Artificial Intelligence (AI) deployments that integrate with the entire ISA-95 technology stack typically win in the game of digital transformation. Typically, this system will involve a plethora of departments and personas, in addition to accommodating existing processes. The result makes AI one of the most challenging systems to successfully implement, especially on your own. Even companies that are ultimately successful must overcome various challenges along the way, both technical and non-technical in nature. So, what can you do now to prepare for these upcoming hurdles? How can you avoid the red flags that often correlate with project failure? By researching best practices and common pitfalls now, you will be better prepared to dive into your own AI projects.
A smart, simple, automated and scalable AI solution, like SORBA.ai, will bring visibility to your inefficiencies and help determine the most important production areas to focus your continuous improvement efforts.
1. Recognize The Need
As the world progresses, the truth is, all companies have and NEED some sort of AI solution. It may be a 100% manual machine data collection with excel regression analyses. Some companies may be so small that all machine tribal knowledge can be tracked in a single person’s mind. There’s nothing wrong with this, however, the fact that you are reading this post reveals there is room for improvement in your current data analytics process. If you are pitching the need to upgrade your existing OT/IT information systems to a more advanced layer of analytics, identifying the key issues that cause production inefficiencies is a critical first step. For example, are your orders filled late? Is overtime required to meet commitments? Do you need to increase capacity? Solidify your understanding of where the highest priority issues lie before approaching the need to others.
2. Sell The Need for Adopting AI Internally
To help unite the gap, every team needs a “Project Champion”, the internal shop floor engineer that pushes the project forward and advocating when it meets resistance. You may not think of yourself as an AI evangelist, but this step must happen in order to achieve a successful ML deployment. Whether early on, or later in the implementation process, everyone will eventually need to be sold on the need for AI. Without question, there will be skeptics and those who are hesitant to change. After all, transitioning to an established AI is full of unknowns, especially for production staff who may not see the high-level need as plainly.
3. Find an Executive Sponsor
Behavioral change is complex, but successfully implementing AI requires convincing those resistant to change to step outside their comfort zone. Gaining buy-in from an Executive Sponsor in upper management provides some valuable forward momentum in this area. The Executive Sponsor is responsible for communicating project goals and gathering support from C-suite executives. This also means advocating for your project to overcome resistance from other senior executives.
4. Define the Scope
It’s important to consider each stakeholder, research the problems or challenges each engineer is experiencing and design your user requirements with the right functionality to help solve these problems. Including functionality that benefits the entire project committee is more likely to result in overall ownership of your ML application. More business groups will have something to gain if the project is a success. Break the project up into lock step gated phases- Each phase should include who the customer (consumer or user of the ML data application) is, the business problems that need to be solved, risk assessment and the success criteria that leads to future project rollouts of the template developed.
5. Infrastructure “Data Plumbing” Audit
The ideal AI system can write data to, and read data from, machine controllers to minimize the manual tasks an operator must perform. In truth, any manual system that does not automatically collect data will be less accurate due to data entry error. An infrastructure audit should be conducted prior to writing the scope so that any required changes are factored into the project funding request. If you are implementing an AI application in a plant that has been around for a while, communications to the machine controllers may not exist, and if they do, they’re labeled as legacy technologies. In some cases, this may require aspects replacing of the control automation layer or adding instrumentation because they don’t have required communications features. Next, the network, ethernet backbone, must be present at each machine controller. In addition to PLC and plant floor device connectivity, you may also need to purchase and install operator terminals, barcode scanners, and overhead displays.
Depending on the interaction with production, servers may need to be on-premises, or you run the risk of stopped production during WAN interruptions. The hybrid solution involves a smaller server that only maintains a limited window of future and past production data. The complete production data is maintained in the cloud or in a private data center. The hybrid solution also helps reduce ongoing cost to maintain backup or redundant databases, backups, etc. because the on-premises server can be redundant and migrated to the cloud.
6. Choosing an AI Platform Partner Wisely
Measure twice, cut one. There are a variety of AI solutions ranging from canned for specific industries, to configurable for any industry. Before committing to an option, it is crucial to document some basic requirements first. These do not need to be a full AI system requirement specification document, but they should provide enough information to vendors to fulfill your basic AI/ML requirements.
