Thermal imaging and continuous ML monitoring address different parts of solar fault detection. One uses an IR camera flown over the array; the other analyzes SCADA data over time. Treating them as competing choices can lead to poor purchasing decisions and weaker operations. The practical question is what each method can detect, what it cannot, and when it belongs in the O&M process.
In brief, thermal imaging is a periodic, physics-based inspection that identifies active cell-level hot spots in one snapshot. Continuous ML monitoring analyzes behavior and detects gradual degradation trends over time. Each finds different fault types, so the decision depends on the failure modes involved and when each tool is useful.
What Thermal Imaging Actually Detects
Infrared thermography detects temperature differences between cells, strings, and modules. A failing diode can bypass a cell, while localized shading can force part of a module to carry excess current through bypass diode operation. That excess energy becomes heat. The temperature difference corresponds to the dissipated power and appears in a thermal image taken under high irradiance.
Thermal imaging reliably identifies bypassed cells with a hot spot from diode activation, cell-level cracking with increased local resistance, substring defects caused by manufacturing or physical damage, and junction-box connector resistance anomalies when enough heat reaches the exterior to register.
Useful results require irradiance above roughly 600 W/m2, low wind speed, which removes heat and narrows temperature differences, and a calibrated thermal camera with suitable spatial resolution for the panel size and drone altitude. At a 30 MW plant with 60,000 panels, a drone campaign takes approximately half a day to fly and another day to analyze with professional processing. A complete survey typically costs BRL 40,000 to 150,000, depending on plant size, access difficulty, and report detail.
Most importantly, thermal imaging is a snapshot. It shows the array's condition on the flight date and under that day's conditions. It provides no information about what occurred between campaigns.
What ML Monitoring Detects
Continuous ML monitoring uses SCADA telemetry differently. It tracks the electrical behavior of each string and inverter, establishes behavioral baselines for irradiance, temperature, time of day, and season, then searches for deviations that continue beyond normal variance.
It reliably detects fault classes that appear as gradual output loss: rising connector series resistance, partial shading caused by vegetation growth, bypass diode failures that reduce output before creating a hot spot, soiling buildup, and declining MPPT tracking efficiency. These conditions develop slowly, creating behavioral drift over days to weeks before a visible thermal signature or threshold alarm appears.
It does not reliably detect physical damage such as cracking, delamination, or broken glass unless the damage reduces electrical output. It can also miss cell-level defects that bypass enough of a cell to generate heat without materially affecting string output. One module with a bypassed cell in a string of 20 series-connected modules might reduce string Voc by 5 percent, within normal variance, while showing a clear hot spot in thermography.
Where Each Method Adds Value in a Real O&M Program
Periodic thermal imaging is most useful at these decision points:
- After commissioning, to establish a baseline and find installation defects such as improper stringing, connectors damaged during installation, or cell damage from transport or racking.
- After hail or a storm, when physical cell damage is the main concern and must be documented for insurance claims or warranty arguments.
- During warranty periods, when thermal evidence may support manufacturer replacement for cell-level defects.
- Before acquisition, when a current snapshot of array condition is needed for valuation and risk assessment.
Continuous ML monitoring supports the ongoing operational mission: finding performance losses between campaigns, identifying strings that have drifted below normal without triggering threshold alarms, and supplying early warning for planned rather than reactive dispatch. It is also the only method that provides historical context, showing whether an anomaly is new or has been present for six months.
Where They Overlap and the Risk of Choosing Only One
This is where poor decisions often occur. A program based only on annual thermography and threshold-based SCADA alerts can miss months of gradual loss between campaigns. A program based only on ML monitoring can miss cell-level hot spots that have not yet materially reduced overall string output.
Consider a string with one module showing 15 percent reduced Isc from a developing bypass diode fault. String current falls by roughly 0.7 percent (15% of 1/20 modules in series), which sits within normal string-level variance. ML monitoring may not flag it reliably until the fault worsens. Thermography in the next campaign, however, would clearly show the hot spot. That hot spot also presents a fire risk that increases with time and operating temperature.
A different case is a plant surveyed by thermography six months ago, followed by a contractor vehicle crossing a cable run and damaging insulation on two string cables. Last year's images cannot reveal the resulting ground fault or rising series resistance. Continuous monitoring is what catches the problem between campaigns.
How to Calibrate Thermal Campaign Frequency
The appropriate campaign frequency depends on plant age, warranty status, local hail and storm risk, and the O&M team's tolerance for cell-level defects. Annual thermography is defensible for a plant under warranty in a region with meaningful hail risk. For a mature plant in a stable climate with reliable SCADA data and continuous monitoring, every two to three years is often reasonable, with triggered campaigns after significant weather events.
Reducing campaign frequency is not only about lowering cost. Fewer campaigns can make each comparison more informative. Reviewing the current survey against the previous one shows which defects developed, progressed, or resolved, creating trend information unavailable from one snapshot. Two campaigns six months apart reveal more than two campaigns a year apart with no intervening context.
Integration in Practice
The most effective O&M programs use thermal imaging to verify specific fault hypotheses and ML monitoring as the continuous early-warning system that produces them. Suppose ML monitoring finds a string group in the northwestern array with a progressive 4 percent generation gap over the past three months. The next thermography campaign can prioritize that area. The image may confirm a developing hot spot or rule out cell damage, directing the investigation toward connectors or cabling that do not create a thermal signature.
This does not mean every plant needs both methods at all times. A small commercial rooftop may not justify periodic drone campaigns. A large ground-mount plant with complex terrain and significant weather exposure likely should not depend on SCADA monitoring alone between annual campaigns. The choice depends on the asset, its environment, and the operating entity's risk profile. These tools are not substitutes; they complement one another by addressing different parts of fault detection.
For more context on how ML-based string fault detection works in practice, see the piece on spotting PV string faults before failure.