Pragmatic approaches to bound industrial AI for safe, reliable and sustainable critical asset operation (Guest blog by AVEVA)

What AI model options to consider? What data to use in training AI? Would AI help improve safety, reliability, and sustainability of operations? Would AI forecast and prevent critical asset failures? How to verify and validate AI model? 

With the rise of no-code usable Artificial Intelligence (AI) systems out-of-the-box, such questions are being asked, and increasingly, by operators of critical industrial assets to understand and instill their confidence in AI systems. Critical industrial assets are essential for functioning of society, such as power generation plants, chemical plants, oil refineries, water supply, transportation etc. And many industrial asset operators are facing challenges due to aging assets and workforce and the need to analyze vast amount of real-time data for decision-making. Hence, deploying trustworthy AI systems for real-time monitoring and early diagnostics of such critical asset-intensive businesses is key to support safe, reliable, sustainable, and economical functioning of the society. And to enable growing AI eco-system to develop policies for safe, assured, and sustainable AI; there is a need to understand it, track it, see precisely how, where, and why it works. 

AVEVA is a UK headquartered global software company, providing design and operations solutions for the world’s industrial sectors, including power, manufacturing, chemicals, oil and gas and shipbuilding, amongst others. These industries typically work with vast datasets so are well suited to AI application. 

AVEVA^TM Predictive Analytics ^[1] enables critical asset-intensive companies to avoid costly failures by identifying asset anomalies weeks or months before failure, forecasting time to failure and giving prescriptive advice such as actions to remediate problems. For the trust part, in any engagement, AVEVA engineers (as the responsible party and assurance provider) leverage procedures and mechanisms built within AVEVA^TM Predictive Analytics software for demonstrating and communicating trust to customer engineers (assurance users). Our pragmatic approach includes satisfying seven ‘V’s or dimensions of industrial big data - Volume, Velocity, Variety, Variability, Veracity, Value and Visualization - across the entire data lifecyle of AI systems to develop, deploy and manage an effective and safe AI solution. 

• Volume: Defining the scope and gathering historical time-series data adequate to the predictive analytics problem being solved. 
• Velocity: Setting sufficient frequency or resolution of imported time-series data to enable users to detect any changes in the equipment’s behavior. 

• Variety: Investigating sources of industrial time-series data including sensors, data lakes, data historians, calculate values etc. 
• Variability: Cleaning the imported data to remove spikes and abnormal behavior for consistency, accuracy, quality, and trustworthiness. Using tools like automatic outliers to exclude/delete data either in time-series or relational comparison mode. 
• Veracity: Building and accessing the operational profile (AI model) in a tabular format containing multi-dimensional cluster of training dataset. Setting filters to activate or deactivate the online AI model to avoid spurious alarm depending on modes of equipment operation. Setting warning / alert thresholds. Evaluating operational profile using data playback tool and verifying/validating that the model tracks well. 
• Value: Configuring alerts to provide early warning when an asset’s current operation deviates from its learned asset signature. Linking alerts to diagnostic, prescriptive, and prognostic information about the asset, or configuring notifications to notify responsible stakeholders of alert conditions. Forecasting time-to-failure. Using transient data module to monitor abnormal conditions during transient events. 
• Visualization: Monitoring equipment health, alerts, and case management via trend visualization in an intuitive web interface. Watching for shifts in overall model residual (OMR) trend and individual sensor changes at that time. 

With these comprehensive safety, security and reliability checks in place for AI models, industry leaders trust AVEVA™ Predictive Analytics to maximize their asset operation safety, reliability and prevent unplanned downtime. To avoid any illegal access and misuse, it integrates with existing enterprise security systems and supports single sign-on authentication, with the ability to limit user access rights and editing privileges at a granular level, delivering stringent cybersecurity protection and the highest operational standards. 

Examples where loss of control and harm to human life was avoided successfully by AVEVA™ Predictive Analytics include sophisticated turbine “catches” where there were step changes of vibration reductions (not increases). Each time, the manufacturer told the customer it was OK because it was a reduction in vibration, not an increase ^[2]. With this particular situation, it turned out to be due to the beginning of blade separation within the turbine stages. The system was nowhere near a control system alarm or warning. However, had it gone on, it would have resulted in a catastrophic failure that could have destroyed the turbine, caused extensive downtime (loss of power production), and a potential for significant injury to personnel. Conservative estimates by the customer showed that over $34 million USD were avoided due to AI systems early warning detection of issue. Another example ^[2] occurred during a major storm with high winds where a transmission grid company leveraged AI and advanced analytics to prevent a catastrophic transformer explosion. The system alarmed due to unusual patterns of dissolved gas analysis (DGA), including methane and carbon dioxide. The company dispatched technicians to investigate and saw that the breakers had tripped open due to the hurricane, leaving the transformer in an energized state with no load. Had the breakers merely been closed per normal procedure, a major explosion could have resulted, destroying the transformer, potentially injuring people, and causing cascading outages. These types of high-voltage transformers cost in the range of €10 million each, and spares are not typically kept on hand as backup. This was a major “catch” that the customer proudly cites as a huge avoided cost in terms of asset damage, lost transmission, and human safety. 

More information and similar examples of enhanced safety and reliability in critical asset operations due to successful implementation on AVEVA™ Predictive Analytics could be found here. 

[1] 

[2]