, a new SOH indicator based on time interval for equal charging voltage difference is designed, which can simulate the random constant current charging segments and simplify the input parameters of the SOH estimation model. Second, by introducing adaptive transfer learning and combining the GRU network characteristics, a GRU network based on an improved domain adaptive transfer learning is proposed to reduce the negative transfer and network training load. Again, based on the new SOH indicator and neural network, the SOH estimation is realized. Finally, the proposed estimation method is validated based on the test data of the self-built experimental platform. The verification results show that, compared with the method based on traditional domain adaptive transfer learning, the mean absolute error and root mean square error of the proposed method are reduced by 3.0% and 3.8% respectively when the test current is 0.75 C. A reduction of 5.8% and 4.3% was achieved at a test current of 0.5 C. Compared with the estimation method based on fine-tuned transfer learning, the error is reduced by 22.9% and 17.4% respectively when the test rate is 0.75 C. At a test current of 0.5 C, the reductions are 25.8% and 14.7%, respectively.