Accurately estimation of the state of health (SOH) of satellite lithium-ion batteries using in orbit measurable parameters is crucial for the safe and reliable operation of satellites. However, due to the inability to fully simulate the conditions of satellite lithium-ion batteries during in orbit operation during ground testing, there are problems such as insufficient training data, significant differences in data distribution, and model mismatch. Traditional data-driven methods for degradation characterization and SOH estimation are difficult to achieve high precision and reliability. Both performance degradation characterization and SOH estimation should be self-adaptive to different operating conditions. Therefore, this paper proposes a probabilistic SOH estimation method based on the statistical deep learning under different operating conditions. Extracting different health indicators (HI) from measurable parameters during satellite lithium-ion batteries charging process to characterize performance degradation, combining Permutation Entropy with principal component analysis to improve feature recognition ability for different tasks. Furthermore, a Bayesian neural network is applied to infer the SOH of lithium-ion batteries and quantify uncertainty. The uncertainty obtained by integrating empirical model with particle filter algorithms further enhances the adaptability of the proposed method to different operating conditions. The experimental results illustrate that the method proposed in this paper demonstrates good adaptability and universality for SOH estimation of satellite lithium-ion batteries under different operating conditions. The cross-validation test results show that the maximum estimation error is less than 0.01, and the estimation interval coverage of most results is higher than 0.95, indicating that the method has good prospects in spatial application scenarios.