The embodiment of the invention provides a method and device for monitoring the thermal state of a crystallizer of a continuous casting machine. The method comprises the steps of obtaining multiple sets of historical sample data in the continuous casting production process; determining a principal component model of historical sample data; determining a control limit of a first square prediction error statistical value of the principal component model based on the historical sample data; determining a second square prediction error statistical value of the principal component model based on each group of online data; judging whether the second square prediction error statistical value at each moment is greater than the control limit value or not; if yes, obtaining a current group of onlinedata corresponding to the second square prediction error statistical value at the current moment, determining a first contribution value of each principal component in the principal component model to the second square prediction error statistical value, and determining a current principal component generating an abnormal condition according to the first contribution value; and respectively determining a second contribution value of each piece of online sample data in the current group of online data to the current principal component, and determining the current sample data causing abnormalproduction according to the second contribution value.