Fault-Tolerance in Ambient Assisted Living

sensorsReviewA Systematic Survey on Sensor Failure Detection andFault-Tolerance in Ambient Assisted LivingNancy E. ElHady ID and Julien Provost * IDProfessorship of Safe Embedded Systems, Technical…

sensorsReviewA Systematic Survey on Sensor Failure Detection andFault-Tolerance in Ambient Assisted LivingNancy E. ElHady ID and Julien Provost * IDProfessorship of Safe Embedded Systems, Technical University of Munich, Boltzmannstraße 15,85748 Garching, Germany; [email protected]* Correspondence: [email protected]; Tel.: +49-(0)89-289-16424

Received: 13 April 2018; Accepted: 20 June 2018; Published: 21 June 2018

CookMyProjectAbstract: Ambient Assisted Living (AAL) systems aim to enable the elderly people to stay activeand live independently into older age by monitoring their behaviour, provide the needed assistanceand detect early signs of health status deterioration. Non-intrusive sensors are preferred by theelderly to be used for the monitoring purposes. However, false positive or negative triggers of thosesensors could lead to a misleading interpretation of the status of the elderlies. This paper presents asystematic literature review of the sensor failure detection and fault tolerance in AAL equipped withnon-intrusive, event-driven, binary sensors. The existing works are discussed, and the limitations andresearch gaps are highlighted.Keywords: ambient assisted living; sensor failure; fault detection; fault tolerance; smart home1. IntroductionAccording to the World Health Organization, the world’s population percentage of people agedover 60 is expected to double in the next decades to increase from 12% in 2015 to 22% in 2050.This phenomenon, known as Ageing Population, can be already witnessed in high-income countries.This demographic shift will induce new challenges to the countries, e.g., preparing the health care andsocial systems to deal with higher capacities [1]. Focusing on healthy ageing is an essential investmentfor facing that shift. Taking care of the elderlies would decrease the chance of further complications totheir health status. This can be achieved by providing care in nursing homes or hospitals. However, it iscostly and the costs increase greatly if the person needs specialized care due to immobilization orother health problems. A cost-effective alternative is using technology for independent living of theelderlies [2].Ambient assisted living (AAL) is defined as “the use of information and communicationtechnologies (ICT) in a person’s daily living and working environment to enable them to stay activelonger, remain socially connected and live independently into old age” [3]. AAL technologies canmonitor the behavior of elderly people at home and provide support whenever required, and hence,improve the quality of life [4]. This would cast some burden away from the family members of theelderlies, decrease the need for qualified caregivers and have a positive impact on the psychologicalstatus of the elderlies, as they would live independently at their homes longer and safer [5].Smart homes and ambient assisted living (AAL) terms were found to be interchangeably used inscientific articles, however, AAL is a special form of a smart home. AAL tools range between health andactivity monitoring tools, wandering prevention tools, and cognitive orthotics tools [6]. The technologyof those tools are based on ambient intelligence, a paradigm that integrates technology in people’senvironment to help them in their everyday lives by learning and adaptively responding to theirbehaviour [7]. Researchers are interested in investigating approaches to track the location and theactivities of the residents, prompting the residents, discovering the abnormal behavior, and predictingSensors 2018, 18, 1991; doi:10.3390/s18071991 www.mdpi.com/journal/sensorsSensors 2018, 18, 1991 2 of 19the future activities [8]. Integrating sensors in an unobtrusive intelligent way in the residents’ homes,allow monitoring their activities of daily living (ADL) to track their health status, and to detect earlysigns of diseases [9].The sensors used to monitor and locate the resident can be classified into intrusive sensors(e.g., camera, microphone) and non-intrusive sensors (e.g., motion detectors, pressure sensors).In practice, the sensors installed in the inhabitant’s place of residence may produce wrong output,e.g., false positives or negatives. A failure in one of the sensors of the AAL could lead to misleadingresult in activity recognition, or in location tracking. This can have dramatic consequences to the healthof the inhabitant [10].This survey paper aims to review the research work done in the sensor failure detection and faulttolerance in the presence of sensor failures in AAL systems equipped with non-intrusive binary sensors.The paper is organized as follows; Section 2 provides an overview of sensor failures, Section 3 presentsan overview of the typical publicly available datasets used in the reviewed works, Section 4 outlinesthe methodology used to conduct the literature survey, Section 5 presents the research work found inthe survey, Section 6 discusses the reviewed works and Section 7 discusses the status of research andhighlights the gaps.2. Background on Sensors Failures in Smart homes and AALA fault can be defined as an abnormal event that can cause an element or an item to fail, while afailure is the termination of the ability of an element to perform a function as required [11]. A fault mayor may not lead to failure.For sensor networks in general, two perspectives for fault type classification in sensor networkswas proposed by [12]:1. Data-centric viewpoint, which is based on the characteristics of sensor readings, e.g., stuck-atand spike.2. System-centric viewpoint, which describes faults causing the malfunction of sensor,e.g., low battery and calibration.The authors in [13] have presented another three perspectives for classification:1. Fault-tolerant distributed system viewpoint, that is based on the behaviour of the failed sensor,e.g., crash and omission.2. Duration viewpoint that classifies faults based on their duration e.g., permanent and intermittent.3. Components viewpoint, e.g., functional and informational faults.Several fault detection techniques have been developed for sensor networks. However,the techniques were mainly designed for time-driven, continuous-valued and homogeneous sensors,e.g., temperature sensors. Thus, those techniques are not suitable for the event-driven, binary andheterogeneous nature of sensors that are needed for the ambient assisted living, e.g., motion detectors,contact sensors, etc. [14].In an AAL system, a sensor failure is considered to be a fault from the perspective of the wholeAAL system. There are two main categories of sensor failures in the AAL terminology:• A fail-stop failure means that the sensor has stopped responding.