Simon Jagers | Samotics

Finding What Needs Attention

Much of today’s discussion about AI concerns generative models. That is not what is being discussed here.

Conventional monitoring relies on thresholds, predefined rules or statistical baselines. Machine learning extends that toolkit. It finds patterns across operating conditions and combinations of signals that are difficult to encode by hand. It can tell whether an asset is behaving differently from its own normal, or whether a change resembles something previously linked to a mechanical, electrical or hydraulic problem.

But finding a change and understanding it are different jobs. Consider an estate rather than a single station. On any given day, the large majority of a utility’s pumps show nothing that warrants human attention. Machine learning is the first filter. It screens the whole population, continuously.

CMAs act as the second filter. They review the findings, compare signals, watch how a condition develops and bring pattern knowledge built from investigating similar assets and similar faults. Sometimes the answer is to escalate. Sometimes it is to keep monitoring. Sometimes an apparent anomaly has a perfectly reasonable operational explanation.

That division of labor is the point. The models screen a volume of data no team could review manually. The CMA spends time on the small number of cases where interpretation and engineering judgment change the outcome. The utility then adds context no one outside monitoring team can fully hold: the station, the catchment, the wider network, etc.

Deciding When to Act

A second case, at a different station on the same utility’s network, shows why that last layer matters. Monitoring identified a developing blockage on one of two pumps. A CMA reviewed the evidence. The second pump was healthy and the station still had redundancy, so there was enough margin to keep monitoring the condition and plan an inspection rather than dispatch an emergency crew.

Later, the second pump began showing similar signs. The technical finding was much the same, but the operational meaning was not. With both pumps affected, the station was losing its redundancy. The finding was escalated and the utility sent a maintenance team. Both pumps were lifted, partial blockages were found on both and the obstructions were cleared before either pump blocked fully.

The blockage diagnosis had not changed. What changed was the station’s ability to cope with it. The same is true of flows, upstream storage and weather. Each can change the consequence of an asset problem without changing the fault. Heavy rainfall makes lost pumping capacity far more serious than the same loss in a dry week. At that point, it stops being a maintenance decision. Asset health has become an input to operating the network.

That last call belongs with the utility. Users may understand the equipment and the developing condition, but the control room knows what taking it out of service means for the network that day.

Earning the Right to Interrupt

Every alert has a cost. A site visit consumes travel time, skilled labor and often specialist equipment. Better information before dispatch changes whether a visit is needed, how urgent it is and what the crew should bring.

False alerts consume the same scarce resource. No one is completely immune. There have been flagged changes that turned out to be a repaired pump settling into a new normal, or there could be a deliberate change in how a station was operated that nobody had told the monitoring team about.

Much of the discipline comes before a finding reaches the utility. A model can surface an unusual pattern. An experienced analyst can challenge it, look for alternative explanations and decide whether the evidence is strong enough to put in front of a user. When something is missed, the outcome feeds back and the missing context gets added. Operators and field crews will not keep acting on findings that repeatedly lead nowhere.

A finding also has to arrive where work is actually managed and be specific enough to act on—which asset, the likely problem, how urgently somebody needs to look, etc. If an operator has to open another dashboard, interpret an anomaly and work out what it means, the technology has created a new demand on attention instead of relieving one.

Learning From What Happened

The blockage case did not end when the pumps were cleaned. After they returned to service, monitoring showed their behavior going back to normal. The inspection had confirmed the diagnosis. The signals afterward showed the expected recovery.

An unusual electrical signature is useful. An electrical signature tied to a physically confirmed blockage is much more useful. The same applies to bearing wear, impeller damage and hydraulic conditions. Confirmations and misses alike are what improve the next model, the next diagnostic rule and the next judgment an analyst makes.

Building that evidence takes time. Every connection between a signal and a fault needs a real operating condition and a field outcome to test the diagnosis against. The sensor data alone cannot confirm whether a diagnosis was right. Someone has to open the equipment. But asset health only goes so far. Knowing a pump is deteriorating does not tell an operator whether to take it out of service. That depends on redundancy and on what is happening elsewhere. Those trade-offs are already the control room’s job. Operators weigh flows, levels, available capacity, weather and other constraints to decide what can wait.

From Filtering Signals to Filtering Decisions

More autonomous wastewater operations will need reliable information about asset condition, hydraulic conditions, available capacity and the limits the network must never cross. They will also need users to understand the decisions people are making today.

This is why the analyst role can be seen as part of the path to autonomy rather than a stage to be removed from it. Reasoning that stays in someone’s head cannot be automated. Once it is made explicit and repeatedly tested against field outcomes, it starts to become clear which parts are consistent enough to encode. The same is true of the calls operators make in the control room.

Take a two-pump station. Today, monitoring may identify a developing problem. An analyst reviews the evidence, and the utility decides whether to change the duty cycle or take a pump out of service. The best-understood parts of that sequence can move into bounded, closed-loop control—changing pump sequencing, or adjusting speed within predefined limits where variable-speed control exists, then observing the result and reversing the change if conditions move outside those limits. Combining that information with hydraulic context could eventually support reliability and energy decisions together.

Start with decisions that are well understood, easy to constrain and easy to reverse. Prove they work. Then widen the envelope. As more routine decisions become safe to automate, skilled attention can move to the exceptions where human judgment matters most.