How PanelProof AI works

One real inspection, followed from the file to the report.

Small models do the looking. Plain rules check them. A person decides. Here is every step, with the numbers it produced on a real roof.

The run

The pictures

Eight drone photos of a rooftop plant in Catanzaro, Italy

The photos are public. They were taken with a DJI Zenmuse H20T thermal camera over a 1.8 MW rooftop plant and published by Danilo Prattico of the University of Reggio Calabria under the CC BY 4.0 licence. The plant, the author and the university are not PanelProof customers and took no part in this.

Nobody has marked which panels on these photos are faulty. So this page shows what the app said, step by step. Where a step has been measured against marked pictures, the measured numbers are given beside it.

8

photos

16

grids placed

1,485

panels cut and checked

638

sent to a person

Step 1

Read the file

Each file holds a picture and, beside it, the sensor's own reading for every one of its 640 by 512 pixels. The app reads those numbers out with its own code and draws a white hot picture from them: the hotter, the whiter. The colour scheme the camera was set to does not matter.

A picture in a colour scheme with no readings behind it is turned back, with a message that asks for the original camera file.

One of the eight photos, as the app drew it. Bright means warm.

What it does not do. It prints no temperatures. Turning these readings into degrees needs the camera maker's own software, which the app does not use.

Step 2

Cut the panels

A technician clicks the four corners of each block of panels and types its rows and columns. The app lays a grid over the block, cuts out every panel, trims a small margin so the frame does not leak in, turns the panel upright and resizes it to 24 by 40 pixels. That is the size of the pictures the model learned from.

Two grids on this photo: 11 rows of 11 panels and 11 rows of 10.

The first panels cut from it, in order.

Step 3

Check the picture

Before a panel picture is judged, the app turns back what it cannot use: a file that is not a picture, a picture in a colour scheme, a picture too small to show a panel.

It also repairs one thing. Some panel pictures are saved with pure black corners round a slightly turned panel. The model reads those badly. The app straightens the panel and cuts the corners away. It did that for all 3,675 pictures of a second public set.

Panels from that set after the repair.

Step 4

Score every panel

The fault finder is a small model with 437,676 adjustable numbers. It learned from 13,994 labelled panel pictures and gives every panel one score. It runs on an ordinary processor.

Measured on 3,002 pictures that were set aside before training:

92.2

faulty panels found in every 100. Likely range 90.8 to 93.5

39.4

normal panels flagged in every 1,000. Likely range 30.7 to 50.5

95

of every 100 faults that show a pattern

61.9

of every 100 evenly warm panels

On this photo: the panels it scored far above the rest of the batch.

Every number, with its range

Step 5

Name the type, when the name has earned it

The app may name a fault type only if that type passed a test: its name must be right at least 85 times in 100, and it must catch at least 80 of every 100 of its kind. One of five passed.

Fault typeName is right, in 100Catches, in 100May the app name it?
Hot cell or hot spot8476.5No
Bypass diode active99.692.8Yes
Whole panel not working90.351.6No
Cracked83.176.1No
Something blocking the panel81.671.3No

So on this roof 6 suspected faults were named "bypass diode active", and 283 read "suspected fault, type not clear". The reviewer may choose any type.

Step 6

Compare each panel with its neighbours

A panel that is evenly warmer than its neighbours looks normal in a picture of that panel alone. So on full photos the app lays a smooth surface through each block of panels and asks two things of every panel: is the whole panel warmer than its place says it should be, and does a small part stand far above the rest of it?

162

evenly warm panels marked

150

of them in 10 groups of neighbours

51

panels with a hot part

15

of the 162 that the fault finder alone had called a suspected fault

A hot part, or a suspected faultEvenly warm

Not measured. No technician has marked these photos, so we cannot say how many of these marks a technician would agree with.

Step 7

Sort the batch for review

A fixed pass mark fails when a new camera's pictures are unlike the training pictures. So the app measures each batch against itself and sorts it into three piles. A batch needs at least 200 panels for this. A smaller one is listed whole, most suspicious first.

289

suspected faults

323

more that need review

873

not flagged

26

of those sent to review anyway, as an audit sample

Measured on marked pictures, the rule was aimed at 50 false alarms in every 1,000 normal panels. It gave 86 on pictures like the training ones and 75 on another camera and site, where a fixed pass mark gave 781. It is steadier than a fixed pass mark. It is not exact.

The app does not say "clear". A panel it did not flag is "not flagged". On the other camera, 11.5 of every 1,000 panels it did not flag were faulty, nearly all of them evenly warm panels.

Step 8

A person decides

Every flagged panel waits for a person. The reviewer sees the panel as the app saw it, its place among its neighbours, and what the app said. One key confirms, changes or rejects. A new decision is added to the record and never erases the old one.

What the app said

Photo 3, row 4, column 6

Suspected fault, type not clear

It stands 14.3 spreads above the ordinary panels of this inspection.

Your decision

No fault NSuspected fault FCannot tell C

Step 9

Sign the report

A report cannot be signed while a flagged panel has no decision. When the lead signs, the app builds a PDF and a CSV with one row for every panel, records a fingerprint of each file, and never changes them. A correction makes a new version and keeps the signed one.

Thermal inspection report

Rooftop plant, Catanzaro

Findings are suspected faults seen in thermal pictures. This report is not a safety certificate and is not a statement that the inspection meets any standard.

PicturesFull thermal photos, cut into panels with the grid
Panels checked1,485
Flagged for a person638
Not flagged by the app847
Photo 3, row 4, column 6Suspected fault, type not clear
Photo 5, row 11, column 2Suspected fault, type not clear
Photo 1, row 2, column 12Suspected fault, type not clear

Model pp-crops-0.1.0-level-camera. Review rule review-chooser-0.1.0.

Get started