• A non-fail-stop failure indicates that the sensor is still responding, however, the reported values areno longer representative of the measured variable, nor the occurring events in the surroundingenvironment that are intended to be detected.Sensor failures can also be classified as single-sensor failures and mutliple-sensor failures.In research works considering single-sensor failures, it is assumed that only one sensor can failat a time [14].In the field of AAL, Flöck has presented an overview of the binary sensors malfunctionsthat were observed during practical AAL implementation, e.g., faulty activation of motionSensors 2018, 18, 1991 3 of 19detectors by sunlight, bouncing of contact sensors, and switch-off delays of motion sensors [15].Also, Rahal et al. have reasoned the false information sent by binary sensors to be either due to anintrinsic error, e.g., the sensor’s error rate, or due to an external error, e.g., an air draft or a pet mayclose the door triggering false events [16]. Different types of non-fail-stop failures have been stated inthe research papers. Examples of the non-fail-stop failures are:• Moved-location failure, which occurs due to moving furniture that have sensors installed on it to adifferent area or re-mounting in the wrong location.• Obstructed-view failure that occurs due to covering the sensors or its dislodgement that may resultfrom regular use, cleaning, other non-residents, etc. [17,18].A set of guidelines and principles for the deployment of large-scale residential sensing systemswas proposed in [19], summarizing the experience gained from installing over 1200 sensors in over20 homes to monitor human activity. The main failure modes were examined to identify the longestacceptable time interval of inactivity for each sensor. For each periodic sensor, the interval is setto 5 times the sampling period, while for event-driven sensors, it is set to 36 h. The root cause offailure is identified based on the set of simultaneous sensor failures, where the considered causes offailure are wireless link loss, dead battery, disconnected plug, sensor sub-system down, internet-down,power outage, and gateway down. The described failure detection and classification approach wasapplied on four deployments for seven months. The analysis of the results showed that sensors are2.3 times more likely to fail due to being unplugged than to dead battery and that wireless link loss is aless cause of failure than the other sources of sensor down time. Failure of an entire sensor sub-systemappeared to be the most common cause of failure. This performed failure analysis enabled the authorsto present guidelines that could avoid some of the pitfalls and failures observed in the deployments.However, a fault detection and diagnosis system still needs to be implemented to deal efficiently withsensor failures.The following is the most common terminology found in the surveyed literature for the evaluationof various systems;• true positives (TP) are the data points reported as positive when they actually are positive• false positives (FP) are the data points reported as positive while they are actually negative• false negatives (FN) are the data points reported as negative while they are actually positive• true negatives (TN) are the data points are reported as negative while they are negative• precision measures the percentage of true positives from the total points reported as positive(TP/(TP + FP))• recall measures the percentage of true positives from the actual positive points (TP/(TP + FN))• accuracy measures the percentage of true positives and negatives from the data((TP + TN)/(TP + TN + FP + FN))• failure detection latency is the amount of time taken to detect a sensor failure after its occurrence.Figure 1 elaborates the terminology with respect to sensor failure detection systems, where theaccuracy, precision and recall values are 85%, 72% and 88%, respectively. The accuracy would still berelatively high if the system does not report as many sensor failures as before (lower TP and higher FN),however, the precision and recall would significantly drop. Thus, only using the accuracy for evaluatingthe system performance is insufficient. The precision indicates the ratio of the correctly reported sensorfailures to all the positively reported sensor failures, while the recall indicates the ratio of correctlyreported sensor failures to the positive sensor failures ground truth.Sensors 2018, 18, 1991 4 of 19Figure 1. Evaluation metrics terminology for sensor failure detection system.3. DatasetsThis section presents an overview of the publicly available datasets that were used in a numberof the reviewed research works. Other publicly available datasets exist for ambient assisted living,but they have not been used in research papers that focus on fault detection nor fault tolerance. It isworth noting that to the best of our knowledge, all the public datasets do not include any labels of thefaulty sensors data.3.1. Kasteren DatasetsTim van Kasteren has collected benchmark datasets (called house A, B and C) [20] from threesingle-resident apartments which were collected over 14, 23 and 19 days, respectively. Wireless sensorsthat gives binary output were installed; reed switches for the doors and cupboards, pressure mats forcouches and beds, mercury contacts for drawers, passive infrared (PIR) sensors to detect motionof resident in different areas of the apartments and float sensors for toilet flushing detection.The number of sensors installed in the three apartments (A, B and C) are 14, 23 and 21, respectively.During the collection of data, the resident performed his daily routine freely in an unscripted manner(i.e., the resident was not told what to do or which activity to perform). Annotation of the start and endof activities was performed by the resident using handwritten activity diary or a bluetooth headset [21].The following data is recorded in the dataset files; start and end date/time of sensor activation, sensorID, start and end date/time of activity and activity label.3.2. CASAS DatasetsThe CASAS research group in Washington State University (WSU) has made 64 datasets publiclyavailable [22]. The recorded datasets were either collected from the WSU smart apartment equippedwith around 90 sensors, residential apartments that has a number of sensors that ranges between 30 to50 sensors or SHib partner lab equipped with 25 sensors, for a duration ranging from hours to years,for single- or two-resident apartments. Some of the experiments were scripted, e.g., adlnormal data andadlinterweave data, and others were unscripted, e.g., aruba data and kyoto data. Examples of sensorsinstalled in the apartments are motion sensors, magnetic sensors, water flow sensors, item presencesensor, stove burner sensor and temperature sensors. The following data is recorded in the datasetsfiles; data/time, sensor ID, sensor value/status. Some of the datasets have labels for the start and endof the performed activities.3.3. Placelab DatasetsThree datasets (PLIA1, PLIA2 and PLCouple1) were collected from Placelab living lab [23] (notethat the Placelab dataset website has been down for months). The living lab is an apartment whereSensors 2018, 18, 1991 5 of 19volunteers live during the data collection process. Two datasets were collected from single residentsfor 4 h who were asked to perform a set of activities, and the third one was collected from a couplewho lived freely there performing their own daily routines for 10 weeks. The datasets were annotatedwith the performed activities using video recordings. The apartment is equipped with around 400sensors that range between reed switches, light sensors, motion detection sensors, water flow sensors,temperature sensors, humidity sensors, electrical current flow sensors, gas sensors, etc. [24].3.4. Tapia DatasetsEmmanuel Munguia Tapia has conducted experiments for two weeks in two single-residentapartments (subject 1 and subject 2) equipped with 77 and 84 sensors, respectively. The sensorsare reed switches attached to the everyday objects, e.g., drawers, doors, containers, refrigerator, etc.The residents carried out their daily activities without any scripts [25]. The following data is recordedin the datasets; activity label, start and end date/time of activity, sensor ID, start and end date/time ofsensor activation.4. Literature Survey MethodologyIn order to conduct the literature survey, the title, abstract and keywords fields were searched inScopus, IEEExplore, Web of knowledge and ACM databases for the following combination of terms;(”fault detection” OR ”sensor failure”) AND (”smart home” OR ”ambient assisted living”). Scopus andWeb of knowledge databases produced the largest number of relevant articles. The search was thenextended on Scopus and Web of knowledge to include more combinations of the keywords shown inTable 1, so that the combination is as follows; ((Group A AND Group C) OR Group D) AND Group B.Only the papers concerned with non-intrusive ambient binary sensors were included in the survey.The obtained articles were cross-referenced, and a total of 30 papers were selected for the review.It was observed that these 30 papers were all published between 2008 and 2017.Table 1. Search keywords.Group A 1 Group B Group C Group D”sensor*” ”smart home” ”fault detection” ”sensor* error””ambient assisted living” ”failure detection” ”sensor* failure*””AAL” ”fault toleran*” ”sensor* fault*””location tracking” ”fault identification” ”sensor reliab*””actvity recognition” ”failure identification” ”faulty sensor*””activity monitoring” ”fault diagnosis” ”*reliable sensor””activity detection” ”FDI” ”uncertain sensor””home* based care” ”fault isolation” ”sensor diagnos*””indoor localization” ”fault prevention” ”sensor node fail*””fault prediction” ”fail* sensor*””fault recover*” ”anomal*” AND ”binary sensor*””self-check*””self-heal*””dependable””failure management”1 * replaces any number of characters, i.e., sensor* will search for sensor, sensors, sensory, etc.The main focus of the research works can be mainly categorized as works concerned with:• sensor failure detection in AAL• fault-tolerant ADL recognition• fault-tolerant abnormal behavior detection• fault-tolerant indoor localization system/location tracking• maintenance scheduling/management• fault detection and diagnosis framework for AALSensors 2018, 18, 1991 6 of 19The reviewed papers classification is shown in Table 2. These papers are presented and analyzedin detail in the next section.Table 2. Main focus of the research works.Focus Research Work

Sensor failure detection Maintenance scheduling/management Fault-tolerant ADL recognition Fault-tolerant abnormal behavior detection Fault-tolerant indoor localization system/location tracking [10,14,17,18,26–36][14,27,30][14,27,30,37–45][37][16,46,47]

5. Literature Survey ResultsThis section provides a state-of-the-art review for the sensor failure detection systems and faulttolerance methods in the presence of sensor failures in AAL systems equipped with non-intrusive, binary,event-driven sensors. The research works are categorized according to the function of the proposedsystems as well as the approach that their methods are based on: correlation-based fault detection,model-based fault detection, fault-tolerant location tracking, fault-tolerant activity recognition or faultdetection and diagnosis framework for AAL, respectively. A glossary of the technical terms can befound at the end of this paper.5.1. Correlation-Based Fault DetectionThe following research papers proposed sensor failure detection systems based on eithersensor-appliance, sensor-sensor or sensor-activity correlations.FailureSense [17] was presented by Munir and Stankovic to detect fail-stop and non-fail stopmutliple-sensor failures. It is based on exploiting the correlation between the trigger of motionsensors and the activation/deactivation of electrical appliances. The correlation is represented by thesmallest interval of sensor firing after and before a turn on/off event within 5 min, denoted byIA and IB, respectively. The distribution of IA and IB is modelled by Gaussian mixture model(GMM), whose parameters are estimated from the training data using the expectation maximization(EM) algorithm. Online failure detection takes place by monitoring the sensor appliance behaviourrepresented by IA and IB. A failure is reported when a deviation occurs in the distribution beyondpredefined thresholds for each sensor-appliance pair. The thresholds are computed using the trainingdataset. Evaluation was performed on three real-home datasets with around two thirds of the datasetused for training and one third for testing. Fail-stop failure was simulated by removing all the readingsof a sensor after its randomly assumed day of failure. For the obstructed-view failure, simulation tookplace for two of the homes by randomly removing a 10-day period during which sensor view isconsidered to be obstructed, and for the third home, physical obstruction of the view of 5 motionsensors was done during the data collection phase. Simulation of the moved-location failure wasdone by replacing the readings of failed sensor with the readings recorded by the sensor at thenewly moved location. The evaluation metrics used are the precision and recall of failure detection,where they represent the percentage of the true failure alerts from the total observed failure alerts,and the percentage of the true failure alerts from the sensor failures, respectively. Experiments of thefail-stop, obstructed-view and moved-location failures produced approximately 82.8%, 90.5% and86.8% average precision, with an average recall of 92.86%, 84.4% and 89%, respectively. The effect ofincreasing the number of sensors that experience fail-stop failures on the percentage of failure detectionhas been also examined, showing an average of 86.6% sensor failure detection. On the other hand,a limitation of the proposed approach is that the average median failure detection latency is 22.08 h.Ye, Stevenson and Dobson presented a technique to detect missing data in event-driven sensorsbased on temporal correlation and time-series analysis [26]. Temporal correlation relationship isdefined to indicate if two sensors fire within a preset time interval. A missing data is reported whenSensors 2018, 18, 1991 7 of 19one of two highly correlated sensors fires