Compression of chemical process data by functional approximation and feature extraction

Effective utilization of measured process data requires efficient techniques for their compact storage and retrieval, as well as for extracting information on the process operation. Techniques for the on-line compression of process data were developed based on their contribution in time and in frequency using the theory of wavelets. Existing techniques for compression via wavelets and wavelet packets are inconvenient for on-line compression and are best suited for stationary signals. These methods were extended to the on-line decomposition and compression of nonstationary signals via time-varying wavelet packets. Various criteria for the selection of the best time-varying wavelet packet coefficients are derived. Explicit relationships among the compression ratio, local and global errors of approximation, and features in the signal were derived and used for efficient compression. Extensive case studies on industrial data demonstrate the superior performance of wavelet-based techniques as compared to existing piecewise linear techniques.