The invention discloses a failure prediction method based on ICA reconstruction, which includes the following steps: step 1, calculating a separative matrix W; step 2, calculating the statistic value I<2>(k), SPE(k) or I<2>e(k) of the real-time data Xnew(k) through adopting the formulas I<2>(k)=S'newd(k)<T> S'newd(k), I<2>e=S'newe(k)<T>*S'newe(k), SPE(k)=(xnew(k)-x'new(k))<T>*(xnew(k)-x'new(k)), S'newd(k)=Wd*xnew(k), and S'newe(k)=We*xnew(k), wherein Wd refers to the matrix formed by the lines expect the first d lines of the separative matrix W, We refers to the matrix formed by the lines except the first d lines of the separative matrix W, and X'new(k)=Q<-1>BdWd*xnew(k), Bd=(WdQ<-1>)<T>, Be=(WeQ<-1>)<T>, and Q refers to a whitening matrix; and step 3, calculating the nuclear density of I<2>(k), SPE(k) or I<2>e(k), and detecting failures as per the control limit. The method provided by the invention solves the problem that the traditional flue gas turbine prediction method can not utilize the multidimensional valid data, takes the multi-channel vibration data into consideration, can be used for directly predicating failures, and improves the prediction accuracy compared with the PCA reconstruction method.