without the other. For each sensor, the next firing time ispredicted using non-linear time analysis technique, and if it does not fire at the predicted time, then it isconsidered as missing data. Evaluation is carried on Kasteren dataset [20] (house A), in which randomlychosen sensors events were removed from the testing data, using precision and recall metrics for eachof the temporal correlation and time series approaches independently, then combined. The effect ofchanging predefined parameters of the algorithm on the performance was also examined. Moreover,the relation of increasing the error rate percentage (percentage of data removed) in the testing set onprecision and recall was plotted along with increasing the percentage of training set. The results on theexamined dataset have shown that the performance of using the temporal correlations for detectingmissing events is better than using the time-series analysis. Also, it was observed from the results thatusing both temporal correlation and time-series analysis simultaneously for failure detection had avery low impact on the performance improvement. Using temporal correlation with data split by halffor training and testing sets, the precision was nearly 70% and the recall decreased from around 80%to 40% with increasing the error rate from 10% to 90%. Increasing the training data to 90%, has madethe precision to be around 78% and the recall to decrease from 85% to 75%. The authors stated that theproposed approaches could not be sufficiently evaluated on the chosen dataset, as it has few sensorsand is collected over a short duration.Kodeswaran et al. aimed to propose a system called Idea, for monitoring the activities of dailyliving while preserving a reduced maintenance overhead [27]. It is based on the assumption that thereare redundant heterogeneous sensors installed for detecting each activity. Maintenance is scheduledaccording to the impact of a sensor failure on the performance of the system to detect ADL. The maincomponents of Idea are; ADL signature Extraction, ADL detection, Impact estimation, Sensor Failuredetection and Maintenance scheduling. Frequent itemset mining algorithm is used to form a rule-basecontaining the frequently occurring subsets of sensors for each ADL, and then the most probable timeof day of occurrence and duration of activity are calculated from the training dataset. The criticalsensors are identified based on their impact on detecting the ADL, which depends on the redundancylevel per ADL using the training dataset. For critical event-driven sensors, a failure alert is flaggedif the time elapsed since the last detection of ADL exceeded a threshold. For non-critical sensors, ararity score is computed as the probability that a sensor has not been triggered while certain ADL, thatshould involve this sensor, has occurred. Experiments were conducted on Kasteren [20] (house A, Band C) and CASAS [22] datasets (aruba, twor9-10, twor2009, tworsmr and adlnormal) using 80% ofthe dataset for training. The accuracy of ADL detection was investigated in the presence of fail-stopsensor failures, emulated by discarding all the events of the failed sensor, and compared to NaiveBayes classifier (NB) and Hidden Markov model (HMM) algorithms. The maintenance efficiency wasalso evaluated in terms of the number of maintenance visits and per-home maintenance inter-arrivaltimes. Across all the datasets, the ADL detection accuracy is reduced in average by approximately0.5%, 1% and 3% in the presence of 1, 3 and 7 failed sensors.Dealing with sensor faults in smart homes using data-driven approach was proposed byMonekosso and Remagnnino [28]. The proposed method aimed to detect sensor faults, mask it,and differentiate between anomalous activities and sensor deviation by combining reconciliationwith failure detection techniques. The approach has two components; one component deals withrandom measurement fluctuations using data reconciliation, while the other component deals withsystematic deviations due to sensor failures or anomalous activities. Models of sensors correlations arebuilt using historical data via principle component analysis (PCA) and canonical correlation analysis(CCA). The models are refined continuously and can deal with heterogeneous sensors types to beused for detecting sensor faults. Experiments were carried out using Kasteren dataset [20] (house A).Two case studies were implemented by injecting intermittent and permanent faults into the dataset.A permanent fault was simulated on a sensor by removing its readings from the testing dataset afterthe assumed failure point of time. A transient sensor fault was injected by corrupting random instancesof sensor readings with wrong values.Sensors 2018, 18, 1991 8 of 19An approach for data-driven failure detection based on clustering was proposed by Ye, Stevensonand Dobson. They address non-fail-stop sensor failures as a clustering-based outlier detectionproblem [18,29]. DBSCAN clustering based outlier detection algorithm is used. The similarity betweenbinary sensor events is calculated using least common subsumer (LCS) based on their semantic features;time stamp, the object to which a sensor is attached, location and user. Data points are clusteredinto groups and then the groups are sorted by their size in descending order. Shoulder-locationmethod is used to select the threshold below which a cluster is considered small. To each data point,a cluster-based local outlier factor (CBLOF) is assigned which is a function in the size of the clusterto which this point belongs, the similarity between the point and the closest large cluster, and thehistoric faulty sensor behaviour. A point is considered as an outlier if its CBLOF is below a thresholddefined by the shoulder location method. The technique was evaluated on Placelab [23] (PLCouple1),Kasteren [20] (house A and B) and CASAS [22] (adlinterweave) datasets with injecting random andsystematic anomalies. Random abnormal events were injected into the datasets by randomly creatingnew sensors events within randomly selected time slots. While systematic abnormal events are injectedby selecting random sensors and creating an event for each of the selected sensors within each timeslot of the testing data. Plots of the precision and recall against the injection rate of abnormal eventswere presented.In another attempt, detection of sensor failures was tackled using classification. Kapitanova etal. proposed simultaneous multi-classifier activity recognition technique (SMART) [14,30], whichuses top-down application level semantics to detect non-fail-stop single-sensor failures. Furthermore,the research work addresses schedule maintenance according to failure severity and improvementof activity recognition accuracy in the presence of failures. Multiple classifier instances are trainedoffline by excluding each time a sensor out of the training set resembling a sensor failure, and onetime with all sensors present in the set. Online detection of a fault is achieved by assessing the relativeperformance of the classifiers that has a missing sensor