Predictive analytics on time series

In today’s digital society, significant quantities of data are being generated at an unprecedented rate, e.g., in health care, energy, or manufacturing. The omnipresence of sensors and connectivity generates significant growth in the complexity of data and offers the potential for a revolutionary transformation. So far, the analysis of data streams coming from sensors and mobile devices is still underexplored and offers a lot of exciting opportunities to enrich systems and make them smarter by running data-driven applications.

Predictive analytics is considered to be one of the key drivers of innovation in the working of such systems. It aims to identify and extract relevant features or patterns from time series data, with the aim of predicting certain outcomes of interest. It lies at the intersection of various interrelated disciplines of computer engineering, such as data modeling, statistics, machine learning, artificial intelligence, deep learning, and data mining. Many systems can greatly benefit from the insights and actionable knowledge that can be extracted from their data, such as, e.g., events, anomalies, and forecasts.

As machine learning models are being incorporated into a workflow, it is often desirable that the models can explain why a certain prediction was made. Therefore, the interpretability of such models is becoming an additional baseline requirement. Furthermore, it is important to have reliable estimates of the uncertainty of predictions, especially when dealing with unobserved variables, incorrect model assumptions, and noisy or imprecise data.

Application areas:

Health care                

Sensor readings provide essential information for a wide range of medical applications. The amount of data to be analyzed grows quickly and human-based analysis is a time-consuming task. To aid in this process, automated deep learning methods are being investigated to improve classification accuracy and interpretability.

 

Energy                       

The increase in computational power combined with advanced modeling and data analytics tools makes it possible to derive new technological solutions that can monitor residential energy consumption in households. Such solutions can provide valuable information to implement energy saving measures and to reduce the emission of greenhouse gases.

 

Manufacturing        

Many industrial processes rely on time-based and hands-on maintenance of equipment, despite recent advancements in predictive maintenance. By embedding sensors on different components, early signs of degradation that indicate the onset of breakdowns and failures may be detected. Predictive maintenance strategies can help to avoid downtime of equipment, the cost of unnecessary replacements, while improving process efficiency, safety and availability.

 

 

 

Selected use-cases:

Sleep apnea

Sleep apnea is one of the most common sleep disorders. It is characterized by pauses in the breathing, or shallow breathing during the sleep. A novel home monitoring solution was developed by combining new imec hardware innovations and state-of-the-art deep learning algorithms [1]-[2]. The imec device measures bio-impedance of the chest, which provides a more accurate estimation of both respiratory flow and effort, as compared to existing solutions that work with indirect measurements. This raw bio-impedance data is then filtered and analyzed using advanced deep-learning algorithms (such as LSTM’s), that are able to learn features and capture temporal dependencies in the data. This makes it possible to achieve a more reliable detection of sleep apnea events. The solutions that were previously developed are patented and valorized through the imec spin-off Onera Health.

 

Press release: Wearable Patch Uses Machine Learning to Detect Sleep Apnea, IEEE Spectrum, 2020

Patent: WO Patent, 2019/179836 (A1), PCT/EP2019/056222, 2018-03-19.

 

Non-intrusive load monitoring

Non-intrusive load monitoring methods aim to disaggregate the total power consumption of a household into individual appliances by analyzing changes in the voltage and current measured at the grid connection point of the household. The goal is to identify the active appliances, based on their unique fingerprint. By disaggregating the per-appliance energy consumption, adequate measures can be implemented to improve energy efficiency and to reduce the ecological footprint. Several new solutions were devised for robust detection of switching events, accurate appliance recognition, novel appliance identification and uncertainty quantification of the models. An extension of the PLAID dataset was released in collaboration with researchers from Carnegie Mellon University [3]-[8].

 

Anomaly detection

Anomalies are events that do not conform to the normal or expected behavior of a process. Their correct detection can be useful as it may uncover events of interest. Traditional algorithms use statistical approaches to find extreme values or correlation changes between features. Several robust, domain-agnostic, and computationally efficient approaches were developed for the detection of anomalies, and validated on manufacturing and signal integrity applications [9]-[11].

 

 

 

Related research projects:

Key publications: