Three-dimensional fluorescence technology is attracting attention in detecting emergency drinking water pollution events. However, some unsolved problems remain, such as being easily affected by water environment fluctuations, low detection rate facing low-concentration organic pollutants, etc. Therefore, in response to the demand for online monitoring, this study proposed a time series double thresholds method for anomaly detection in drinking water using three-dimensional fluorescence. This method applied principal component analysis (PCA) to extract the feature spectrum of the detected samples and trained the linear autoregressive (AR) model to predict the principal component of the water samples in the future. The eigenvalue difference was then obtained by comparing the predicted and measured ones. At the same time, combined with the change rate of the measured eigenvalues, the double threshold for time series was set to finally determine the start and end points of the pollution event to determine the entire pollution event. The research validated the proposed method by simulating high-concentration pollution events, low-concentration pollution events, and fluctuations in water background. The experimental results show that this method maintains the detection accuracy for high-concentration pollution events. Moreover, compared with conventional methods, the proposed method improved the detection performance in low-concentration pollution events and low-concentration pollution in high-interference environments. The detection accuracy is increased by 9.4% and 20.7%, respectively.