Want to know more?

We realize algorithms and models that, thanks to Optimization and Artificial Intelligence, create innovative solutions to solve problems.

AI & Optimization Software and Applications

We want to analyze the weather forecasts to make an estimate of the wind energy offers that the various operators make to the "next day market", in Italy.

**The Challenge**

Predict the offer strategies of the operators and plan an offer taking into account bonuses and penalties that the market regulation foresees according to the demand/supply ratio.

**The Solution**

We use clustering techniques to reduce the dimensionality of the problem by capturing useful information from weather forecasts and train artificial intelligence models to model operator behavior based on "next day market" historical data.

**Results and Benefits** 

The analysis techniques adopted enable the planning of an "intelligent" bidding strategy.

It is possible to use weather forecast to anticipate energy market trends thanks to artificial intelligence: visit the website and find out how!

Renewable Energy Market Forecast

Forecasting fluctuations in the renewable energy market

The objective of this research project, conducted in collaboration with the Policlinico Gemelli, is the development of an automatic system for the analysis of EEG (ElectroEncephaloGram) signals in order to predict seizures in patients resistant to drug treatment.

**The Challenge**

The solution requires the capture and recognition of EEG signal changes typical of the onset of a pre-critical phase. This allows the necessary precautions to be activated to avoid or manage the actual seizure. One of the difficulties to be faced in the development of an automated solution is the strong noise of the EEG signals.

**The Solution**

We filter the EEG signals managing to extract the most relevant features through autoregressive models (AR). Based on the features thus highlighted we build classification models of type Support Vector Machines (SVM), which we train to distinguish between a normal phase and a pre-critical phase.

**Results and Benefits**

The predictive model helps to identify the pre-critical phase of a seizure.

Find out how a mathematical model uses electroencephalogram data to predict seizures onset and other adverse events: visit the website!

Forecasts of Adverse Events from Medical Data

Identify pre-critical phase from EEG data

Adverse Events Prediction

Predict the behavior of complex systems to take the best decisions

Forecasting

The objective of this research project is the development of a detection algorithm using thermal images, which have the advantage of preserving the privacy of individuals and allow operation even at night.

**The Challenge**

Training a neural network would require a large sample of annotated thermal images. Unfortunately, the availability of such datasets is extremely limited, especially when compared to the amount of data available for networks that, in contrast, use RGB images.

**The Solution**

We exploit domain adaptation techniques to adapt neural networks to operate on new domains. In this case we start from networks trained on RGB images and modify the first layers through a bottom-up process of convolutional network adaptation. In this way the new model is able to operate detection in the thermal spectrum. Furthermore we use a task-conditioned approach to automatically recognize if the image under examination is day or night allowing the algorithm to adapt to the specific task.

**Results and Benefits**

Our approach allows us to overcome the limitations imposed by the lack of labeled data. Transferring knowledge from a network trained on RGB data to the thermal domain, allows us to achieve performances similar to those achieved with the use of RGB images but with the advantage of preserving privacy and the possibility of use in night conditions.

Find out how we can train neural networks without data thanks to domain adaptation, to extend neural network usage. Visit our website!

Small Data Image Analysis Software

Adapt a neural network to a new domain without data

Image Analysis with Small Data

The objective of this research is to automatically update GIS vector maps by extracting meaningful information of the territory of interest, captured from SAR imagery.

**The Challenge**

The objective is to insert within GIS vector maps, spatial information (roads, terrain, forested areas, lakes, buildings, etc..) collected by synthetic aperture radar (SAR) equipped on aircraft, overcoming the problems of registration of images with respect to the map, the presence of artifacts in the images and the need to distinguish between many possible classes.

**The Solution**

We combine image processing techniques for segmentation and registration with machine learning techniques for classification of identified areas.

**Results and Benefits**

Our approach enables the creation of GIS maps with included information from SAR imagery, in an automated manner.

Learn how to automate the recording and classification of SAR satellite data with GIS maps: data fusion to the nth degree!

