The invention discloses a Markov process-based time series stream data anomaly detection method. The method comprises the following steps of s1, selecting a training data stream; s2, carrying out dimensionality reduction on the training data flow through an LPP algorithm; s3, determining a current mode category through a K-Means clustering algorithm based on an elbow strategy, and performing mode division on the training data stream; s4, constructing a Markov anomaly detection model, and s5, processing to-be-tested data, inputting the processed to-be-tested data into the Markov anomaly detection model, and outputting an anomaly detection result. According to the method, the time series stream data is clustered, and the data is divided into different modes to construct the Markov-based anomaly detection model; in the model, normal conversion between the different modes is identified as concept drift, only mode conversion which cannot occur is identified as anomaly, the distribution rule of the model and the time series flow data is more stable, and the detection accuracy is higher.