Abstract: In order to satisfy the outlier detection requirements in one kind of high-speed, small-variance unlabeled industrial time series, an online outlier detection and confidence estimation algorithm based on Bayesian posterior ratio was proposed. The algorithm combined prediction and hypothesis testing, establishing the autoregressive model firstly and then using Bayesian posterior logarithm of residuals to identify outliers. To reduce misjudgment, the state transition probabilities were calculated by self-organizing map neural network and the reliability of detected outliers was evaluated afterwards. It updated models periodically to dynamically adapt to data changes, thus improving accuracy. Experimental results demonstrate that the online algorithm can effectively detect outliers in time series provide reliable confidence evaluation, bringing higher adaptability and practicability.