Detecting Anomalies in Bus Passenger Data with Rules and Machine Learning

Sammanfattning

Public transport planning and revenue allocation depend on accurate passengercounts, but raw data from Automatic Passenger Counting (APC) sensors often contains quality issues that simple thresholds cannot detect. This thesis developed andevaluated a prototype that combines three deterministic rules with an unsupervised Isolation Forest detector to identify anomalies in raw APC stop-event datafrom a Swedish bus operator network. The detector was validated through synthetic anomaly injection and through manual review of the highest-scoring eventsthat no rule flagged.Applied to one month of data containing almost 1.6 million stop events, thepipeline flagged 78,044 anomalous events. Manual review of the 50 highest-scoringnovel events confirmed 44 as sensor faults and zero as false positives. The rulebased and machine-learning approaches were found to be complementary: rulesdetect physically impossible values and clear extremes, while the detector identifies events that are anomalous only in the context of their trip