Data Augmentation on Satellite Images

Fusion of GIS maps with SAR images

The goal of this research project is to develop a convolutional model to robustly count the number of people present in images and videos of extremely dense crowds, using the least amount of data possible for training.

**The Challenge**

Training a neural network that can robustly count individuals typically requires a huge amount of annotated data. The process of retrieving this data and tagging it is particularly onerous and expensive.

**The Solution**

We develop a self-supervised algorithm based on an appropriate crowding degree sorting function. This allows us to train and adapt a convolutional network using very little human supervision.

**Results and Benefits**

The network trained using the self-supervised approach achieves, at a lower cost, a level of accuracy comparable to that of networks trained on large amounts of tagged data.

In the context of self-supervised networks, discover how it’s possible to train a neural network to count people even without large amounts of annotated data.

Self-Supervised Networks - People Counting

Self-supervised algorithms for people counting problems

People Counting

The goal is to monitor the traffic of passing vehicles within an area of interest (e.g. highway) through the installation of a set of cameras and other types of sensors (e.g. radar) capable of observing the area from different viewpoints and with partially overlapping fields of view.

**The Challenge**

The system needs to integrate information from multiple viewpoints and be able to detect all vehicles in the area of interest in real time, classify them, accurately locate them and track their movement. There is also the need to synchronize video streams, calibrate different viewpoints, and manage occlusions, all in real time.

**The Solution**

We integrated artificial intelligence and computational vision techniques for vehicle recognition and classification with multi-target tracking, stochastic filtering, and data fusion techniques for vehicle tracking.

**Results and Benefits**

Intervention methods enable effective vehicle identification.

Find out how Data Fusion and Computer Vision work together to effectively identify, locate and track vehicles. Discover more!

Vehicle Tracking via Computer Vision

Vehicle detection and tracking using multiple cameras

Vehicle Tracking

Gain powerful insights from images and videos through object recognition and tracking algorithms

Image Analysis

Adaptive optics is a technology that allows a deformable mirror to be used to correct optical distortion due to atmospheric turbulence in real time and is of critical importance in ground-based telescopes.

**The Challenge**

Mitigating atmospheric turbulence requires the design of a decentralized system to control the deformation of the telescope mirror, ensuring stability and performance in terms of response time and accuracy.

**The Solution**

We applied a data-driven approach that uses frequency profiles of turbulence and vibration directly obtained from experimental data. In this way it is possible to obtain a real-time response prediction and optimize the telescope mirror control system.

**Results and Benefits**

The solution allows a clear improvement of the telescope performance, thanks to the real-time stabilization of the mirrors even in adverse weather conditions.

Data-driven control: how to take advantage of experimental data and frequency analysis to optimize a system in real time. Find out more!

Data-driven Control System

Real time correction for optical distortion of deformable mirrors

The goal is to determine the real-time position and velocity of various moving objects based on Doppler travel time and displacement measurements provided by a passive radar.

**Problem Analysis**

The project solution requires analysis of the information received from the passive radar and exclusion of false and lost measurements, as well as analysis of the unknown association between measurements and objects the measurement from multiple (multi-static) sources.

**Method of Intervention and Solution**

We proposed a solution that combines optimization techniques for trace initialization with advanced multi-target tracking and sensor fusion techniques.

**Results and Benefits**

The methods that were used allow to obtain the effective identification of objects.

Object tracking in real time: find out how Data Fusion techniques allow you to identify and track objects using radar data analysis!

Real-Time Object Tracking

Object tracking using passive radar

Real Time Object Tracking

Leverage real-time data and forecasts to optimize control system decisions

Model Predictive Control

The goal is to develop a software able to support the operator in the ordinary maintenance operations by automatically identifying the engines with an higher risk and therefore higher priority.

**The Challenge**

In order to classify the risk category of the device we use the available data from the analysis of chemical and physical properties of the oil performed on samples taken at different stages of the engine life. The main difficulty is due to an incomplete and inconsistent data history.

**The solution**

We developed a classification algorithm on Support Vector Machine (SVM). To handle the missing data we built an ad-hoc modification of the SVM kernel was used to handle the missing data in order to maximize the separation between classes.

