In terms of the non-stationary systems existed in engineering systems, cointegration model can be established only when the system variables are integrated of first order. The cointegration coefficients matrix is proposed as the characteristic parameter of non-stationary system fault diagnosis. System variables have to be verified integrated of first order through unit root test, then use Johansen method to estimate the cointegration coefficients matrix, which can be used as the diagnosis characteristic parameter. The classification algorithm of supporting vector machine (SVM) is used to train the model and test the accuracy of classification. Thus, a new approach applied in non-stationary system is established towards finding a new characteristic parameter of fault. This new approach is applied to the fault diagnosis of a simulated hydraulic flap servo system, 5 system variables including input orders, flap angle, etc. are used as the cointegration variables. The test results indicate that the cointegration coefficients matrix model has great fault diagnosis ability in a typical non-stationary system.