versus the one trained with all sensors, thus afault is detected and identified. Severity analysis is performed to evaluate the impact of sensor failureon the accuracy of activity detection. As the level of sensor redundancy increases per activity, theurgency of repairing a faulty sensor decreases. Fault-tolerance of the activity recognition is achievedby updating the classifier ensemble with the classifiers that were previously trained to deal with aparticular sensor failure. The system was evaluated using CASAS [22] and Kasteren [20] (house A andB) datasets considering only prepare breakfast, lunch and dinner activities. NB and HMM classifierswere used. Stuck-at failures and misplacement failures were introduced manually to the datasets.To simulate stuck-at failure, the value of the failed sensor is set to 1. For simulating misplacementfailure, the data of failed sensor is replaced with the sensor in its new location. The results showedthat this approach could decrease the number of maintenance dispatches by 55%, identifies non-failstop failures by 85% accuracy, and improve activity recognition accuracy in presence of sensor failuresby 15%.5.2. Model-Based Fault DetectionThe following researchers have used model-based fault detection based on localization systems.An indoor human localization (IHL) system with fault detection focusing on hardware as well ashuman-made single faults was presented by Veronese et al. [31]. The IHL system consists of three maincomponents; an RF-based localization subsystem, an off-the-shelf modular wireless home automationsubsystem and a fault detection subsystem. The types of sensors chosen for home automation werecontact sensors and passive infrared (PIR) sensors. A model-based fault detection approach wasapplied based on the concept proposed by Isermann [48] which states that a fault can be detectedusing the dependencies between different measured signals. The activation of the home automationsensors and its features were used to estimate the resident’s location. Also, the position of the residentis estimated independently with the localization subsystem. The fault detection subsystem comparesthe two estimated location areas, and flag a fault whenever there is no intersection between the twoSensors 2018, 18, 1991 9 of 19areas. Experimental work was done, where 19 fixed LAURA anchors and 7 Z-wave devices were fixedacross the rooms of the university building. Two fault scenarios were considered; forgotten worndevice and blinded PIR motion detector. The results showed that the faults could be detected using theproposed approach. As a continuation of the work, multi-user simulation was conducted using threevirtual users trajectories, the faults could be detected in the presence of multiple users with specificityand sensitivity above 90% [32].Danancher proposed model-based location tracking of single as well as multiple inhabitantsin smart homes [10]. He treated the location tracking of inhabitants as a problem of discrete eventsystem modeling. Finite automata was used to model the observable motion of inhabitant, where eachstate represents a zone in the apartment, each event represents the rising or falling edge of binarysensor, and each transition is the observable location change. A case study was presented for anapartment equipped with motion detectors and door barrier sensors. The impact of sensor faults on theperformance of location tracking was discussed. The applicability of three model-based fault detectionand isolation (FDI) approaches; diagnoser, template and residual approaches, were investigated forfault-tolerant location tracking. An adaptation to the residual-based approach was applied to a casestudy of tracking a single inhabitant. Three fault scenarios were considered; spurious activation of amotion sensor, failure of power supply of door barrier sensor and a failure of motion detector sensor.The approach could not detect nor isolate faulty sensors in the proposed faulty scenarios. The authorconcluded that the industrial FDI approaches are not suitable for sensor faults in smart home and thata new FDI approach designed specifically for smart homes should be developed.Another discrete event system approach for location tracking was proposed by Wu et al. [49].The motion of the resident is modeled using an automaton model and the observations of motionevents from sensor signals are described using the state tree of Graph theory. An Observer is thenused to estimate the location of the inhabitant. Dealing with transient sensor faults is performedby adding a reset procedure to the state tree and the observer so that they return to the initial statewhenever blocking occurs due to missing or disordering of a sensor event. This ensures that thelocation tracking returns to output correct estimation results after deviating due to the transient sensorfault. However, false location estimation still occurs. A scenario of the motion of inhabitant in thepresence of a missing sensor event was presented.Amri et al. have proposed fault detection approach for indoor localization based onset-membership fault detection using the q-relaxed intersection method [36]. The random walkmodel is used as the mobility model of the resident. The PIR sensor activation leads to the activationof a box representing its coverage area. At one second time step, the measurement boxes are observedand the predicted boxes are deduced using the mobility model. The q-intersection method deducesthe location zone of the resident using these boxes. Outlier detection takes place by comparing thesolution set obtained and the measurements. Experiments were conducted in a living lab equippedwith PIR sensors.5.3. Fault-Tolerant Location TrackingA fault-tolerant location tracking system was presented by Rahal, Pigot and Mabilleau, which aimsto localize single inhabitant using the already installed sensors in smart home [16]. The authors aimedto provide a reliable location tracking system that can estimate the location of inhabitant accuratelydespite the false trigger of sensors that may occur due to various factors. The adopted approach isbased on sensor fusion, in which particle filters approach is used to estimate the new inhabitant’slocation using the last known position and the last sensor event. To evaluate the system, experimentswere conducted in the DOMUS apartment, where non-intrusive unobtrusive sensors (infrared (IR)presence sensors, tactile carpets, smart light switches, contact sensors and pressure detectors) areinstalled. A daily routine scenario was performed by 14 subjects, one subject each time, and the resultsshowed an accuracy in location tracking above 85%. The system performance was also investigatedwith respect to the inhabitant’s profile, sensor configuration, inhabitant’s dynamics and in the presenceSensors 2018, 18, 1991 10 of 19of noise. The results showed that the accuracy of the system is profile-independent. The accuracyof localization when using only infrared sensors is similar to using all the sensors. However, the IRsensors are more prone to false triggers, thus, the authors