**Results and Benefits**

Our classifier is able to correctly recognize the risk category of the engine with more than 85% accuracy. In addition, the model is interpretable, thus allowing the identification of the chemical and physical parameters considered most relevant for classification.

Find out how to build a classification algorithm for predictive maintenance even in the absence of data. Trust the experience of Deix!

Classification Algorithms in Predictive Maintenance

Predictive Maintenance using engine oil measurements

Engine Classification

Monitor and predict the condition of machinery, products, and tools through data-driven and simulation based models

Predictive Maintenance

The goal is to design the double hull of a catamaran in order to reduce as much as possible the expected value of the average resistance in the field of irregular waves increasing, at the same time, the maneuverability of the boat.

**The Challenge**

We want to take advantage of the HF-URANS (High-Fidelity Unsteady Reynolds-Averaged Navier-Stokes) simulator to get very accurate results of the boat design performance. The challenge is to insert the simulation model into an automatic optimization procedure while keeping computational costs and time low.

**The Solution**

By training a network of radial basis functions, we build a surrogate model (called SRBF) of the simulator. This model has very low computation times but accuracies comparable to those of the high-fidelity model. We can thus exploit it within a global and multi-objective optimization method that, properly hybridized with local techniques, quickly produced an optimal solution.

**Results and Benefits**

The new simulator allows to support the design through the use of a virtual prototype for rapid simulation and an optimizer to identify the optimal parameters.

How to integrate a high-fidelity simulator and an optimizer to support virtual prototyping design? Find out more on our website!

Virtual Prototyping and Simulation

Design of the hull of an open-sea racing catamaran

Virtual Prototyping

We want to improve the reaction time (delay) to the on/off command of AMOLED displays.

**The Challenge**

A number of design variables (PMOS and NMOS transistor characteristics) can be acted upon to minimize the reaction time. Circuit performance, however, also depends on temperature, operating voltage, and process variables subject to uncertainty and statistical fluctuations related to the electronic component manufacturing process.

**The Solution**

Using machine learning techniques, a "yield simulator" has been defined, that is a software tool that allows to accurately predict the percentage of integrated circuits "in full working order". The simulator is then used to calculate the objective function to be minimized and, in order to take into account the variability of the parameters, the algorithm uses robust optimization techniques, which can take into account the variance of the process and manage the "worst case scenario".

**Results and Benefits**

Leveraging robust optimization techniques we identify the optimal configuration without the need to produce prototypes on which to perform tests. This significantly reduces design time and costs.

As part of robust optimization, discover how uncertainty quantification techniques enable optimization of variable parameters.

Robust Optimization - AMOLED Display Design

Robust optimization of a DC-DC converter for AMOLED displays

AMOLED Display Design

The goal is to optimize the design of an industrial three-phase induction motor (7.5 KW) used for industrial lifts and hoists.

**The Challenge**

The objective is to improve the performance of the motor by acting on a number of variables that define the shape and size of the rotor and stator and the windings in the stator slots, while simultaneously taking into account several factors: cost of construction, energy efficiency, power factor, and start-up current.

**The Solution**

Given a configuration, we calculate the performance of the motor using finite element simulation software. Through a purpose-developed interface, the simulator is embedded within a multi-objective optimization procedure. The optimizer takes into account all performances simultaneously arriving at a set of optimal solutions.

**Results and Benefits**

Multi-objective optimization allows to propose different engine designs that represent the best tradeoffs between constraints and performance.

Find out how to manage multiple factors in the design of an electric motor. Visit our page dedicated to multi-objective optimization!

Multi-Objective Optimization - Electric Motor Design

Multi-objective design of an electric motor

Electric Motor Design

Streamline processes and improve product design through single and multi-objective optimization algorithms

Process & Product Optimization

We apply our varied skills  to create original solutions to business problems. Our models translate into software and web applications for image and video analysis, for the solutions of forecasting problems and predictive maintenance and for simulation-based optimization.