In this paper, after investigation of the possibility of misidentification of faults with traditional structured residuals approaches, a new optimal structured residuals design criterion for fault isolation is proposed. To maximize fault isolation ability, a matrix of optimal structured directions has been designed. Each of them is insensitive to one subset of faults while being most sensitive to one of the remaining ones. The maximum of all structured residuals in each row is then selected for fault identification. After the faulty sensor is identfied, normal sensor measurements are reconstructed through maximum likelihood estimation. The utility of the proposed method is demonstrated using data from an industrial boiler, which shows improved fault diagnosis performance.