recommended the usage of at least one othertype with IR sensors. The system accuracy remained at 84% when 2.5% and 5% noise were applied tothe collected data.A similar system was proposed by Ballardini et al. that is based on estimating the resident locationin the presence of false positive or false negative sensor readings via Bayes filtering [46]. The systemuses a probabilistic model of the sensors and a motion model of the inhabitant. The proposed approachwas tested on two noisy datasets that use PIR sensors (observed frequent false triggering of a motionsensor when no person is moving, and trigger of atrium’s motion sensor when motion occurs in thedining room), producing 5% and 9% error rates in localization.A fuzzy set-based approach for localization tolerating sensor failures was proposed byAhvar et al. [47]. The approach relies on using several functionally redundant sensors at specificnodes. The system is composed of sensor nodes and context broker based on the fuzzy set theory.The apartment is divided into zones and equipped with various types of ambient sensors. The sensorssend context information, then the membership values for each zone is computed. The highest valueindicates the user location. A case study was presented and simulated using the DPWsim simulatorwith different sensor error rates. However, the system was not verified using a real dataset.5.4. Fault-Tolerant Activity RecognitionIn addition to the fault-tolerant activity recognition implemented by SMART system [14,30]and Idea system [27], a framework of fault-tolerant activity recognition was addressed byHong et al. [38–40]. First, the effect of sensor failures on the accuracy of activity recognition wasinvestigated. Only binary sensors were considered for monitoring the ADL in smart homes. Sensorevidence reasoning network was designed based on activity hierarchy of ontology for activityrecognition while tolerating uncertainty in the sensors’ measurements. The discounting values dependon the manufacturer statics on the sensors. To validate the proposed approach, a case scenario waspresented. In addition, sensors data recordings were collected from smart laboratory environmentof a kitchen area for four weeks, and then, offline analysis was performed to verify the sensor datawith video recordings. The sensor data was fed to the evidential reasoning network that is based onthe Dempster-Shafer theory. The performance of activity recognition was assessed with respect tothe number and combinations of sensor failures. Mckeever et al. [41] have extended the evidenceof theory to incorporate temporal features and evaluated their proposed framework on Kasterendataset [20] (house A). A limitation of the approach is that expert knowledge is needed for the sensormass functions and sensor quality. Also, knowledge from users is used to get information about thetemporal features of activities.Liao et al. [42–44] have proposed an activity recognition framework that deals with uncertaintyin sensor measurements based on Dempster-Shafer theory of evidence while considering the effect ofhistorical information and activity patterns. This is implemented through a framework with a latticestructure, which has a context layer that includes combinations of sensors derived from the historicaldata of inhabitant. Two types of uncertainty sources were considered; sensor hardware and contextuncertainty due to the variability in human activities. A case study was presented in addition toapplying the proposed approach to a publicly available dataset (Tapia dataset, subject 1) [25] collectedfrom an apartment equipped with binary sensors. The performance was evaluated using precision,recall and F-measure metrics for activity recognition.A Weighted Dempster-Shafer theory was presented by Javadi, Moshiri and Yazdi [45], where aweight for each sensor is assigned based on the historical data and activity patterns of the resident.In the training phase, 10 models are built for each sensor, and then in the testing phase, a weight foreach sensor is calculated based on the membership degree of each sensor signal to the sensor’s models.The proposed method is applied to a dataset (Tapia dataset, subject 1) [25] and evaluated through theSensors 2018, 18, 1991 11 of 19accuracy detection rate of toileting activity. A drawback in the experiments is that, sensor faults werenot injected to the dataset.Abnormal behaviour recognition in the presence of sensor failures/uncertainties was addressedby Marhic et al., it is based on the evidential approach using transferable belief model (TBM) [37].It is assumed that there are three or more heterogeneous redundant sensors per each monitored activity.The system consists of Sensor FDI and the Abnormal behaviour detection modules. The Sensor FDIanalyses the conflict between the heterogeneous redundant sensors using sensor fusion calculated bythe Smet’s operator and two experts. Abnormal behaviour is then detected by comparing the normalbehaviour of inhabitant represented by the Markov chain model (MCM) and the detected/predictedbehaviour within the TBM framework. Experiments were conducted on datasets collected fromperforming sitting, lying and standing activities with various single sensor failures, during whichpressure sensor, omni-directional vision sensor and an accelerometer were used. The authors showedthe ability of the system to detect abnormal behaviour in presence of sensor failures (unpluggingsensor for a period of time) and highlighted some limitations that could be addressed in the future.Methods for fault tolerance in Ambient Assisted Living were suggested by Ahvar et al. [50].Data from binary sensors, e.g., movement sensors, may be corrected using a model of the inhabitantbehaviour. While fault tolerance for analog data from sensors, e.g., temperature sensors, may beimplemented using sensor fusion. However, the system was not verified against faults in a case study.5.5. Fault Detection and Diagnosis Framework for AALA fault detection and diagnosis framework for Ambient Intelligent systems was presented byMohamed, Jacquet and Bellik [33,34], however, it is concerned only with the tasks performed by thesystems through the actuators. The approach is based on modeling the physical phenomena that aresupposed to occur in the environment due to the activation of a particular actuator. The system thenautomatically creates links between actuators and sensors at run-time using the models. It predicts theexpected sensor reading due to the activation of an actuator and compares it with the actual sensorreading to detect if a fault has occurred. Simulations were performed to illustrate the operation of thesystem and show the ability of the system to discover new components at run-time. The basic idea ofthe diagnoser model was presented without details.A self-diagnosis framework was proposed by Oliveira et al. [35], where a Bayesian networkconstruction algorithm is used to create a Bayesian network for each scenario that is supposed tobe fulfilled by the AAL system to assist the user. The algorithm takes as inputs the rules file thatspecifies