The New Automation: Industrial AI

• Probabilistic approaches like machine learning AI with existing data is not ideal for many industrial use-cases
• Most industrial use-cases need a deterministic approach to be effective
• The recommendation is for I&C engineers to deploy analytics with mechanistic AI based on known cause & effect and first principles together with sensors for additional real-time data
• The result is earlier, more precise, and more robust predictions and diagnostics which in turn drives many operational benefits

AI in plants to transform maintenance, reliability, integrity, sustainability, production, and quality assurance work has higher demands on analytics being more predictive, more specific, and with little or no false alarms and therefore more actionable than AI in the office such as for customer care, human resources, sales & marketing, and supply chain management. The insight from this is therefore that the AI approach in the plant must be different from the AI approach in the office. 28 April is the International Automation Professionals Day #AutomationProDay by the #ISA so let’s look at how, with the right budget, I&C engineers can deploy an AI analytics framework for the plant, different from the AI analytics framework deployed by the IT team for the office. It is a question of the right tool for the job to transform work in the plant. What is the recommended practice? Here are my personal thoughts:

Office: Business AI

The McKinsey article “An executive’s guide to AI” lists many machine learning AI use-cases: Buying preferences like assess product sales drivers such as prices, distribution, and advertising. Optimize price points, predict client churn, assess which product attributes make a product most likely to be purchased. Assess sentiment and product perception, predict likelihood of someone clicking on an online ad, and forecast product demand and inventory levels. Predict the price of used cars. And customer segmentation. Finance like detect fraudulent activity in credit-card transactions, and classify customers based on likelihood of repaying a loan. HR like segmenting employees based on likelihood of attrition. All these are business examples and attempt to predict human behavior. Machine learning AI is used because we don’t understand how the human mind and emotions work, and there are no sensors on people’s brains, so for use-cases involving humans, statistical correlation data science analytics is used. This is the kind of analytics you’d expect to find in an office, not in a plant.

Machine learning AI essentially is automated curve fitting statistics like regression and support vector machine (SVM). The predictions are not precise for a specific individual but works out for larger groups of people. For business purposes, statistical methods like machine learning (ML) are accepted because they are better than nothing at all. It works by first ‘training’ the ML app by feeding it large datasets of historical data to find correlation between variables. But, as the saying goes; “correlation does not mean causation” meaning ML does not understand causality or underlying mechanisms so there will be false positives and missed opportunities because there are confounding variables and missing variables which skews the correlations. Before ‘training’, all manually collected data must be ‘cleaned’ to reject outliers that don’t make sense. Any free format manually entered text records must be classified into simple codes.

The one use-case in the McKinsey article with an industrial connection is to predict power usage, or any utility for that matter, where machine learning makes sense because consumption is a human activity subject to human whims.

Use-cases for deep learning AI examples (also from the McKinsey article) include Image like help assess diseases from medical scans, assess brand perception and product usage in media, detect a company logo in social media, detect defective products on a production line, assess disaster area images for damage claims, and generate image caption. Sound like language translation.

Plant: Industrial AI

AI use-cases in plants include reliability and maintenance like predicting failure in rotating machinery like pumps, fans, and gearboxes. Integrity like predicting corrosion and erosion failure in piping and vessels. Sustainability like optimizing the time of cleaning of equipment like heat exchangers, cooling towers, and air-cooled heat exchangers to restore performance after fouling. And production and quality such as predicting process upsets. The good news is that we know in detail how these equipment work, fail, and foul. This well-established knowledge is what mechanistic AI is using for prediction and diagnostics. This is what is known as cause & effect relationships, causality, or sometimes referred to as failure mode and effect analysis (FMEA). It of course also abides by first principles (1P) physics and chemistry.

