Case study

Eight drone photos of a 1.8 MW rooftop plant.

Public thermal photos of a rooftop plant in Catanzaro, Italy, put through the app as it is. It shows the one thing a single panel picture cannot: panels that are evenly warm.

The facts

Pictures
8 full thermal photos
DJI Zenmuse H20T
640 by 512 pixels each
Checked
1,485 panels
16 grids
Marked
162 evenly warm panels
51 panels with a hot part
10 groups of neighbours

One of the eight photos, drawn white hot from the camera's readings, with the app's marks. Blue is evenly warm. Orange is a hot part or a suspected fault. Lime needs review.

The run

Overview

A real roof, and no answer sheet

This is a public set of pictures, put through the app as it is. The plant, the author and the author's organisation are not PanelProof customers, took no part in this and have no connection with PanelProof.

The photos come with no marks of any kind: no panel outlines and no list of faults. Their author says they show hot spots and panels that are not working, but not which photo shows which. So this page shows what the app said. It cannot show how right the app was.

The question

Can the app see a panel that is warm all over?

A panel that has stopped working is evenly warmer than the ones beside it. Cut out on its own, it looks like any other panel. On another test set the fault finder found 8 of every 100 such panels. On a full photo there is something to compare it with: its neighbours.

What we did

The same four steps a customer takes

  • Uploaded the eight camera files. The app read the sensor's own reading for every pixel and drew each photo white hot.
  • Placed 16 grids: four corners, a number of rows and a number of columns for each block of panels.
  • Let the app cut out and score 1,485 panels and compare each with its neighbours.
  • Read the result. No panel was reviewed by a technician.

What the app said

Photo by photo

Opening the eight photos

The finding

The neighbour check saw what the fault finder did not

162

evenly warm panels marked by the neighbour check

15

of those that the fault finder alone called a suspected fault

51

panels with a hot part

49

of the hot part panels that the fault finder also called a suspected fault

The two checks see different things. The fault finder is good at a pattern inside one panel. The neighbour check is the only part that can see a whole panel, or a row of panels, running warm. 150 of the evenly warm panels sit in 10 groups of three or more in a line. The app labels a group "several panels in a row not working, check the string".

The review list

A long list, and the app said why

The app sent 289 panels as suspected faults and 323 more for review. With 26 audit panels that is 638 of 1,485 for a person to look at. It named 6 of the suspected faults "bypass diode active" and left 283 as "type not clear".

The app's own warning on this inspection. This batch scores high overall. Either the camera or the site is unlike the pictures the app learned from, or many of these panels are faulty. The app cannot tell which from the pictures alone.

What it cannot tell

Three limits of this run

  • How many of the marks are right. That needs a technician to mark the same photos.
  • Any temperature. The app worked from the camera's raw readings and printed no degrees.
  • Why a group of panels is warm. "Check the string" is a pointer for the person on site, not a diagnosis.

Credit

Whose photos these are

Danilo Pratticò, University of Reggio Calabria. "UAV radiometric thermal imagery of PV module faults, sample dataset (1.8 MW plant, Catanzaro, Italy)", version 1, Mendeley Data. Licensed under Creative Commons Attribution 4.0 International.

Shown changed: redrawn white hot from the camera's readings, cut into panels, with PanelProof's marks on top. Full credits and links.