the causal relations between variables, and the scenario description file that specifies therequired assistance and the home description. Conditional probability distribution is calculated foreach child node. The real values are then compared with the predicted ones and a fault is flagged ifthe readings are different. Using the causality relations and conditional rules, a diagnostic is reached.A case study was investigated to show the ability of the proposed framework to detect and diagnosefaults. However, like the previous system [34], the framework would only work fine for the tasks thatinvolve a sensor-actuator feedback.6. Discussion6.1. Correlation-Based Fault Detection SystemsNext, we discuss the pros and cons of the correlation-based fault detection approaches.FailureSense [17] has good average precision and recall for the examined fail-stop, obstructed-viewand moved-location failures. Also, the experiments show consistent performance for failure detectionwith increasing the number of sensors that had fail-stop failures. However, the method does not workwell if the failed sensor is not associated with any electrical appliance. In addition, its average failuredetection latency is not suitable for emergency situations. Another limitation of the system is that, itis based on the assumption that the resident has to be physically beside the appliance to turn it on.Sensors 2018, 18, 1991 12 of 19In addition, the system performance depends on the behaviour of residents (i.e., the residents turnon/off electrical appliances remotely or their behaviour pattern in using electrical appliances).Using temporal correlations and/or time-series analysis in [26] only relies on sensors firing todetect missing sensor events. The temporal correlation method achieves better results than usingtime-series analysis. However, the average precision and recall on the examined dataset with randomnon-fail-stop failures are not as good when increasing the error rate percentage, except when thetraining data percentage was increased to 90% . This makes the performance of the proposed methodstill questionable and needs to be evaluated on other larger datasets.The approach of the Idea system [27] is designed to suit homes equipped with functionallyredundant sensors per activity of daily living. Otherwise, it will not work as expected. In this work,only fail-stop failures were considered. The reduction in the ADL detection accuracy in the presenceof sensor failures is relatively low. Thus, an efficient fault tolerant activity recognition seems to bepromising using the proposed approach. However, the effect of monitoring multiple ADLs to detectsensor failures on the failure latency detection, and the effect of rarity threshold on the false positivealerts were the only assessments used for the sensor failure detection subsystem. Those assessmentsare not enough to be able to see the efficiency of the sensor failure detection. Also, non-fail-stop failuresneed to be considered in the experiments. In our opinion, detecting failures using time elapsed is notan efficient solution and using the rarity score assumes that the system has not misclassified the activity.Similarly, the detection of sensor failures using the proposed approach in [28] was not thoroughlyevaluated. The experiments were only concerned with the ability of the system to detect and isolate afaulty sensor, without any quantitative evaluation of the performance. Another drawback is that theinjected faults in the experiments were applied on only a single sensor.The advantage of using clustering approach as in [18,29], is that no training phase is required.However, the proposed method aims to detect false sensor triggers, but it can not detect missingsensor data. Another limitation is that the failure detection takes place in a post-processing stepon the collected data. Also, the false positive trigger is less likely to be detected if it is associatedwith a sensor that has similar features to other correctly working sensors. Using multiple classifierinstances [14,30] produced promising results for sensor failure detection and fault-tolerant activityrecognition. However, the disadvantage of this approach is that the training effort is large and itincreases proportionally with the number of installed sensors, thus the system is non-scalable.6.2. Model-Based Fault Detection SystemsThe reviewed model-based fault detection systems do not seem to provide better results thanthe correlation-based fault detection systems. The approaches mainly rely on checking if the sensortrigger is consistent with the predicted location of the resident. The method proposed in [31] thatuses RF-based localization system in addition to the environmental binary sensors installed at home,can not identify if the fault source is the localization system or the installed sensors. In another researchwork [10], applying residual-based fault detection to the location tracking finite automata model of aninhabitant could not detect nor isolate the faulty sensors. Only preventing the transient sensor faultsfrom blocking the discrete event location tracking model was proposed in [49]; however, it was noteven capable of tolerating those faults. In [36], comparing the motion sensors triggers with the randomwalk mobility model is not reliable, since this mobility model can produce unrealistic patterns as itdoes not keep track of the past locations and speed.6.3. Fault-Tolerant Location Tracking SystemsThe fault-tolerant location tracking systems reviewed are based on attempting to estimate thelocation of the resident under uncertainty of sensors whether through sensor fusion [16,46] or fuzzytheory [47]. The results seem promising, however, the systems need to be investigated more thoroughlyin real-time experiments.Sensors 2018, 18, 1991 13 of 196.4. Fault-Tolerant Activity Recognition SystemsIn addition to the SMART [14,30] and Idea [27] systems discussed before for proposing sensorfailure detection and fault-tolerant activity recognition, fault-tolerant activity recognition based onrecognizing activities under sensors uncertainty were reviewed. The works that used the evidencetheory [37–45] have the disadvantage of requiring lots of expert knowledge.6.5. Fault Detection and Diagnosis FrameworkThe reviewed fault detection and diagnosis frameworks [33–35] were designed to only suit AALsystems involved with sensor-actuator feedback.7. ConclusionsIn the last 10 years, an increasing interest in tackling sensor failures/faults in AAL has beenobserved. However, there is still much to be done in this area to offer a dependable system for the users.Tables 3 and 4 summarize the work reviewed in Section 5. For each research work; the contributedsystem, its method, algorithm(s), experiments conducted and performance metrics used are listed inthis table.The overall general limitations of the existing works can be categorized as follows:Limitations regarding the approaches:• Most of the existing works have developed their approaches considering only single failures.However, it may happen that more than one sensor fail simultaneously.