Issue link: https://www.e-digitaleditions.com/i/1483907
14 BioPharm International ® Manufacturing and Facilities 2022 eBook www.biopharminternational.com OperatiOns machine learning works and how it can be demon- strated in simpler terms (1). In t h is v ideo, Rober has a g uest who discusses a hy pothetical kid named Timmy who has prefer- ences on toys that are large and complex rather than small and simple. One can then provide the size and complex it y of t he toy a long w it h Ti m my 's deci- sion to the software, and it will accurately depict if Timmy will like a certain toy. The same approach was then applied to baseball signaling in an attempt to "crack the code" on when a runner will steal a base or not. For this to be accomplished, one must provide t he sof t ware inputs and outcomes unti l t he sof t- ware has enough training data to identif y patterns and begin predicting the result. Rober found that sometimes in as little as three signals, the AI was able to accurately identif y if the runner was going to steal or not (1). How can these same principles of machine learn- ing be applied to provide knowledge management to companies with operating equipment? Essentially, one can provide the sof t ware inputs and out puts based on operating parameters until there is enough data to begin predicting if the system is operating at peak performance or has begun to deviate. This is accomplished by creating models that monitor certain transmitter inputs against the desired KPI to understand if those operational changes predict a process a noma ly. A f ter a na lyzi ng t he t ra i n ing data, the model can predict a change in the KPI and alter its expected value and alert range. One such example in a water system is conductivity, as shown below in Figure 1, of a pharmaceutical water gener- ation system in commercial operation. Figure 1 presents a model that is analyzing prod- uct conductivity while using a multitude of different transmitter inputs to model the expected value. The blue trend (dark blue line) is the actual operating data. The numerous inputs fed into the model lead to f luctuations in the expected value (red line) and alert (light blue region). This model is monitoring all the system's inputs and predicts the KPI while comparing it to the live data. The model has been trained on historical data to understand the trends and standard operating nature of the system. This training data allows the software to learn from its previous performance and begin building a baseline to understand process anomalies. This is only one of many models that can be monitored to understand system operations and process anomalies to provide feedback to the customer and store historical knowl- edge of the system. Models can be built using cloud-based machine learning platforms that remotely communicate with water purification equipment on the end-user's pro- duction line. Data is sent to the cloud to update the models in real time. However, leveraging only the data and machine learning isn't enough when it comes to a complex piece of equipment. Machine learning models can predict known process anomalies only if their train- ing data includes that anomaly. Much like a person, the models can only learn from experiences and data that are available to them. When the model produces an anomaly it has never seen before, it will send an alert that something is wrong, and an SME is needed to interpret the results. SME involvement for anomaly and alarm triage One story from a software company, which declined to be named along with its customer, provides some clarit y on t he need for experienced person nel to FIGURE 1. Reverse osmosis product conductivity model. FIGURE COURTESY OF THE AUTHORS.