For example, your basic AI requirements should include process and/or condition monitoring variables, historical CMMS data and records, SCADA/Historian systems of record in place, environmental regulatory compliance requirements. Try to keep the list as concise and relevant as possible. Avoid vague requirements like “must support reporting”; all systems likely support reporting, so this requirement will not help narrow down the best option for your company.
A major requirement to consider is how AI will integrate with other ISA-95 technologies in place such as Enterprise Resources Planning (ERP), Human-Machine Interface (HMI), Supervisory Control and Data Acquisition (SCADA), and others.
The more custom integration that is necessary to connect everything, the more complex and labor-intensive the implementation will be. This means more development labor and longer rollout timeframe, not to mention the need to dedicate more ongoing resources to maintain the system and custom integration in the future. We want to avoid the high total cost of ownership (TCO) resources!
7. Your Favorite 3-Letter Acronym, R.O.I
After an AI implementation, every company discovers benefits they never expected at the start of their implementation. For example, by successfully implementing a formal Overall Machine Intelligence, OMI (AI/ML integrated with OEE) solution, a major food & beverage producer was able to increase its filtered beer volumes by 20%, resulting in both significant financial savings and increased employee engagement. If an AI solution is implemented correctly, there will undoubtedly be savings. But the million-dollar question is, “Will there be enough savings to pay for the project in a short amount of time or pay-back period”
Validating ROI for AI is challenging when a hypothetical return is hard to estimate. Answering subjective questions like “how much productivity can be gained,” “how much throughput will be increased,” and “how much machine downtime will be decreased” is difficult without historical data “baseline”. If a funding approval depends on a project’s ROI, how should you proceed?
The best approach to determining pre-implementation ROI is to look for inefficiencies such as reworked/scrapped product or production interruptions caused by lack of coordination. There is a very high probability that inefficiencies like these can be resolved by installing an AI application, so an ROI can be calculated in advance. Once you have rolled out the grouping of machines in a facility or area / production line, closely monitor, record and share the KPIs to help calculate ROI on future rollouts. This will likely include the improvements that were not expected at the beginning of the project.
8. Sourcing / Procurement Guidelines
The procurement process involves the terms between the end customer, vendors, and any integration companies. It also involves pricing and legal negotiations that may cause potential delays in the AI implementation. A recommendation for this step is to start the procurement process early, even before the final AI vendor is chosen. Overlapping this step with prior steps will save time.
When fielding a solution for your AI project, consider not only the initial purchase price, but any ongoing costs and labor required to expand scope and maintain the system. TCO includes all ongoing costs, including support renewals, service subscriptions, operator terminals, number of required servers, etc. By conducting an upfront cost analysis, the goal is to minimize unforeseen surprises down the road. Low initial costs look great unless they are accompanied by hidden fees years later.
9. Program / Product Management
Far too often, AI implementations may fall in the hands of an OT single engineer or IT department. Sometimes this works out; however, without proper software training or onboarding, an engineer may overcomplicate the implementation or fail to use techniques that can be easily rolled out. This may result in project delays and a system that is difficult to maintain after the initial AI implementation is complete.
Because AI implementations involve many departments, the need for a program or product manager (PM) that handles the coordination between all of the different entities is crucial. The PM monitors the implementation progress and ensures all disciplines are progressing smoothly. If any discipline is not progressing smoothly, the PM mission is to bring the project back on track by whatever means necessary. Without this role being filled, the risk of delay or failure greatly increases by 10x.
Although not always the case, it is fairly common that IT and OT are not aligned with their technical & business goals. The two business groups must be married to the hip during an implementing for an AI solution because data is passed back and forth between the shop floor, edge/fog computing and cloud. Also, IT is usually responsible for managing the servers running the AI software platform. If the two departments are like oil and water, then the PM is the babysitter to help keep both IT and OT “children” behaved and out of “time-out”.
