Since the rise of data science and Industry 4.0, a lot has been written about predictive operational performance in the pharmaceutical industry. The bulk of the discussion is related to predictive maintenance, where applied sensors are used to analyze and predict equipment behavior. Predictive maintenance is mainly focused on the most critical assets through centrally led, time consuming, and expensive predictive data modeling projects.
The newest maintenance concept, prescriptive maintenance, takes it a step further. Instead of just predicting impending failure, it aims to produce outcome-focused recommendations from prescriptive analytics.
But a fully artificial intelligence led prescriptive analytics system for running an autonomous factory is still a bridge too far today. However, organizations can start by improving the reliability of their asset performance. Self-service industrial analytics can bring asset reliability to the next level, helping to increase overall profitability and plant safety.
From predictive to prescriptive
Predictive maintenance can be defined as the effort to reach a high level of asset reliability while reducing unnecessary maintenance costs by decreasing the maintenance frequency. A good predictive maintenance practice ensures that maintenance work is planned at the right time, before a failure mode manifests itself or before the economic impact of reduced asset performance becomes too large.
There are various approaches to predictive maintenance that vary in complexity but have the same goal, such as failure finding maintenance (FFM), risk-based maintenance (RBM) and condition-based maintenance (CBM). Although predictive maintenance is similar to CBM — many even consider it simply a more advanced form — predictive maintenance will typically leverage data from more sensors, detect when operating zones are shifting, automatically detect patterns and issue warnings earlier compared to simple logic rules. This allows the maintenance team to plan further ahead and reduce costs even more.
Recently, prescriptive maintenance has emerged, typically imagined as an artificially predicted form of maintenance, scheduled automatically and with instructions for the required maintenance. This type of maintenance requires even more data from more sources than the “few” sensors on the equipment. Historical operational contextual information is required to artificially assess all circumstances to generate the adequate prescription for the maintenance instructions required.
For both predictive and prescriptive maintenance, deep process knowledge is necessary to analyze the data, build the data models and assess all situations to create appropriate prescriptions for solutions. With data scientists already scarce in the market and those on your central analytics team most likely overbooked, a new analytics approach is required.
One such approach is self-service industrial analytics for process and asset engineers. With self-service industrial analytics, the operational process experts will be equipped to analyze the process and asset performance themselves by using sensor-generated time-series data. They can analyze the data quickly through pattern recognition and build monitors that help them predict future performance for both the production process itself and the equipment performance, all within their operational context.
There are three different prediction approaches through self-service analytics (Exhibit 1):
- The first is event-based: If a certain signature behavior is detected that can affect another part in the process that typically occurs later, a notification can be generated. This notification can include instructions for the required preventive actions or required maintenance.
- The second approach is probabilistic: The current behavior is interpreted, and a likelihood of future behavior is calculated, optionally resulting in automatically scheduled maintenance work orders with the needed instructions.
- The third type is regressive: The prediction is based on certain conditions that must be met and verified, and in case of deviations, the instructions can be given to the control room, or used to schedule required future maintenance.
For all three situations, the events can be captured in case they occur, providing more information for improving future predictive and even prescriptive maintenance work. For a full picture, events or more generic contextual information that may reside in other business applications can be taken into account. Operational data within the maintenance system, lab system, OEE system and/or batch system gives the context for proper assessment of the issues in the production process and the required future maintenance.
Prescriptive maintenance approach
Asset performance, or overall equipment effectiveness, greatly depends on the process in which the asset operates. Instead of just using equipment-related sensor data for performance analysis, all process-related sensor data should be taken into account. This is called “contextualization of asset performance with process data,” and with this, prescriptive maintenance based on data becomes possible.
The goal of prescriptive maintenance is to be able to perform maintenance at a time not only when it is the most cost-effective, but also when it will have the least impact on operations. That requires a good understanding of the process performance. Process engineers (or other subject matter experts) are in the best position to analyze good and bad performance. Modern self-service analytics tools provide descriptive analytics features to quickly explore and filter data visually and search through large amounts of process data (easily up to multiple years of historical data). The advanced analytics capabilities also allow the user to do root cause analysis (RCA), test hypothesis (discovery analytics), and quickly find similar behavioral occurrences.
Through diagnostic analysis, the process engineer can understand effects of process changes (comparing before vs. after) and find potential influence factors for a specific issue. By understanding the difference between good and bad behavior, the basis is created to understand when maintenance is required. With this information, monitors can be created to safeguard best operating zones, increasing asset reliability, improving plant safety and predicting when maintenance is required.
Considering all operational data, all contextual information residing in various business applications and all the knowledge possessed by process and asset experts, a fully artificial intelligence-led prescriptive analytics system for running an autonomous factory is still a bridge too far. Therefore, a human-interacted artificial intelligence system is currently a much safer bet.
Enabling engineers to analyze the process data within its operational context themselves leads to many opportunities to optimize plant performance. The engineers can also collaborate with the maintenance engineers to focus on asset reliability and predictive maintenance, which has led to additional use cases, such as:
1. Heat exchanger lifecycle control for planned cleaning
One example where asset performance is directly related to process behavior is the fouling of heat exchangers. In a reactor with subsequent heating and cooling phases, the controlled cooling phase is the most time-consuming, and it is almost impossible to monitor fouling when the reactor is used for different product grades and when a different recipe is required for each grade. Fouling of heat exchangers increases the cooling time. However, scheduling maintenance too early leads to unwarranted downtime, and scheduling too late leads to degraded performance, increased energy consumption and potential risks.
In this case, a monitor was set up to look at the cooling times of their most produced products. If the duration of the cooling phase starts to increase, a warning is sent to the engineers who can then schedule timely maintenance, sometimes two to three weeks in advance. The gained benefits are extended asset availability, predictive maintenance leading to operational and maintenance cost reduction, and reduction of safety risk.
2. Prediction of equipment before filter breaks
Another example is the prediction of the breakthrough of a filter in a suspension tank. A filter is used for removing impurities in the medroxyprogesterone acetate product — a hormonal medication sold under the brand name Depo-Provera among others — before they enter into the batch. This process happens in batches where the temperature increases, and the pressure increases slightly. Sometimes one of the valves can leak, and gas can enter the system. But sometimes the valve can get stuck due to solids, and the pressure keeps building until the filter eventually breaks.
With the use of self-service analytics, the process engineers can diagnose the root causes in the production process themselves. In that way, they can set up the monitors to identify the problems much more quickly and automatically. When valves are leaking, the equipment can be replaced sooner, or the process can be controlled differently. This pattern can be an early indicator of a filter breakthrough, which could contaminate an entire batch of product. With the predictive monitors, this pharma company could avoid destruction of entire batches of their high-cost product.