Guest contribution by Cinn Song, founder of SOLARTODO, a solar street lighting manufacturer.
Editor’s note. This is an unpaid guest contribution from a manufacturer, with no product names or links in the body. I’ve added the sections on battery state and the cellular layer, because on a solar column running over 4G those are the first two places I’d look. Anyone who has done call-outs to remote kit knows the most expensive fault is the one you diagnosed from the dashboard. Peter Green
Anyone who runs a remote lighting estate has seen it. A resident reports a dark column, the dashboard says the lamp is on, and someone has to decide whether to send an engineer, and which one. Get it wrong and you send a lighting crew to fix a modem, or a comms engineer to a failed driver. Either way you’ve paid for a site visit that fixed nothing.
The cause is usually the same: the dashboard is answering a different question from the one you’re asking. This guide sets out how to tell a telemetry outage from a lighting fault using evidence most estates already hold: timestamps, session behaviour, cellular records, battery data and driver diagnostics. It uses MQTT 5.0 and DALI as worked examples because both are publicly specified, not because every estate uses them.
Three questions, answered separately
Before you classify anything, split the problem in three.
- Can I hear the column? Is the communications path working, or has something between the controller and the server gone quiet?
- Is what I’m seeing current? Is that “lamp on” tile a fresh observation, or the last thing the column said before it went silent?
- Is the light actually failing? Is there evidence, from the controller, the driver, the schedule or someone standing underneath it, that the lamp didn’t do what it should have?
The answers can share a cause. On a solar column, a flat battery can take out the lamp and the modem together. They can also diverge: a lamp can run its dusk-to-dawn programme perfectly while the uplink is down, and a healthy uplink can faithfully deliver a stale picture of a dead lamp. Answer each one on its own evidence.
A worked example
A hypothetical case runs through the rest of this article. It is an illustration, not a real installation.
An operations desk looks after around 80 solar columns. Shortly after dusk, a resident reports column SL-047 dark. The dashboard still shows “lamp on”. Its neighbours on the same gateway have also stopped updating; columns on the next gateway along are fine.
That pattern points at something shared: the gateway, its backhaul or its power. It doesn’t tell you the lamp has failed, and it certainly doesn’t tell you it’s still on. At this stage, communications look impaired, the displayed state isn’t current, and a lighting fault is possible but unproven.
On a solar column, check the battery first
On mains-fed lighting, power is somebody else’s problem. On a solar column, the lamp, the controller and usually the modem all run from the same battery, and that changes the diagnosis.
Charge controllers protect the battery with a low-voltage disconnect, and designs differ in what they shed first. Some dim or cut the lamp to keep the controller and modem alive. Others drop the modem early to save energy for the light. Either way, a column running short of charge can go quiet, go dark, or both, and none of that is a communications fault or a failed lamp.
UK winters make this worth checking first. Short days, low sun and a run of overcast weather can pull a battery down over a week without any single bad day to point at. If several columns sharing the same weather go quiet together around dusk, power is at least as plausible as a failed gateway.
The evidence is usually already in your telemetry, if you look at the trend rather than the last value:
- Battery voltage or state of charge over the past few days, not just the final reading.
- Daily charge from the panel, which shows shading, soiling or a spell of poor weather.
- When the column went silent. Dusk matters, because that’s when the load starts and the charging stops.
A steady decline that ends in silence around dusk points to power. Silence that arrives abruptly with a healthy battery reading points back to communications. Don’t turn that into a universal voltage threshold: the right levels depend on the battery chemistry and the controller, so use the manufacturer’s figures for the equipment actually installed.
When was this actually true?
A status without a reliable time is just a label. Two times matter:
- Observation time: when the controller or driver took the reading.
- Receipt time: when your server received it.
Store-and-forward, retries and batched uploads can put hours between the two. Show only the receipt time and an old reading looks fresh. Show only the observation time and you hide how long the path was silent. Store both.
RFC 3339 defines the timestamp format most systems use.1 A complete timestamp states its relationship to UTC: “Z” or “+00:00” for UTC itself, or a numeric offset such as “+01:00”. “-00:00” means something different: the UTC time is known but the local offset isn’t. A time with no offset at all is a poor basis for comparing records between systems, and in a country that changes its clocks twice a year, an unlabelled local time is an argument waiting to happen.
A correct format doesn’t make a correct clock. Outdoor controllers drift, boot with unset clocks, or stamp messages in a local time nobody configured. Flag any clock you can’t vouch for.
In our example, SL-047 last reported an observation time of 2026-11-07T16:10:22Z, received at 16:11:04Z, and nothing since. The honest reading is: “the column reported the lamp on at 16:10 UTC, and there has been no observation since dusk.” How old is too old is a freshness policy each asset owner should set per site, not a universal number.
What MQTT can and can’t tell you
MQTT 5.0 makes a good worked example because four of its features are regularly misread on dashboards.2
Keep Alive. The client agrees a maximum interval between control packets and sends a ping if it has nothing else to say. If the server hears nothing for one and a half times that interval, it closes the connection as though the network had failed. So “connected” means packets, possibly only pings, are arriving on time. A ping says nothing about the lamp.
Retained messages. A message published with the retain flag is stored by the broker and handed to anyone who subscribes later. That’s how a dashboard opened after dusk can show “lamp on” from a column that went silent at teatime. A retained message is the last thing the column said, not evidence that it’s still connected or still lit. Clearing it doesn’t reset anything at the column.
