Europe Union

Automated Fibre-Optic Installation Validation with Computer Vision and Machine Learning

Automated Fibre-Optic Installation Validation with Computer Vision and Machine Learning
Learn how our AI Engineers helped a leading Dutch fibre-optic provider to save time and money on infrastructure installations.

Our Client is a Leading European Fibre Optics Provider

Our client is one of the leading software development providers delivering digital solutions for utility companies in and outside the country, extending its services to neighbouring countries such as Germany. As one of the biggest providers in the country, they hire external companies to carry out the installations at homes and public institutions.

Manual Installation Checks Drained Resources and Wasted Time

To ensure that the installations are carried out correctly and to prevent any additional challenges before they happen, they required the work to be verified in multiple stages based on photos of the installation. Manual verification was tedious and time-consuming, involving many installations within various neighbourhoods and districts. 

They’ve determined that the best solution to this challenge would be to automate the verification process to save time and resources. That’s where our experts came in.

AI for Automated Fibre-optic Installation Verification

Our Machine Learning Expertise Was The Missing Ingredient

Our client met us at our discovery meeting with complete knowledge of what they wanted to achieve business-wise. However, since their company didn’t deal with the required technologies, they hired us to design a complete AI-based solution. The solution’s task was to analyse and automatically verify the photos at two stages of an installation:

An open box shows if the wires are installed correctly and intact.

A closed box to confirm the final stage and ensure the box was correctly mounted on the surface.

We had a discovery meeting that helped us understand the expectations and develop a method to achieve them. Our experts proposed a deep learning model to recognise the correct and incorrect states of the devices. Such a model would process and analyse a picture uploaded by the contractors and verify if it meets the conditions for a correct installation.

To train such a model, we needed quality data from photographs of correctly and incorrectly installed equipment. Our client expected the rate of correct verification to be 60-70%, corresponding to the approximate level of manual verifications. We ultimately exceeded these expectations.

How To Extract a High-Quality Training Data Set from Millions of Images for Deep Learning?

Once we aligned our vision with the client, we were ready to start working on the project. After designing the model’s base, it was time to feed it data for training.

To help us achieve this, our client prepared an impressive amount of 4TB of data, totalling several million photographs. However, since it was sensitive in-action data showing actual customer installations, it had to be transferred securely. To deal with that weight, we employed several technologies to connect cloud drives and access data without risking errors and unsuccessful transfer attempts. Making a training set from such a huge number of photographs would require too much resources and time. Our data analysts chose a more lightweight, representative set of 10,000 images to make it easier.

How to Use a Lightweight Dataset for AI Model Training?

We divided the data set into smaller parts and used the cross-validation method for more reliable validation metrics. In every iteration, we fed the model with different, non-repetitive parts of the data, and the rest was used for evaluation. Metrics from all iterations were aggregated to obtain accurate model performance indicators on unseen data. 

Our methods proved highly efficient, and we achieved much higher accuracy than expected, namely 90%.

It was possible thanks to the extensive expertise of our team and the quality of data our client provided.

Grzegorz Ozymko Grzegorz Ozymko Machine Learning Project Manager at DAC.digital

We Used Explainable AI to Back Up the Accuracy with a Showcase of the Model’s Workings

After achieving considerable success on the accuracy level, we’ve decided to strengthen it by showing the client a sneak peek behind the scenes. Explainable AI (XAI) was implemented to show how the model looks at the images and which elements it considers during the inference.

What is Explainable AI, and How It Works?

Explainable AI allows the detection of any inconsistencies in the model and fixing them at an early stage. XAI is about creating models that can show their work. It aims to make the decision-making process of AI systems transparent, understandable, and interpretable. It enables the AI to provide reasons for its output, such as whether it thinks a particular image is a correct installation.

Adding explainable AI also helped our client understand the model’s rules and assess whether it included every essential element of the photograph. It reassured them that the model was trained well and that its accuracy was consistent and reliable.

What is Explainable AI (XAI)

Working from the Inside Out: How We Built an Additional Model for Outdoor Verifications

Our collaboration continued after ensuring the box verification model and service were finished and accurate. Using similar processes and methods, we built an additional model to verify the outdoor installations – the photographs of wire spools installed in the ground.

Our Process for Data Engineering and ML Model Training

Photo set icon

Our data analysts chose over 10,000 photographs to use as a training set for the model.

Machine learning icon

After carefully selecting the data set, our machine-learning experts used it to train the first model, using cross-validation for reliable metrics. It was a success, exceeding expectations with 90% accuracy.

AI photo icon

To show the client how the model works and build trust that the image classification is based on the boxes’ regions, we implemented explainable AI to show which image parts are considered and ensure that it covers all the critical elements.

Web Settings icon

We built and integrated a web application that allows the client to upload pictures and analyse them for predictions on the box type, its status (open or closed), and the state of the photograph (correct or incorrect).

Machine learning icon

We created the second machine-learning model to recognise and validate wire spools in the ground.

One of the most formidable challenges was selecting a representative training set covering all relevant cases. However, I’m very proud of the final effect and the fact that we vastly exceeded the client’s expectations in terms of accuracy.

Jan Glinko Jan Glinko Machine Learning Researcher at DAC.digital

The Faces Behind the Success: Our Expert AI and ML Team

Our team included two data analysts who ensured a selection of quality data from a vast client-provided batch. Two machine learning engineers designed and trained the validation models with outstanding results. They were supported by MLOps and DevOps engineers who helped with implementation and service development. A project manager coordinated the whole team, and we also had one point of contact to streamline the communication with our client.

Technologies for Successful Photo Validation Solutions

  • Python: the primary programming language
  • PyTorch: Deep Learning
  • Hydra: configuration file management
  • GradCam: Explainable AI
  • FastAPI: services
  • OpenCV: image processing
  • Numpy: matrix operations
Python
PyTorch
FastAPI
NumPy
Hydra
OpenCV

Results:  Solid AI MVP Preparing a Successful Product Launch

Using computer vision and machine learning for the project was a success. We have completed the project in both stages – validation for indoor boxes, outdoor excavations, and optic-fibre spools. We used expertly trained machine learning models fed with large amounts of data our client provided, making it reliable and accurate at 90%, which was more than initially expected. 

The client can now test the complete MVP in the wild and compare the results with the manual photo validation. So far, the results provided by our machine learning-based solution are promising and exceed expectations. The feedback was very positive, and the company executives are preparing to roll out our implementation in the upcoming months. 

Once live, the new innovative photo validation process will be completely automated, helping our client save enormous time and resources. It will also ensure consistent accuracy and eliminate the risk of human errors.

Machine Learning and Computer Vision for Photo Validation

Automate Your Business with Our Expertise in AI and Machine Learning

Do you plan to take your business to the next level and need technological expertise to carry it out? We will happily share our experience and build innovative solutions to transform your operations. Don’t hesitate to contact us.

Build your technological future. Let’s discuss the details.

Contact us

Read More of Our Case Studies.

Let’s connect!

Send us an e-mail: [email protected]