Industrial AI can also be used to track equipment, discern abnormal operational states, increase asset transparency, and enable the manufacturing operations team to make more informed decisions.
But there are two questions that should be asked before we get to the right model, sensor, or automation platform:
What type of operational pain point would this system fix?
Jumping to technology instead of the problem can end up with costly systems that generate data rather than solution. It's better to understand the process, assess the available data, then choose technology as it applies to the use case.
Here's a framework for how to approach an industrial AI project.
- Describe the problem operationally
What if you are running a manufacturing plant and face unplanned machine down time.
This could be the first reaction to introducing predictive maintenance through machine learning, but the facility should explore exactly what is causing the disruption.
Useful questions include:
- What machines have the most downtime?
- Have the failures been documented systematically?
- PPM: Is maintenance carried out based on schedule or equipment condition?
- Do you have sensor data?
- How much are the production hours on hold during downtime?
These responses indicate whether the root cause is the state of equipment, the maintenance schedule, records not being kept up to date, or there is no monitoring process.
If you are reading this, it's probably because you are new to the idea of AI and wondering if it's a good fit for a problem you face. Well, not everything is a nail, and you don't need a hammer.
- Verify if the data is in prepared form
Industry AI learning systems are based on data obtained through physical devices and processes.
Data may originate from vibration sensors, temperature sensors, PLCs (programmable logic controllers), IDs, production software and/or maintenance.
But just because data has been collected doesn't mean it is it is still suitable for analysis.
What areas should you look at before building a model?
- Accuracy: Are the measurements representative?
- Consistency: Are values captured in the same units and formats?
- Completeness: Is there a loss of critical record or readings?
- Up-to-dateness: Does it have the information when I need it?
- Context: Is it possible to assign the reading to a specific machine, process, and operating condition?
A temperature value without a machine tag or timestamp might be confusing.
This means that quality and context ought to be integrated into the system design. They should never be discovered 'after the fact'.
- Understand how existing systems connect
There are usually a blend of older machinery and connected plants.
A production line could have PLCs controlling the machinery, sensors for monitoring, and to keep track of stock and for servicing, separate software.
If these systems are unable to safely communicate, an AI application could receive an interrupted or incomplete data.
A basic integration assessment should consider:
- Available communication protocols and interfaces.
- Compatibility with existing equipment.
- Data collection frequency.
- Network reliability and latency.
- Access permissions and cybersecurity.
- What the application will do in cases of incomplete or interrupted data.
The aim is not for you to scrap everything that is already in place. Instead, you want a trusted method of connecting to the information you need for the use case you selected.
- Select the technology once you know the use case.
Different operational problems require different technical approaches.
For example:
| Operational problem | Possible technology |
|---|---|
Limited asset transparency | RFID, ID system, or location monitoring |
| Equipment condition monitoring | Sensors and analytics |
| Inspection of visual quality | Computer vision |
| Material flow | Industrial automation and robotic systems |
| Process simulation | Digital twins |
These technologies are not interchangeable. It depends on the environment, the volume of data, the manner in which it needs to be integrated, and the desired outcome.
AIoT can embed intelligence in physical devices and sensors, and connect those with analytics so data from industrial processes can be used to make better decisions.
Physical AI, on the other hand, is where the decision made by AI is tied to physical things (by physical I mean things like robots, machine, and actuators). These applications need to be tightly controlled, validated and monitored by humans.
5. Use a small, measurable pilot to test approach
Rather than rolling out a solution facility-wide and immediately, start with one well-defined use case.
As another illustration, an assembly-line team might track a specific cohort of machines to flag unusual vibrations.
Here is a simplified workflow:
``text
Machine
v
Vibration Sensor
v
Data Collection
v
Data Validation
v
Anomaly Detection
v
Maintenance Alert
v
Human Inspection
``
The system will alert staff to abnormal readings for additional analysis. A maintenance technician then inspects the equipment and responds accordingly.
This is a generalized workflow, not a production system architecture. A real implementation would require correct sensor placement, validated models, operating-condition baselines, and correct alarm thresholds.
Set up metrics for success before the pilot gets started.
Possible Signs Could Be:
- Unexpected downtime.
- Time required to identify equipment abnormalities.
- Useful alerts versus false alarms.
- Maintenance response time.
- Integration and maintenance costs.
A model that is technically correct makes sense, but does it make operations better?
- Designing with reliability and human control in mind
Industrial uses of this machine include operations in which the failure can impact production, machinery, or worker safety.
A recommendation made by an AI should not be construed as an automated mechanism to take control of machinery.
Before connecting AI decisions to physical actions, teams should establish:
- What actions the system has permission to undertake.
- Which decisions require human approval.
- Avoidance of commands which are unsafe or unexpected.
- What do we do if a sensors data is missing or suspicious.
- The impact on the system in the event of equipment or network failures.
- How access and security are managed.
Testing should not be limited to normal operating conditions.
Human intervention becomes critical if the application could affect critical equipment or safety.
- Evaluate the operational result
Post-pilot: check technical results and business effect.
In a predictive maintenance application, model accuracy is not all that important. The team needs to test if alerts really help maintenance personnel catch issues earlier and take appropriate action.
In an asset tracking use case, what matters is: have assets become more visible, and are your employees locating tools and assets more efficiently?
The test assessment allow us to decide if the original pilot should be enhanced or if it should be developed further, or if we should go back to the drawing board.
Final Thoughts
AI system is not just about choosing a model or linking a sensor.
It involves the knowledge of the problem to be solved, data validation, interfacing with the existing data infrastructure, selecting the right tools, and finally corroborating performance in the field.
By following a problem-first approach, developers and decision-makers have a more objective reason for making decisions. It also provides an advantage by avoiding unnecessary complexity in the process and keeping the development and measurement of results on target.
If you're working on industrial AI, what's the most difficult aspect to overcome when transitioning from a prototype to full-scale operation: quality, integration, or dependability?
This article was written with the help of AI and edited to clarify technical details. The author has to confirm the facts and have a good understanding of the subject matter before publicising.
https://apertureventurestudio .com/
Top comments (0)