Mechanistic (or physics-based) models are built using knowledge and information about the way the world works. They encode well-established laws and relationships and can be used to generate testable predictions. Mechanistic models can project beyond the data they are fed, elucidating how (and, at the mechanistic level, why) certain scenarios are likely to arise. - Nova Discovery

Mechanistic AI, sometimes called physics-based AI or by other names, is apps where subject matter expert domain knowledge in the form of the well-known cause & effect relations and first principles has already been codified into the app, such as an app for pump analytics. Only relevant variables are used. This way false positives and missed opportunities are minimized. No historical data cleansing or app training required because we already know in detail how these equipment work, fail, and foul. Such readymade apps are referred to as engineered analytics. This makes mechanistic AI software easier to deploy and use. You don’t have to become a programmer or data scientist.

Mechanistic models are based on the fundamental laws of natural sciences, including physical and biochemical principles. Less experimental data is needed to calibrate the model and determine unknown model parameters, such as adsorption coefficients, diffusivity, or material properties. An essential benefit of mechanistic versus statistical models is that the model parameters have an actual physical meaning, which facilitates the scientific interpretation of the results. - Cytiva Lifesciences

?The most interesting fact is that most likely you are already familiar with mechanistic AI, just never thought of it by that name. Vibration monitoring systems for protection of large turbomachinery have used software with embedded cause & effect principles for decades to predict and diagnose wear to successfully guard critical turbines and compressors. Process simulation systems have used software with first principles (1P) for decades to predict the outcome of process or controls changes to safely allow operators to practice their skills and process engineers to optimize. Now we are replicating these successful approaches to all equipment and unit operations across the plant. Readymade mechanistic apps are also available for complete pieces of equipment such as steam traps, pumps, heat exchangers, air-cooled heat exchangers, pressure relief valves, cooling towers, corrosion/erosion, industrial lighting, and control valves etc. The recommendation is to use mechanistic AI for process and equipment analytics.

Since we don’t put sensors on human skulls and we don’t know the first principles or cause & effect of the human mind, mechanistic AI is not used in the prediction of human behavior.

Multivariate Analysis (MVA)

Many processes such as those in petroleum refining, petrochemicals, bulk chemicals, and many more are well understood with decades of study and research, often licensed processes. Predictive agents and models for simulations can therefore easily be created for the associated end-to-end plant process units and may even be part of libraries in readymade software. However, some newer industries use new process technologies such as in lithium refining, rare-earth metals refining, new hydrogen electrolyzer membrane technologies, and new carbon capture membrane technologies or new amines to name a few. To better understand the dynamics of these processes at scale so they can be optimized for throughput, yield, energy efficiency, consumables, cost, purity, and other quality parameters, these plants run process analytics in the form of Multivariate Analysis (MVA) such as regression, Principal Components Analysis (PCA), Partial Least Squares (PLS) regression, and Principal Component Regression (PCR).

MVA process analytics uses data from the control system which is all collected automatically. There is no manually entered data from the ERP required. This avoids issues normally faced with manually entered data, especially free format text, such as maintenance records.

The recommendation is to use an industrial MVA app which allows you to adjust for process dynamics like lags between cause and effect when looking for correlations.

Based on the analysis, the MVA app can then monitor multiple process variables to predict process upsets while the process is running. This gives operators a chance to adjust parameters to keep the process on the correct trajectory. The MVA app can also be used for “soft sensing” of variables that cannot be measured directly. Key to robust MVA is for a subject matter expert to validate the relevancy of variables used against first principles.

Industrial Software

There is one more important attribute of industrial analytics software which makes it industrial, or in the grander scheme of things – is part of making all industrial software industrial; it uses standard automation software interfaces. People in the plant do not like custom coded APIs because they are costly and time consuming to develop and tend to break when software is upgraded. Industrial AI analytics just like all industrial software instead use the standard OPC-UA interface, or in some cases the classic OPC-DA or Modbus/TCP. Because these interface technologies are extremely common in other industrial systems and software, the analytics can easily tie in with the existing automation in a plant. No custom coding required. The I&C team is very familiar with industrial software integration. The recommendation is to use software which supports OPC-UA.

