Artificial intelligence (AI) with Tensorflow for building PV remote monitoring

The automation of construction processes, as well as the need to remotely monitor large projects due to mobility restrictions, has led to the improvement of technological capabilities and the development of new products based on the use of Artificial Intelligence (AI) applied in Industry 4.0, which will benefit the sector in its digital transformation process.

In this article we give an overview of how to implement an Artificial Intelligence (AI) model using Tensorflow to automatically detect the progress of works in the execution of photovoltaic projects.

This new approach tracks the in-situ evolution of works, based on deep learning techniques, using Tensorflow, an open source machine learning framework, which makes implementation flexible and increases development capabilities. Tensorflow facilitates the implementation of the neural network on various types of devices (embedded and mobile devices, mini computers, etc.). In addition, it supports different types of neural networks that can be tuned and retrained for particular purposes. The results presented are promising, since the retrained networks identify each of the phases of progress, with this information the system can be controlled to correctly follow the initial schedule of activities.

How does it work?

The model is based on a series of images from IP video cameras (video) located on-site, which are used to train an AI model that performs for each dataset an analysis on more than 10,000 images. This allows to classify in real time the progress of construction activities in 3 phases or stages of the PV project.

A machine learning model created with Teachable Machine, a tool that makes it easy for anyone — teachers, students, artists, creators of all kinds — to train machine learning models, has been developed.

Structure of the learning model by classes or categories.

Parameters

Learning rate of = 0.00101 Cycles or epochs = 60 number of lots / images = 16

The model is based on a series of images from IP video cameras (video) located in-situ, which serve to train a model of IA which performs for each dataset an analysis on more than 10,000 images. This allows to classify in real time the progress of the construction activities in 3 phases or stages of the PV photovoltaic project.

  1. PHASE 1. Poles and Base works
  2. PHASE 2: Assembly of structures
  3. PHASE 3: Solar Panel Instalation

The model can calculate the pressure for each class, in order to adjust the training images as needed:

Confusion matrix

During training we calculate the accuracy or classification percentage of the model. Assuming only results above 0.70:

Precision by epoch

We also calculate the learning loss or learning level of the model, in order to predict the correct classification for the data set:

Lost by epoch

The architecture of the model is based on the one used by TensorFlow.js: Link

Thanks to DS team.

Phd in Geographic Information Technologies and Visiting Scientist. Bonn Office, UN-SPIDER Programme United Nations Office for Outer Space Affairs (UNOOSA)