Abstract: Considering the non-stationary nature of chemical process, this paper presents a multi-resolution analysis (MRA) methodology for data processing. It is based on the time-frequency localization of wavelet analysis, which can extract temporal features localized in both frequency domain and time domain of process signal. In this algorithm, multi-scale wavelet transform (WT) of the process data is made, and the process data is decomposed into scaled signal and detail signal on each scale. According to their unique characteristics created by WT in the detail signals across time or scale, different variations are identified, such as steps, peaks, and noises. Especially, abnormal sudden changes, which are the most harmful temporal features to process analysis, are detected from a couple of WT modulus maxima with abnormal large amplitudes. With different variations modified with appropriate strategies, process data are reconstructed from the modified scaled signals and detail signals. It is a close approximation of the original process data, with noises discarded and abnormal sudden changes treated effectively, which traditional Fourier Transform of filtering is powerless to do. Using this method, the real process trends are extracted from the raw process data correctly. It is a powerful tool for dynamic process trends analysis, such as process steady state detection, faults diagnosis on-line and real time detection of gross errors in dynamic process data. This method has been applied to processing practical data of crude oil distillation plant. The result shows that the method is essential and effective for process trends analysis and treatment on-line.