Treffer: Anomaly Detection in Multivariate Sensor Data Using Forecasting Models : A Data-Driven Approach for Detecting Irregularities in Multivariate E-Scooter Sensor Data ; Avvikelsedetektering i multivariata sensordata med hjälp av prognosmodeller : En datadriven metod för att upptäcka oregelbundenheter i multidimensionel sensordata för elscootrar

As shared micromobility systems become more prevalent in urban environments, so does the need for robust monitoring tools capable of identifying irregular or unsafe vehicle behavior. This thesis explores how forecastingbased anomaly detection methods can be applied to multivariate sensor data collected from electric scooters. Unlike rule-based systems, which are often uncertain in the face of noisy and complex real-world data, forecastingbased approaches learn the normal signal dynamics from historical data and flag deviations as potential anomalies. The study is motivated by the operational challenges faced by micromobility providers, including crash detection, technical faults, and risky rider behavior, all of which are difficult to detect manually or through threshold-based systems. The work was conducted in collaboration with a e-scooter operator and uses real-world ride data recorded at high temporal resolution. A modular detection pipeline was implemented using Python and the Darts time series library, supporting model training, forecasting, scoring, and evaluation. Three forecasting modelswere evaluated: MovingAverage (as a baseline), XGBoost, and the Temporal Fusion Transformer (TFT). These models were trained on normal ride data and evaluated on datasets that contain confirmed events such as downforce, hardware failures, and accidents. A novel precision evaluation using synchronized camera footage was also carried out to validate the model output against the observed rider behavior. The results show that the TFT model performed the best in terms of interpretability and precision, with approximately 52% of flagged anomalies corresponding to verifiable events. XGBoost also showed promising results with slightly lower precision. The study demonstrates that forecastingbased anomaly detection is a viable and scalable approach to unsupervised monitoring in micromobility systems. It also highlights challenges in thresholding and false positive control, suggesting directions for future work in adaptive scoring ...