New Toshiba AI only needs a single photo

Toshiba Corporation has developed an inspection information management AI for equipment inspections.

The company claims the management AI can recognise the location and signs of deterioration in equipment — including the size of a crack — through analysing a single photo taken with a camera.

The system operates through combining two of the Toshiba’s technologies: position-recognition AI which identifies locations from images, and 3D-recognition AI that estimates sizes.

Toshiba aims for the new AI to contribute to the efficiency of maintenance inspections at power plants and other facilities with weak or no GPS signals.

Traditional inspections and maintenance of infrastructure start by photographing and recording the condition of the site and equipment.

GPS data is then used to record equipment locations in facilities such as power plants. However, when signal is weak or non-existent, inspectors generally photograph signs of deterioration and cracks, and manually record the location and measure the size.

They then have to be organised, and the photographs associated with a floor plans. It is a time-consuming process, and is possible for error to creep in — such as an employee forgetting to measure cracks.

There are existing methods for identifying indoor positioning including installing wireless base stations and the use of markers, but the former solution is expensive, and when markers are used both the area and marker need to appear in the same photograph.

Toshiba’s inspection information management AI automates the task, and enables information sharing in remote work.

Photographs captured by inspectors, robots or drones are uploaded to a server, and managed as a batch. The AI looks for and recognises the location of images, the sizes of cracks, and any other signs of deterioration.

The process

First, photographs taken in the infrastructure facility are used to create a positional database that the location-recognition AI uses to link photographs with locations on a floor plan.

The AI uses a deep neural network model to recognise imaging directions and spatial positions on the floor plan from camera images and, according to Toshiba, is highly accurate even with low-resolution images.

Deep learning is used to analyse image blurring that occurs due to distance from the subject — allowing the 3D recognition AI to measure distances.

Having the images — and the additional information generated — on an online database allows inspection information to be stored with online floor plans without additional equipment. Accumulated inspection information can then be obtained by accessing the relevant floor plan.

In the future, the company aims to expand the scope and utilisation of the inspection information management AI by adding anomaly-detection, meter-reading and lifecycle prediction through accumulated data examination AI.

Alongside power plants, Toshiba aims to expand application to other structures where GPS is unavailable or unreliable — warehouse or bridge inspections.

Toshiba plans to release the AI via its IoT platform for energy systems, with practical application aimed for 2022.