(314f) Statistical Evaluation of a Data Exception and Compression Algorithm Applied to Industrial Data Management Systems

The integration and availability of data in the industry have played a crucial role in the operational and managerial activities of various industries over the years. However, with the intensification of data exploration through advanced statistical techniques and machine learning methodologies, data quality has become a critical priority (Xie et al., 2020). Such methodologies and techniques applied are predominant in industries that have reached or are moving towards the Industry 4.0 paradigm, highlighting the need for high-quality data that represents the real observed process behavior. Specifically, industrial data management is of paramount importance to industrial data infrastructure. Data historians play a crucial role in collecting, filtering, storing, and making available data collected from various sources (Thornhill et al., 2004; Alsmeyer, 2006). The objective of this work is the evaluation of the data filtering stage that occurs before data storing using the PI System infrastructure. In this infrastructure, data filtering is performed by the combination of two distinct algorithms (Bristol, 1990): the data exception algorithm, traditionally known as the deadband algorithm, and the data compression algorithm, named by Bristol (1987) as Swinging Door Trending.

The methodology presented in this paper aims to define the configuration parameters for both algorithms, quantifying the impacts of various parameter sets. As a tool for verification and comparison of data quality, some key performance indicators (KPI) have been implemented: (i) centrality indicators, such as Compression Ratio (CR) and Mean Squared Error (MSE); and (ii) variability centrality indicators, such as Ratio between the Variance of the Reconstructed Values (Ei), Percentage Difference between the Mean Values (PDM), and Pearson Correlation Coefficient (Rxy). Such KPIs were used to compare the quality of the resulting data with the original process data. With this, it is possible to evaluate the data quality after the filtration process, enabling the reproduction of the original data's process dynamics without attenuation or distortion of the variable profiles stored in the PI System Data Archive.

As a case study to showcase the framework, a multi-component hydrocarbon flash unit has been developed in AVEVA Process Simulation, with the flash vessel level dynamics considered under changing operating conditions. The data filtration process was explored by comparing the raw data to the data stored through PI Points in the PI System. By applying the methodology to different process intervals, appropriate parameters for the filtration process were estimated. Results on this method show that the simulation and validation of the data capture and filtration processes in the PI System have provided a reliable and effective approach for optimizing the compression and exception process parameters. In particular, it was possible to achieve an 80% reduction in the volume of the original data, while maintaining the original profile observed in the raw data and reproducing the process dynamics under consideration. The accumulated error was only 0.000638%, demonstrating the effectiveness of the data reduction approach.

Keywords: Industrial data infrastructure; Data historian; Data filtering; Process Dynamics.

References

Xie, H., Li, Y., Chen, G., Liu, Z., Chen, X. (2020), IEEE Transactions on Industrial Informatics, 16(5), 2964-2974

Bristol, E. H., (1987), US Patent 4,669,097.

Bristol, E. H., (1990), Proceedings of the ISA National Conference, 749–753.

Alsmeyer, F., (2006), Computer Aided Chemical Engineering, 21, 1533-1538.

Thornhill, N., Shoukat Choudhury, M., Shah, S., (2004), Journal of Process Control, 14(4), 389-398.