The invention discloses a monitoring data intelligent sampling method based on relevancy analysis, which includes the four key steps: time series data encoding, relevance relationship mining, state transition matrix calculation, and state prediction. According to the method, a monitoring cycle can be dynamically adjusted according to the prediction on future resource usage of a main unit, thereby reducing sampling frequency while resource usage varies stably, and increasing the sampling frequency while the resource usage varies sharply to save computing and storage resources. Compared with the prior art, the method has the advantages that the monitoring cycle can be enlarged and sampling frequency can be decreased while a machine runs stably; when the machine running fluctuates, it is required to decrease the monitoring cycle and increase sampling rate; more meaningful monitoring data are acquired, the quantity of gibberish to be collected is decreased effectively, spending major computing resources to on gibberish acquisition, computing and other processing is avoided, efficiency is improved, and high accuracy is maintained while gibberish collection is reduced.