A lot of companies have slowly become “data rich” but they are still “wisdom poor,” as much of their collected data goes unused and un-managed. How can manufacturers move ahead with AI projects efficiently and successfully? This post covers a few of the key concepts for industrial AI in engineering and manufacturing.
Common issues with manufacturing data
Data pipeline engineering is paramount. Very often, data is combined from multiple sources (telemetry and ERP system records, plus video streams) and recorded at different frequencies (milliseconds, seconds, hours, days). As a rule, industrial data is never clean. To complicate things further, the root causes of dirty data are difficult to distinguish. The problem could be sensor drift or malfunction, of which there is plenty. Or it could be related to process drift or raw materials changes, which make data pipelines and related safeguards truly essential. “Data custodian” is a term and function you will need to get acquainted with if you want to materialize the AI-Ops and ML-Ops benefits everyone is talking (or mostly dreaming) about.
The solution is simple: assume your data is dirty until proven clean. Start by engineering data pipelines with the proper safeguards and assign a data custodian to keep the resulting data clean for usage downstream. Then, and only then, can you deliver AI models that will reliably deliver good results–results that will come close to the performance metrics your data scientists will have trained the AI model to achieve.
AI models and a discrete difficulty
Getting AI models to reliably deliver good predictions is complex, but particularly in discrete manufacturing. You first need to establish automated manufacturing data traceability between manufacturing stages. This means that if a technician or robotic element inadvertently changes the order of processing, AI models cannot make reliable inferences or predictions if the data is not also re-ordered.
In continuous manufacturing, it’s a little easier but still requires transformation. This can be dynamically changing in time (i.e., when process speed is not constant, such as when you have conveyer belts or temperature oven with varying soak times). You can typically run automated lag analysis on the data to arrive at the correct inputs, each with a different timestamp based on where the sensor is located along the continuous manufacturing process chain. While a little easier than discrete manufacturing, it is still not as trivial as other AI applications.
An industry 4.0 approach
Data-driven manufacturing can and should extend beyond internal teams. Despite conflicting financial incentives that may impede full data sharing, your AI success may depend on data, such as raw material properties data that would have to come from your key suppliers. With this understanding firmly in hand, the discussion can shift to exploring the mutual benefits of data sharing and cooperation. The growing trend is toward more open discussions about data sharing when the mutual business benefits are clear and well grounded.
Extended data reach
Another aspect of data sharing is acquiring data from products after they have shipped. Engineering teams have a strong desire to get more actual usage data from the systems they designed, as this will feed improvements to the next generation design. This disruptive development in manufacturing means design teams retain a data-driven interest in the future of the product as it is used in the wild. For example, John Deere, Caterpillar, and other heavy machinery manufacturers no longer sell “tractors” –they provide equipment-as-a-service, complete with Service Level Agreements on performance of these assets. This approach changes the way customers view the product and provides a new perspective for engineers and manufacturers whose connection to the product performance doesn’t end with delivery.
AI implementation: sources of failure and solutions
AI implementation failures generally boil down to three sources:
- Data pipelines were not engineered with the right input-data safeguards, for example, against dirty or drifting data and wrong engineering unit or sensor clean-up or recalibration or complete change to a new sensor.
- The technical metrics on which the AI algorithms were trained were not appropriately related to the business results sought.
- The result of the AI prediction is not communicated properly to the human in the loop or to the systemin the loop and is interpreted incorrectly.
The solutions are straightforward and require a concerted effort towards digital transformation:
- Take the time needed to properly engineer your data pipelines and assign a data custodian to keep them maintained and updated moving forward.
- Define your business metrics well upfront, ensuring a cause-and-effect relationship with the technical metrics, or at least an understanding of why AI technical metrics are different and only indirectly related to the business metrics.
- Involve stakeholders and recipients of the prediction or control value in the project as early as possible to ensure the AI data consumption and distribution is tailored to the right format and frequency.
The future potential of AI in engineering
Leveraging reinforcement learning from system-level simulation engineering models for downstream operations shows great promise. The focus on this area is likely to continue along with the current trend toward cloud-based AI and machine learning (ML-Ops) deployments and interest in AI-based machine vision.
Ready to embrace the future with successful industrial AI? Contact a Maya HTT expert today.