Data Quality Assessment

Data Quality Assessment Yuri A.W. Shardt Data quality assessment seeks to provide a framework for determining the quality of stored data for use in process system identification. The benefit of using historical data for identification is that the potential exists for developing models without needing new (open-loop) tests. This can reduce the costs associated with process identification, as there are no costs associated with loss of productivity or safety concerns. Furthermore, data quality assessment can provide guidance for why the resulting models are not sufficiently good when tested on new data. This often is the result of poor data quality in the initial, modelling data set. In any data quality assessment framework, the following topics must be considered: 1) Motive: For the purposes of this report, the motive is system identification from routine operating data. Therefore, there is a need to specify the model structure and time delay in order to obtain a measure of the data quality. 2) Outliers: Outliers are data points that are somehow different from the expected value. If it is assumed that the number of outliers is small in comparison to overall data length, the impact on the overall metric should not be sever. If it is desired to remove outliers beforehand, then recourse can be had to any standard methods, such as 3σ-edit rules or box-plot analysis. 3) Data Length: For routine operating data and due to the fact that certain modes may not be excited often, the data length is especially important. Empirically, it has been determined that less than about 1,000 data points can provide unsatisfactory parameter estimates. 4) Model Changes: When examining data from the historian, it is possible that the process conditions and hence the underlying model has changed. This implies that there is a need to partition the data into its constituent regions in order to avoid fitting a single model to a multimodel region. As well, model changes can inflate the quality of the data with respect to a single model, but provide unreasonable parameter estimates.