The reliability of electronic control units is crucial in automotive engineering for ensuring vehicle functionality and safety. Malfunctions in these control units can be detected by analyzing log data. As vehicles are becoming increasingly software-centric, the volume and complexity of text-based log files in these systems are reaching record levels. This poses significant challenges for traditional error detection methods that rely heavily on expert intervention. In this thesis, we examine a transition towards automated, data-driven anomaly detection using deep learning techniques to address these challenges. Key contributions of this thesis include setting up a data pipeline, adapting and testing the two existing deep learning architectures LogBERT and LogLLM, evaluating performance on a real-world dataset from the Dr. Ing. h.c. F. Porsche AG and the two open-source datasets BGL and Thunderbird. Our experiments demonstrate that although LogBERT benefits from significantly shorter training and testing times (less than six hours and five minutes respectively per dataset), it has low Precision values with the best being 0.193. With the best F1 value of 0.322, it is not possible to reliably detect anomalies with LogBERT either on the public data sets BGL and Thunderbird or for the data of the Dr. Ing. h.c. F. Porsche AG. In contrast, LogLLM outperforms LogBERT across all data sets, with larger model variants with 8B instead of 1B parameters showing significant improvements. On the BGL data set, the F1 value increases from 0.419 to 0.912 as a result of this scaling. Despite its superior performance, LogLLM’s extensive training and inference times limit its use for ad-hoc analysis in vehicles. The longest experiment executed exceeds 400 hours for training and over 24 hours for evaluation on the Thunderbird dataset. On the two open-source datasets, we successfully replicate the high F1 scores reported in the original paper, achieving 0.912 for BGL and 0.982 for Thunderbird. However, the performance on real-world data from the Dr. Ing. h.c. F. Porsche AG needs improvement, with the best F1 score reaching 0.351. These insights highlight the critical role of scalable models and high-quality data in advancing the application of deep learning for industrial anomaly detection.