Message expiry. A publisher can give a message a lifetime in seconds; if it isn’t delivered in time, the broker deletes it. Without the property, a message never expires. An expiry on status messages that roughly matches your freshness policy stops stale states lingering on screen. It still tells you nothing about the lamp.
Last Will. When a client connects, it can lodge a Will message with the broker. If the connection drops without a clean disconnect, including a Keep Alive timeout, the broker publishes it. Paired with a retained status topic, it’s the simplest way to make a dashboard say “offline” rather than quietly showing the last good state. MQTT 5.0 also adds a Will Delay Interval, so a brief drop-out doesn’t raise an alarm if the client reconnects in time.
Lost the session, or lost the network?
MQTT tells you the broker lost the client. It doesn’t tell you why. On a cellular estate, the next question is whether the device lost its network or only its session, and the records for that sit outside the dashboard:
- Signal history from the modem or router: RSRP, RSRQ and SINR over time, not a single snapshot. A gradual decline suggests something physical, such as a new obstruction, a damaged antenna or water in a connector.
- Registration and data session state. A modem can be registered on the network with no active data session, which looks exactly like an outage from the server’s side.
- The SIM’s connectivity platform. Most IoT SIM providers show the last network attach, last data session and recent usage for each SIM. If the SIM attached and passed data after the dashboard went quiet, the problem lies between the device and your broker, not in the radio network.
- Local modem or router logs, if the device keeps them, collected on the visit.
One trap with LTE-M and NB-IoT: power-saving modes such as PSM and eDRX make devices deliberately unreachable for long periods. A device that sleeps longer than its Keep Alive will drop its session every cycle and look offline between reports. Design the reporting interval, the Keep Alive and the freshness policy together, or the dashboard will cry wolf every night.
Evidence from the driver
A commanded “on” isn’t the same as light on the road. Where the gear supports it, the controller or LED driver can report more directly on the lamp, and that evidence has to be read against what the schedule or photocell should have been doing at the time. A lamp off at 9pm in November is a fault candidate. A lamp off at 2pm is working.
DALI Part 253, Diagnostics & Maintenance, is a public example of what a supporting driver can expose: failure conditions for the control gear and lamp, plus operating time, start counts, supply voltage, temperature, output current and similar data.3 For DALI-2 certification, Parts 251 to 253 are optional; D4i certification requires all three. Even within Part 253, some fields are optional.
So treat driver diagnostics as useful when present and silent when absent. A missing fault field is not evidence that the fault isn’t there.
In our example, a current lamp-failure flag from SL-047’s driver during burning hours would be strong lighting evidence. A retained “lamp on” from 16:10, with no diagnostics since and a silent gateway, is not.
Reading the evidence together
| What you see | Most likely | Check first |
|---|---|---|
| One column silent, neighbours fine, battery healthy at last report | Device comms: modem, SIM or antenna | SIM platform session history, signal trend |
| Several columns on one gateway silent, batteries healthy | Shared path: gateway or backhaul | Gateway power and backhaul |
| Silence around dusk after a falling battery trend, often several columns | Power, usually weather-related | Charge history, panel shading or soiling |
| Reporting normally, driver lamp-failure flag during burning hours | Lighting fault | Driver and lamp on site |
| Reporting normally, no fault flag, resident says it’s dark | Unconfirmed: schedule, photocell or unsupported diagnostics | Schedule configuration, then a visual check |
The handover ticket
The aim isn’t to finish the diagnosis from the control room. It’s to give the engineer a ticket that separates what’s known from what isn’t.
- Asset: column and controller ID, location from the asset register, gateway and firmware. Leave blanks rather than guess.
- Times: observation and receipt times for every value relied on, with UTC offsets, and whether the clock is trusted. If nothing has arrived since a given time, say so.
- Comms: session state, last packet seen, whether the displayed state came from a retained message, and what the SIM platform shows.
- Power: battery trend and recent charge history.
- Faults: any controller or driver alarms with times, alongside the schedule that should have applied. If there are none, say diagnostics are unavailable, not that the lamp is fine.
- Uncertainty: what the desk can’t conclude. For SL-047: “Gateway comms impaired. Lamp state last known 16:10 UTC, not current. Battery trend falling. Lighting fault possible, not confirmed.”
That leaves the engineer the jobs only a site visit can do: confirm whether the column is lit, read the local indicators and check the supply against the as-built record, under the asset owner’s method statement. Remote resets, isolation procedures and numeric limits are deliberately out of scope here.
Done this way, a silent modem doesn’t get written up as a failed lamp, a flat battery doesn’t get blamed on the network, and a stored status doesn’t get mistaken for a live one.
References
1. Klyne, G. and Newman, C., Date and Time on the Internet: Timestamps, RFC 3339, July 2002.
2. Banks, A., Briggs, E., Borgendale, K. and Gupta, R. (eds.), MQTT Version 5.0, OASIS Standard, 7 March 2019.
3. DALI Alliance, DALI Data (overview of Parts 251, 252 and 253).
About the author. Cinn Song is founder of SOLARTODO LIMITED, a manufacturer of solar street lighting and smart infrastructure systems. solartodo.com
Disclosure. This is an unpaid guest contribution. The author is affiliated with SOLARTODO. AI tools assisted with drafting and source research, and the author is responsible for the final text. Edited by IoTPortal, including additional material on battery state and cellular diagnostics.



