Unmanned Aerial Vehicles (UAVs), commonly known as drones, are increasingly deployed across diverse application domains, raising critical challenges for digital forensic investigation following safety incidents and system failures. In drone investigations, systematic analysis of flight logs is essential for reconstructing events, identifying root causes, and supporting reliable incident attribution and risk mitigation. Because a message may contain multiple sentences, message-level analysis cannot precisely pinpoint which log segment indicates a problem. Therefore, this paper proposes DroPTC ( Dro ne P roblem T ype C lassifier), an end-to-end framework to identify and classify problems at the sentence level. A rule-based segmenter is designed to segment log messages into sentences based on historical log characteristics. Using the resulting log sentences, a pre-trained embedding is fine-tuned using contrastive learning for semantic alignment. The integrated gradient is employed to enhance the model's interpretability, enabling admissible and trustworthy analysis. Sentence deduplication is utilized to identify unique log events, thereby reducing the analyst workload. Quantitative and qualitative analysis of the experimental results show that DroPTC outperforms the baselines in three aspects: performance, trustworthiness, and efficiency. This paper also presents a working open-source tool as the tested implementation of the proposed framework. The tool accepts the decrypted flight log file and produces a forensic report in HTML and PDF format.
Silalahi et al. (Sun,) studied this question.