To meet voltage and capability needs, batteries are grouped into packs as power sources. Abnormal ones in a pack will lead to partial heating and reduced available life, so removing anomalies out during manufacturing is of great significance. The conventional methods to detect abnormal batteries mainly rely on grading systems and manual operations. Current data-driven methods use statistical, machine learning and neural network approaches, building models, then applying them on the unlabeled. However, both cannot make full use of multiple source data and expert knowledge. Therefore, how to use these multi-source data and knowledge to improve the effect of battery anomaly detection process has become a research focus. We put forward a data-driven multi-source data feature fusion and expert knowledge integration (FFEKI) network architecture that follows encoder-decoder structure with multiple integration units and a corresponding joint loss function. First, we collect multi-source data and obtain fusion features. Then, we refine filters from expert knowledge and transform them into neural network layers as components of integration units. By this way, supervisory knowledge is integrated into our network. We evaluate our scheme by sets of experiments comparing with most widely used approaches on real manufacturing data. Results show that FFEKI obtains a maximum 100% anomaly detection rate (ADR). Meanwhile, when the number of detection T is greater than the actual number of anomalies in the testing set, our method can achieve full ADR faster. It is concluded that the proposed FFEKI achieves effective performance on power lithium-ion battery anomaly detection.