Leen De Baets

Optimized statistical test for event detection in non-intrusive load monitoring

Event detection plays an important role in non-intrusive load monitoring to accurately detect when appliances are switched on or off in a residential environment. Besides being accurate, it is important that these methods are robust on real-life power traces. This paper shows that some state-of-the-art event detection methods may miss events when there is a substantial base load caused by active power consuming devices. In order to address this problem, this paper extends the existing chi-squared goodness-of-fit test with a a voting scheme. Furthermore, a work flow is proposed using surrogate-based optimisation for tuning the parameters of these tests in an efficient way. Results on the BLUED dataset indicate that the novel voting chi-squared GOF method outperforms the standard chi-squared GOF test when applied to traces with a higher base load.

Handling imbalance in an extended PLAID

The ability to classify appliances, given the current and voltage consumption of a household is useful for a variety of applications, including demand response verification, and eco-feedback technologies. To support research efforts in this problem domain, this paper presents an extended version of the Plug-Level Appliance Identification Dataset (PLAID), which is called PLAID 2 and contains 30 kHz voltage and current measurements of different residential appliances as they are switched on. As an extension to PLAID, this dataset adds appliance instances as well as measurements for multiple operating modes (e.g., low or high fan settings for air conditioners). As with other datasets in this problem domain, the appliance classes are not equally represented in PLAID 2. Different techniques for handling this imbalance and avoiding biasing the classifiers during training are investigated. The results indicate that performance improvement depends on the classifier type, when binary VI images are used as input.

Unsupervised Trajectory Inference Using Graph Mining

Cell differentiation is a complex dynamic process and although the main cellular states are well studied, the intermediate stages are often still unknown. Single cell data (such as obtained by flow cytometry) is typically analysed by clustering the cells into distinct cell types, which does not model these gradual changes. Alternative approaches that explicitly model such gradual changes using seriation methods seems promising, but are only able to model a single differentiation pathway. In this paper, we introduce a new, graph-based approach that is able to model multiple branching differentiation pathways as continuous trajectories. Results on synthetic and real data show that this is a promising approach which is moreover robust to parameter changes.