Models used

In this section, we will explore the technical aspects of the models used in the project. Although I do not personally participate in their development

The Machine Learning and Deep Learning models used for the implemented features include time series forecasting models and autoencoder models.

Time Series Forecasting

Illustration de la prévision de séries chronologiques

Source : Lightning AI

Time series forecasting models allow for predicting future values over a given period. They are developed from historical data to predict future values.

A time series is a set of data points ordered in time, such as temperature measurements taken every hour during a day.

The predicted values are based on the analysis of past values and trends, assuming that future trends will be similar to historical trends. It is therefore crucial to examine trends in historical data.

The models make predictions based on a window of consecutive samples. The main features of the input windows include the number of hourly points and their labels, as well as the time offset between them. For example, to predict the next 24 hours, one can use 24 hours of historical data (24 points for windows with an hourly interval).

Auto-encoder

Example autoencoder

Source : IBM

Autoencoder models are neural networks designed to copy their input to their output. They first encode the input data into a lower-dimensional latent representation, then decode this representation to reconstruct the data. This process allows for compressing the data while minimizing the reconstruction error.

In our case, these models are trained to detect anomalies in the data. Trained on normal cases, they exhibit a higher reconstruction error when faced with abnormal data. To detect anomalies, it is sufficient to set a reconstruction error threshold.

Conclusions

These models play a crucial role in the project by enabling accurate predictions and anomaly detection. Although I did not directly participate in their development, this internship offered me a valuable opportunity to deepen my knowledge of machine learning and understand the importance of these models in real-world applications. These tools not only help anticipate future trends but also identify unusual behaviors, contributing to more informed and proactive decision-making, which is relevant for IoT.