Fault detection and diagnosis of a continuous process with input fluctuations such as load fluctuations can not be simply performed because of the difficulty of detecting the abnormality in the process, where normal values of state variables are uncertainly time-varying according to the input fluctuations. In this study, normal values of output variables are predicted by the use of the approximation function that is drawn from a family of orthogonal wavelets. Measured value of an output variable is classified to five-range signs (+, +?, 0, - ?, -) by simultaneously performing two sequential probability ratio tests (SPRTs) based on the error residual between predicted and measured values: one test examines whether it is normal or higher than normal, the other examines whether it is normal or lower than normal. A combination of signs given to all the output variables is called a pattern, which is considered to represent an abnormal situation occurring in the process. Then, the fault diagnosis algorithm, based on signed directed graph (SDG), can deduce the fault origin that causes the pattern. The effectiveness of this approach is demonstrated through experiments by a tank-pipeline system.