Industrial AI might be entering the market as one of the key technologies at play. However, machine-learning models trained on production data can detect patterns, identify anomalies, approximate the state of an asset, and advise on maintenance.
Which does raise an interesting engineering question:
How far can an AI go on the basis of what it knows
Manufacturing circumstances have a habit of defying expectations.
The Limits of the Data Pipeline
An idealized depiction of the flow of data within an industrial setting could be something along the lines of
sensors -> connectivity -> data -> ML -> alert -> humans.
Within this chain, sensors might measure temperature, vibration, pressure, state of a machine, energy use, and cycle time.
Those signals might be transmitted and received in a number of different ways through networks, gateways, and processing engines before eventually getting to a data science or machine-learning pipeline.
Each element adds potential points of failure: wrong values, delayed data, mapping errors, timestamp offsets, and incompatible data formats all contribute to a model that has less information about the current state of a process than it might need.
A complex enough model can find ways to make up for missing or corrupt data.
It can't defend against inputs that don't make sense in the first place.
Unfamiliar Ground for Predictive Maintenance
Let's say that there is an AI meant to advise on the maintenance of critical assets within production.
A machine fails unexpectedly, and the cause involves a mixture of factors that were never part of the model's training set.
It will know on a basic level that something is wrong, but determining exactly what requires deeper inspection and a level of understanding well beyond the basics of anomaly detection.
At one level, a predictive-maintenance model is dealing with two different questions:
Anomaly detection: "Is this different from what I expect?"
and
root-cause analysis: "Why is this happening?"
Each has value, but they aren't the same task.
Context Isn't Always Easy to Model Within an AI Engine
Industrial data has a unique set of challenges when it comes to modeling a situation accurately.
Take temperature: it might affect a machine's overall health, but it could also represent a change in a production batch, a different material, ambient conditions, or an unrelated process happening at the same time.
If there isn't enough richness within the input data, the model doesn't have enough information to make a judgement.
Correlating telemetry data with environmental data, production schedules, maintenance logs, PLC state, quality, material, machinery position, and operator action can supply needed context for an AI engine.
The intent here isn't to gather more data.
Rather, it's about providing meaningful context and understanding.
Industrial Edge Computing Adds Response Time
Latency can be a factor in the time needed to react.
A given process in manufacturing might have to respond to a change or event in near-real time, without waiting for data to travel to a distant server, receive analysis, and return a response.
Edge computing handles some part of this problem by bringing computation and storage closer to the machines generating and consuming data.
These systems can still rely on the same principles of machine learning and statistical analysis that power data science, but edge computing helps to reduce delays and assure that appropriate responses still happen even when communications get disrupted.
With an AI model, no matter how powerful or sophisticated, these are the limitations that need to be accepted and accounted for:
Accuracy of input data
Training regime
Feature selection from available data
Monitoring of the model's performance
Ability to recognize unfamiliar data and respond to it appropriately.
Uncertainty Isn't Always Unacceptable in Industrial AI
One key engineering practice for industrial AI is to recognize that a black-and-white certainty about any given recommendation or prediction has practical limits.
A system can provide confidence levels, anomaly detection, and the underlying signals that contributed to a given result without attempting to mask those uncertainties.
In practice, this means that someone inspecting a system has options: continue with normal operations, deepen and broaden the investigation into an anomalous event, perform a known procedure in response to a high-confidence trigger, or escalate an unknown or nonsensical combination of events to a human operator for review.
Engineering Reliable AI for Industrial Applications
For any developer building AI systems or tools for industrial applications, the practical reality is that a production environment involves more than one moving part.
In addition to the machine-learning models themselves, monitoring of data, reliability of the data pipeline, and inclusion or exclusion of a relevant set of sensors for telemetry each plays a role.
At scale, it also involves model and data drift, anomaly detection, and escalation pathways for alerting and response procedures.
Industrial AI and industrial IoT applications attempt to combine many of these capabilities in practical ways, including the automotive manufacturing environments that OEMNex AI describes within its portfolio in intelligent industrial technologies and manufacturing applications .
AI Has a Limit That Isn't Automation
When developing or evaluating AI and industrial applications, it's important to remember that an AI engine doesn't necessarily have a fundamental ability to operate machinery.
It does a much better job at data processing, pattern recognition, and surface analysis.
It has less competence at digging down into a situation or applying physical reasoning to the movements and functions of machinery.
These represent limits to what an AI engine can reasonably do.
A more practical application of AI within manufacturing is to give machines and processes the ability to reason about collected data while also providing workers and operators with the insight and situational awareness when an AI engine isn't certain about a given assessment, allowing for proper escalation and human review.
In these practical terms, knowing when an AI model knows and doesn't know is as much a part of the system as when it has certainty.
Top comments (0)