In manufacturing and production processes, analytics are increasingly being used to help these facilities operate more safely, reliably, and efficiently. In an Analytics Association of the Philippines presentation, The State of AI in Singapore (Facebook video beginning at 5:30, run-time 19 minutes), Emerson’s Jonas Berge touched on process plant challenges, analytics solutions, and deploying operational analytics.
Process manufacturers, depending on their industry, may face numerous challenges. These include hundreds of tons of flammable substances and tons of other toxic substances, very high pressures, thousands of pieces of plant equipment, kilometers of piping, corrosive and abrasive fluids, and hundreds of unit processes.
Operational analytics can play a significant role in addressing these challenges. Some analytics solutions can predict process upsets, process properties, equipment fouling and inefficiency, equipment failure, emissions, and pipe corrosion. Jonas contrasted various types of artificial intelligence—engineered analytics based on programmed rule-based analytics, machine learning (ML)-based analytics, and deep learning (DL) neural network-based analytics.
Deterministic analytics based on first principles and well-established cause and effect relationships are fast and great for equipment to monitor performance and failure conditions. Jonas’ presentation addressed rule-based analytics and ML analytics. Rule-based cause & effect (RBCE) analytics provide failure prediction, are deterministic, are descriptive in distinguishing failure modes and not just anomalies, and are prescriptive in recommending actions to address the issue. The first principles can be inference-based—fouling, performance, efficiency—or integrity-based—corrosion and erosion, for example.
Readymade apps and templates, such as Plantweb Insight applications, can be applied, which do not require data cleansing, coding, modeling, training, or testing. They are built for purpose and don’t require programming and data science skills. The input for these analytics comes from real-time sensors.
Machine Learning predicts human consumption and uses statistical approaches, including regression, principal components analysis (PCA), support vector machine (SVM), and neural networks. The results are probabilistic. These analytics require data cleansing, coding, algorithm selection, training, and testing. Data science and programming skills are required. This type of analytics is not the best for equipment health prediction because it provides anomaly detection but not prescriptive guidance unless there are many years of equipment historical data.
ML analytics works well in large end-to-end systems with complex interactions between many unit processes and equipment with no established cause and effect patterns. Inferential sensors can be added where no direct measurements are in place. These analytics use fault trees for failure prediction, describe failure modes and recommend failure modes’ actions. ML analytics need a long history of sensor data to uncover strong correlations.
Wireless devices are available to easily add sensors where needed and come in a wide variety of measurements. These include corrosion, erosion, vibration, position, level, discrete (on/off, open/close, etc.), flow, pressure, temperature, acoustic, toxic gas, geolocation, oxygen depletion, etc. These sensors can be applied for safety & health, reliability & maintenance, production & quality, and sustainability applications.
Jonas highlights steps in the journey for a successful analytics project implementation.
Watch the video for more as Jonas shares his thoughts on required skills development and case studies from manufacturers who undertook this journey to more predictive operations. Visit the Plantweb Digital Ecosystem section on Emerson.com for more on the operational analytics to drive improvements in safety, reliability, and efficiency.
