While attaching a sensor to the thing being sensed and plotting it on a dashboard is straightforward, actually engineering a pipeline of data that translates the event into meaningful operational decisions is rather non-trivial.
An industrial AIoT architecture will have to grapple with the following:
Data collection, validation, processing, inference, and downstream action given an alert.
- First: come up with a sensible event model
Let us think of a normalized example of sensing some industrial equipment. The event schema we need could be something along the lines of
asset_id, timestamp, temperature, vibration, operating_state, and sensor_status.
The kind of information collected, of course, depends on the equipment and the particular use case. What is rather important is defining common fields, timestamps, units of measure, identifiers (the way assets and even sensors are identified), and event format.
- Data validation, checking for required fields, outliers, staleness, and other vector for errors.
Some very simple validation code in Python would be something like:
def validate_reading(event):
"""Check if event has all expected fields and they are not NaN."""
required_fields = ["asset_id", "timestamp", "temperature"]
for field in required_fields:
if field not in event:
return False
if event["temperature"] is None:
return False
return True
Again, this is something very simple: actual production-grade code will have to do more comprehensive type checking, timestamp validation, unit validation, possible error reports at the sensor level, and other checks that make sense in context.
- Consider decoupling event ingestion from processing and rule evaluation for the purposes of reliability, observability, and maintainability.
There are many ways that such a pipeline can be structured:
gateways can receive the events,
the system can ingest them into a storage or processing layer (rules engine or a model),
a separate service can consume processed or raw events and generate alerts,
and there should also be some infrastructure-level monitoring for such a pipeline.
The more generic the infrastructure for processing these events is, the easier it is to use it across different use cases. Depending on the scale and the SLAs required, data may be processed on the edge, on the cloud, or a combination of both.
- Choosing the right tools for the job at hand.
Sometimes a simple rule can be better than a machine learning model, if the required performance, interpretability, and maintenance budget are taken into account. Statistic process control-like approaches can be much better at detecting outliers.
At the same time, more complex rules or even black-box ML models may be necessary for more nuanced use cases when operating conditions are barely predictable.
It all boils down to identifying the false-positive budget and business-driven priorities.
- Make sure alerts actually have actionable information and can get to the right people or processes.
An alert, if necessary, should have clear recommendations on what to do next, identify all relevant measurements and assets involved, and have all the relevant contextual information such as timestamps and causes. Again, what constitutes an actionable piece of information depends on the use case and the downstream workflow.
- Finally, make sure the system is monitored and tweaked in production.
Such a system will benefit tremendously from having dashboards that monitor ingestions and alert rates, data quality, latencies in processing, model performance (if applicable), and business-related metrics when possible. The operating conditions of any industrial setting, equipment, and sensors will constantly change, so the models and pipelines will need to adapt.
Conclusion
Industrial AIoT systems are seldom about the models themselves; building an actual production system is a data and operations-heavy endeavor just as much as it is a ML challenge. Data validation, observability, rules, and even downstream actions must be designed with care. Aperture Venture Studio works with companies across different areas of industrial applications and digital operations; get in touch if you are building or scaling an industrial AIoT solution of your own!
Top comments (0)