10. Validation of AI Solutions
This step is often overlooked but is very important to achieving success. Any time operators, supervisors, management, maintenance, and others experience incorrect data or bugs, the credibility of the AI system is diminished. Negative experiences can even reinforce the cynics who were opposed from the beginning.
The most important method to keep credibility high is to fully test, end-to-end, the various AI/ML applications with team members who are not a part of the original implementation team. This can be done in a series of open/closed loop testing scenarios (I.e., writing back optimized setpoint control) in high/low demand production states where risk can be easily mitigated.
11. Technology Knowledge Transfer
Training and onboarding are commonly overlooked but have a significant impact on the overall success of adopting, maintaining and scaling your AI implementations. Operators need to be prepared to quickly handle the various production issues that may arise. Otherwise, a system that was commissioned to increase efficiency may accomplish just the opposite. It is wise to include operator training in advance of startup, but not until after the user interface and experience is stable and drives in such a way that provides a new sense of “convenience” in day-to-day operations.
Training the maintenance staff on basic operations, solution architectures, and new data flows are encouraged. When production outages are scheduled or instrumentation is re-calibrated by someone who doesn’t understand the potential upstream impacts on the AI application, or who doesn’t understand how to handle unexpected environmental or seasonal conditions in data flow, this will ultimately reduce efficiency causing potential re-work.
Other disciplinaries may need training include IT, quality, management, and scheduling staff. Identify those that will be using or supporting the AI software platform and proactively educate them prior to going live.
Training is an excellent time to pre-test your AI/ML solutions. Usually, if something was overlooked or not working correctly, it will surface during training. It is also a good time to see how people interact with the AI software to make usability improvements.
12. Time To Value in a Production Pilot
Online pilots are generally not crucial unless you fall into a high-risk category. In these riskier scenarios, a production pilot is highly recommended to avoid risking both credibility and finances. If a project is technically challenging or hasn’t been done before, then an offline proof of concept/value is critical. In other words, testing a theory is wise prior to gambling on that theory.
If a project doesn’t fit into the culture of your company, then use a pilot to test if the behavior change is achievable. If you position the test as an open pilot where ideas are welcome, then most contributors will be focused on what will help make operations run smoother. It is much less risky to consider new ideas at this stage versus after a large investment and substantial development have already been dedicated to a project.
Pilots are also valuable in collecting ROI data. The only way to accurately estimate what you can expect for your ROI is by doing a pilot.
The rollout phase can determine the success or failure of an AI implementation with a validated template approved by upper management. Depending on the quantity of tags required for monitoring and the size of your production facilities, the time required to roll out AI to each facility will have a major impact on the implementation’s success. Factor in your economies of scale when determining what internal resources can deploy the solutions quickly versus engaging a 3rd party systems integration partner that can help speed up delivery implementation schedules.
Some common items that increase rollout time include hard coding, baseline scripting, data massaging, delayed funding, over-complications, and scope creep to accommodate each facility’s desires. The term hard coding represents a structure that requires a lot of manual code changes for each manufacturing facility. Wherever possible, aim to streamline configuration of a system to the equipment data profile/model of each manufacturing facility using NO-CODE approaches and design/integrate Unified Namespace (UNS) data foundations from the get-go.
Instead of including every feature on your wish list in the initial rollout, start with an MVP (Minimum Viable Product), and roll out updates and new features over time. Whenever possible, utilize out of the box functionality and automated production ready components rather building custom in house applications resulting in future costly technical debt. Start small and avoid over-complications. By aiming for an MVP and keeping it simple, you can learn from early experiences. Remember Machine Learning gets better with time, digital transformation is not a destination but a JOURNEY.
Ready to Start a Free 30-Day Pilot Program, No Strings Attached?
A full AI/ML solution requires technical systems, training, analysis, new procedures, behavior change, and more. Chances are your company is changing constantly, and as a result, your desired AI/ML application/requirements will be in constant flux as well. New products are added, machines are changed or moved, and new initiatives are implemented that all affect AI's ability to generate value.
Being flexible with changes from the start will enable you to better accommodate them. As more employees are trained, more momentum is gained, resulting in a greater ROI and increased bottom lines.