BENEFITS OF ARTIFICAL INTELLIGENCE WORLD PULP&PAPER 96 to anticipate an effective knowledge transfer to the younger generation. So how does one utilise the current data streams to successfully achieve both corporate and customer-driven goals? Turning data collection into actionable insight is key to successfully exceeding expectations and reducing costs. Predictive analytics supported by machine learning can provide real-time quality measures that remain robust and accurate in the face of changing machine conditions. These adaptive quality “soft sensors” allow for more informed, on-the-fly process changes, rapid change detection, and process control optimisation without requiring periodic model tuning. THE EVOLUTION OF DATA ANALYTICS Analytics has been in use in the paper industry for decades and, as Figure 1 illustrates, has experienced a gradual evolution. In the early days, critical measurements of both quality and process variables were collected and used for improved machine control, customer specification setting, and benchmarking. This descriptive data collection then evolved into process centerlining, a statistical approach that analyzed where a process’s key variables should be set for machine stability and improved product performance. When the process deviated from the ideal state, a more involved analysis task called root-cause corrective action (RCCA) was conducted to identify what caused the deviation. The goal of RCCA was to identify process improvement opportunities that reduced the chance of failure reoccurrence. However, this approach was diagnostic rather than proactive, meaning it relied on a failure to occur rather than preventing failure in the first place. A newer generation of analytics emerged in which driving process and product optimisation relied on understanding how process variables influenced a machine’s key quality and performance variables. The most basic understanding of these relationships was found by calculating simple correlations, where an important dependent variable was inspected against a single independent variable. This improved the insights an operator could achieve, but it still oversimplified the papermaking process, where the dependent variables of interest, such as quality parameters and production tonnage, are highly related to numerous variables within the process. Further, these multidimensional relationships required the use of nonlinear, multidimensional techniques to fully grasp the relationship between a dependent variable and many interrelated process variables. The need to understand the true nature of these multidimensional relationships ushered in an era of more accurate predictive modelling, allowing for proactive process control of the dependent variables being predicted. Predictive analytics spread quickly across the paper industry and continues to be a valuable alternative to diagnostic analytics. Predictive analytics can provide insight on the impact a change in machine conditions will make on a critical dependent variable of interest. Unfortunately, few of these descriptive, diagnostic and predictive analytics tools, though widespread in papermaking, have been successful in the predictive analytics space, and virtually no prescriptive analytics tools are available. Having a robust predictive model opens the door to the most advanced form of process control, which is prescriptive optimisation. Prescriptive analytics answers the question, “what can I do to achieve my desired dependent variable result?” While predictive analytics uses relevant information about the process and the critical dependent variable, prescriptive analytics provides insight on what process variables need to change and by how much to achieve the desired output. It takes paper machine control to new levels of accuracy, minimising materials and energy waste and maximising first quality production. LIMITATIONS OF CURRENT STATIC PREDICTIVE MODELS Predictive analytic products have been gradually introduced to the papermaking industry. While an array of predictive analytic tools can be purchased from a variety of sources, including chemical suppliers, software companies, and equipment suppliers, these products are developed for generic purposes and frequently lack papermaking focus. Simply put, predictive models are developed using historical tag data to infer a mathematically driven outcome once the new tag data is calculated by the model. A successful predictive model employs the correct techniques by considering the type of data and the true relationships between the response and the predictors to make accurate predictions. Static predictive models ultimately fail when the variables influencing the predictive model constantly change. Machine runnability, grade mix, fibre availability, chemistry usage, water characteristics, Analytics has been in use in the paper industry for decades and… has experienced a gradual evolution.