Page - 372 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Wehave also developed a driver to import the datasets acquired bymeans of the UniMiBAAL suite [13]. UniMiBAAL includes twoAndorid apps that ease both the acquisition of signals from sensors in a controlled environment and the labeling tasks requiredwhenbuildingadataset. All thedrivershavebeenimplementedinthePythonlanguage,whilethewebservice has been developed relying on Laravel2 and exposing the services through RESTFul APIs. 4. DataManagement Once the signals have been standardized in terms of structure, they can be included in theUnifiedDataset tobedistributed.For this reason, theresponsibilityof theDataMan- agement component is twofold: it integrates the new labelled signals into theUnified Dataset (Data Storage) andmakes available sets of labelled signals to thosewho need them(DataAccess). Before being inserted into theUnifiedDataset, signals require another elaboration tomake themhomogeneousboth in termsof representationand label. Sensors record data at a given sampling frequency,with a given range of intensity values, and soon.Eachmanufacturer design their own sensorwith operating specifica- tions that may be different from the typical ones. For instance, wemay deal with ac- celerometer sensors thatworkat verydifferent sampling frequency ranging fromfew to hundredsofHertz.Machine learningmethods require inputdata in agiven format (e.g., number of samples per second and intensity range) that is consistent over time [4]. For this reason, raw data acquired by sensors need to be pre-processed before being pro- cessedbymachine learningmethods. Inevitably thisproduces anoverheadofdata tobe handled and stored. Storage spacemanagement is carried out using cloud storage tech- niques that reducespaceconsumptionbyusingcapacityoptimization,datadeduplication anddatacompression tools [14]. Signals fromdifferent datasetsmay have assigned different (but semantical equal) labels for the sameADL(e.g., ’walk’ vs ’walking’), same label for differentADL(e.g., ’sitting’mayrefer to thestateofbeingseated inachairor the transitionfromstanding to sitting), anddifferent labels for the sameactivity (e.g., ’running’vs ’jogging’). Thus, the aimof theDataManagement component is to harmonize signals and to make themavailable for exploitation.Theorganizationof the component is sketched in Figure3. The Data Aligner module is in charge of pre-processing the data from the Data Collection component in order tomake themusable by anymachine learningmethod. For example, an activity that is in charge of theDataAlignermodule is the conversion toa samemeasurementunit. TheLabelConsolidatormodule is in chargeofuniforming the labels of thedataset to include to a commonunified set. For example, if a dataset uses the label ’sitting’ to label signals related to the transition (fromstanding to sitting down) and in theUnified Dataset isused’sitdown’tolabelsignalsrelatedtothetransition(fromstandingtositting down), then the labelwill bechanged tobeconsistentwith theUnifiedDataset. Inview 2https://laravel.com/ A.Ferrari etal. /AFramework forLong-TermDataCollection372