Abstract： Incipient fault detection is of significant importance for preventing the occurrence of accidents. A multivariate analysis method named transformed component statistical analysis (TCSA) was proposed to solve the incipient fault detection problem. The algorithm processes the data in the sliding time window to extract transformed components. Statistics (mean, variance, skewness, kurtosis) of transformed components were monitored to realize the detection of incipient faults. The transformed components extracted by the approach are linear combinations of the normalized data. Statistics of transformed components can reflect some invariants under normal condition. Some incipient faults break the balance and therefore can be detected. Numerical simulation and Tennessee Eastman process (TEP) simulation indicate that TCSA is able to detect both incipient sensor faults and process faults effectively.