Particle accelerators, used in foundational research and cancer treatment, are complex machinery comprising many different components. In this work, the ion source CAPRICE ECRIS is examined with 48Ca operation. The most common problem with 48Ca at the ECR ion source is the instability of the created plasma inside the source, which leads to increased material consumption and lower quality of the resulting beam. Beamtime data of 2020 and 2021, which consists of multiple device settings and readings, was accumulated and labeled with normal or anomalous state accordingly. For automatically detecting plasma instabilities, 1D-convolutional neural networks are investigated to perform time series classification. The results show the effectiveness of convolutions, which leads to sensitivity of 0.74 and specificity of 0.79. A visual evaluation of the prediction shows good detection of longer anomalous sequences, but the model struggles with smaller anomalies. Given the small nature of the dataset, more data is needed to improve classification performance. Furthermore, better metrics for anomaly detection with time series have to be investigated to perform high-quality evaluations.