<p>Accurately estimating the state of health (SOH) of lithium-ion batteries is crucial for ensuring safety, improving economic efficiency, and optimizing battery system operation. However, existing data-driven SOH prediction methods rely on expert experience to extract artificial features. Meanwhile, unsupervised learning-based SOH prediction methods are difficult to capture the long-term dependence of capacity degradation and are prone to overfitting. To address these issues, this paper proposes a novel SOH estimation method based on automatic feature extraction and BiLSTM-SA. In this method, firstly, a Squeeze-and-Excitation block is added to the Temporal Convolutional Network to dynamically adjust the channel weights, and combined with the multi-scale technique to enhance the extraction of battery features, so as to construct an Improved Temporal Convolutional Network (ITCN). Subsequently, the ITCN is fused with an Autoencoder (AE) to form the ITCN-AE structure, and the representative features can be automatically extracted by directly taking the data after wavelet denoising and truncation-alignment operations as the input to the ITCN-AE. Next, the extracted features are fed into a Bidirectional Long Short-Term Memory Network (BiLSTM) to further mine the features, and mapped to the battery health state using a fully connected layer after assigning feature weights through the self-attention (SA) mechanism. Finally, on both the NASA and Oxford lithium-ion battery datasets, the proposed model achieves an average estimation error within 1%. Compared with transformer-based method and the latest LSTM variants, the error is reduced by at least 25.2%.&#xa0;</p>