Plants Need More Data

“Having too much data” is presented as a ‘problem’ by those that want to sell analytics as a solution to that ‘problem’. However, “having too much data” is not a problem. If having too much data was the problem, you could simply delete it and disconnect your sensors. The problem is not too much data so please do not delete your data and do not disconnect your sensors. The ‘problem’ may be restated as “we don’t know what to do with all the data we have” as in not knowing what problems their data could potentially solve. That is also not the right way to think about it. Analyzing data just because it is there would be tantamount to analytics for the sake of analytics. Instead think about what operational problems the plant has in the domains of reliability, maintenance, integrity, sustainability, production, and quality, that you need to solve. Identify the problems you can solve with analytics, and then check what data you need to solve those problems. The documentation for the mechanistic AI apps from your automation vendor tells you which data you need, depending on the failure modes you want to predict, so no research required. Some of that data you already have, but many measurements are missing and need to be added. Remember, you need real-time data from sensors to make prediction and real-time diagnostics. You can’t do prediction or real-time diagnostics using historical data.

Plants have lots of data, from thousands of sensors, collected over many years. But this tends to almost exclusively be process data. Process data is excellent to predict and diagnose process problems including quality issues like off-spec product, but you can’t predict or diagnose equipment problems using only process data. You also need equipment data. For this reason, again additional sensors need to be installed in plants to get real-time equipment data like vibration, acoustic noise, and ultrasonic thickness etc. Data collected manually from time to time using portable testers is not adequate to make prediction or real-time diagnostics. And you will not find real-time data in the ERP either.

One luxury we have in plants is that we can deploy as many sensors as needed, and permitted by the budget, to get early and accurate predictions. Thus the automation budget is one of the most important line items for a plant. The recommendation is for the I&C team to deploy a wireless sensor network (WSN) based on the WirelessHART (IEC62591) standard and to deploy non-intrusive wireless sensors often referred to as monitoring and optimization or M+O sensors or IIoT sensors. Wireless and non-intrusive are key sensor attributes for easy deployment. I&C engineers are very good at figuring out exactly what type of sensor is required. Because mechanistic AI does not need historical data for ‘training’, you can deploy the analytics the same day as you install the new sensor. There is no need to wait months or years for representative data to amass.

Soft Sensors

“Soft sensors”, refers to software apps which compute ‘complex’ product properties which cannot be measured online, such as Reid vapor pressure, using multiple ‘simple’ measurements. This gives a much faster result than grabbing a sample and bringing it to the lab for offline analysis. Soft sensing requires multiple measurements like pressures, temperatures, and flows etc. Some of these may already be available from sensors connected to the DCS. However, very often it is necessary to install additional process sensors to provide inputs to the soft sensor app.

Operational Excellence

The result of successful analytics based on mechanistic AI and additional sensors is greater sustainability thanks to greater energy efficiency, lower emissions, reduced losses, and lower energy cost. Greater reliability and reduced maintenance cost. Greater throughput and reduced off-spec product. Thanks to automatic data collection there is also an overall improvement in productivity as well as greater safety as fewer people work in high-risk areas. The new automation solutions make every-day work in the plant easier.

Automation Budget

So AI is not taking the place of traditional plant automation. Industrial AI is a new additional part of plant automation, hyper-automation if you will. The recommendation is for the operational departments in the plant to work with their I&C team to put forth a proposal for greater automation budget for more industrial software and more sensors. Start with a workshop to uncover the plant challenges.

Lead the way. Schedule a meeting for 28 April, the International Automation Professionals Day by the ISA, or today.

Share this essay with your sustainability, reliability, maintenance, integrity, and I&C managers now.

And remember, always ask vendor for product data sheet to make sure the software is proven, and pay close attention to software screen captures in it to see if it does what is promised without expensive customization. Well, that’s my personal opinion. If you are interested in digital transformation in the process industries click “Follow” by my photo to not miss future updates. Click “Like” if you found this useful to you and to make sure you keep receiving updates in your feed and “Share” it with others if you think it would be useful to them. Save the link in case you need to refer in the future.