We present and evaluate a Bayesian method ('Surprise Theory') for detecting changes indicative of sensor malfunction within data measured by autonomous sensor networks. Surprise Theory is evaluated under simulated conditions representative of known anomalies to test its detection capability. Real world performance is evaluated by comparing expert classification of anomalies to surprise model classification of anomalies within a dataset comprised of sensor data obtained from a wireless sensor network measuring thermal profiles, chemical variables, and meteorological conditions in northern temperate lakes. Within this two year dataset comprised of a diverse range of sensors and containing many distinct types of sensor malfunctions, 91.5% of the errors classified by experts were correctly classified by Surprise Theory using conservative parameters. We conclude that Surprise Theory has potential uses as a screening tool to help users identify plausible problems in streams of sensor data in ecology.