• The majority of the developed algorithms use parameters or thresholds that need to be chosen byan expert rather than being deduced automatically.• Differentiating between failed sensors and anomalies in human behaviour is still a challenge thatneeds to be addressed.Limitations regarding the datasets:• The public datasets used for the training and testing phases are limited to short duration,low sensor node redundancy and single resident apartments.• Also, the data in the publicly available datasets was originally collected for activity detectionwith labelled activities, thus, failures or anomalies were not labelled. Instead, sensor failureswere manually injected and simulated by the researchers, which may not be representative ofreal-home sensor failures rate and percentage.Limitations regarding the experimental methodology:• It is difficult to compare between the efficiency of the presented approaches because not allthe authors use the same evaluation criteria and same testing data. Thus, there is a need forstandardized evaluation criteria.• Beside the accuracy, precision and recall, the sensor failure detection latency is an importantcriterion to be considered.• Real-time online evaluation of the algorithms was not carried out, instead the data collected fromprevious experiments or datasets were fed to the algorithms.• The proposed approaches should additionally be evaluated on data collected from elderlies withphysical and/or cognitive deficiencies.Sensors 2018, 18, 1991 14 of 19Table 3. Summary of the reviewed work in Sections 5.1 and 5.2.Source Contribution Method AlgorithmExperimentsPerformance MetricsData Failure Type[17] sensor fault detection sensor-appliancecorrelations GMM & EM custom datasetsinjecting fail-stop and non-fail-stop(obstructed-view and moved-location)failuresprecision, recall &failure detectionlatency[26] sensor fault detection sensors correlations mutual information and non-lineartime series analysis techniquespublicly available dataset(Kasteren, house A)injecting non-fail-stop failures(removing random sensors events) precision & recall[27]sensor fault detection,fault-tolerant activityrecognition &maintenance schedulingsensor-activitycorrelationsfrequent itemset mining algorithm& rarity score calculationpublicly available datasets(Kasteren; house A, B & C, andCASAS; aruba, twor9-10,twor2009, tworsmr &adlnormal)injecting fail-stop failuressensor failure falsealert rate, failurelatency detection& reduction in ADLdetection accuracy inpresence of failures[28] sensor fault detectionand masking sensors correlations PCA & CCA (Kasteren, house A) publicly available dataset injecting permanent and intermittent faults (i.e., fail-stop and non-fail-stop) ability to detect faults[18,29] sensor fault detection clustering-based outlierdetection DBSCAN clustering algorithmpublicly available datasets(Placelab, PLCouple1, andKasteren; house A and B, andCASAS, adlinterweave)injecting random and systematic falsepositive sensor triggers (non-fail-stop) precision & recall[14,30]sensor fault detection,fault-tolerant activityrecognition &maintenance schedulingsimultaneous use ofmultiple classifiersNB, HMM, hidden semi-Markovmodel (HSMM) & decision treespublicly available datasets(Kasteren, house A and B, andCASAS not specified))injecting non-fail-stop failures(stuck-at and moved-location)failure detectionaccuracy & failurelatency detection[31,32]indoor localizationsystem with faultdetectionmodel-based faultdetection using RF-basedlocalization & homeautomation subsystemsestimating the location using theactivation of home automationsensors and the RF-basedlocalization subsystemcustom dataset collected with blinded PIR sensor andforgotten worn devicesensitivity &specificity[10] location tracking withsensor fault detectionmodel-based faultdetection using a modelof the observed motionof the inhabitantfinite automata & residualcalculation scenario of motion of inhabitant in the presence of fail-stop and non-fail-stop failures ability to detect faults[49] location tracking dealingwith transient faultsstate estimation withreset procedureautomaton model & state tree ofgraph theory scenario of motion of inhabitant scenario of the presence of missing sensor event (non-fail-stop)location estimation inpresence of transientsensor faults(non-fail-stop)[36] localization system withsensor fault detectionmodel-based faultdetection using therandom walk model ofinhabitantset-membership fault detectionusing the q-relaxed intersectionmethodcustom data collected fromLiving lab not specified (outliers) ability to detect faultsSensors 2018, 18, 1991 15 of 19Table 4. Summary of the reviewed work in Sections 5.3–5.5.Source Contribution Method AlgorithmExperimentsPerformance MetricsData Failure Type[16] fault-tolerantlocalization systemstate estimation based onsensor fusion particle filters approach custom data collected (non-fail-stop) injecting random sensor noise localization accuracy & mean belief[46] fault-tolerantlocalization system state estimation bayes filtering custom dataset data collected in presence of noise localization error rate[47] fault-tolerantlocalization systemfuzzy-based approach usingvarious types of ambientbinary sensorsfuzzy-set theoryscenario and simulation ofmotion of inhabitant onDPWsim simulatorin the presence of sensor nodefailure fail-stop and non-fail-stop localization accuracy[38–40]fault-tolerant activityrecognitionframeworkevidential approach forreasoning under uncertaintysensor evidence reasoningnetwork & dempster-shafertheoryscenario and custom datacollectedinjecting different combinations ofsensor failures belief in activity inference[41]fault-tolerant activityrecognitionframeworkevidential approach forreasoning under uncertaintytemporal evidence theory& dempster-shafer theorypublicly available dataset(Kasteren, house A) no faults injected precision, recall & F-measure activity recognition[42–44]fault-tolerant activityrecognitionframeworkevidential approach forreasoning under uncertaintyevidential lattice structureconsidering historicalinformation and activitypatterns & dempster-shafertheoryscenario and publiclyavailable dataset (Tapia,subject 1)no faults injectedactivity recognitionprecision, recall andF-measure of activityrecognition[45]fault-tolerant activityrecognitionframeworkevidential approach forreasoning under uncertaintyweighted dempster-shafertheory & fast fouriertransformpublicly available dataset(Tapia, subject 1) no faults injected activity recognition accuracy[37]fault-tolerantabnormal behaviourdetectionevidential approach forreasoning under uncertainty inthe presence of heterogeneousredundancy per activitysensor fusion based onSmet’s operator, experts,TBM & MCMcustom data collected with inducingnon-fail-stop sensor failureability to detect abnormalbehaviour and/or failedsensor[33,34]fault detection anddiagnosis frameworkfor AALmodeling the physicalphenomena that are supposedto be detected by sensor due tothe activation of an actuatornot applicable simulating a scenario inpresence of sensor failure not specified ability to detect system fault[35] self-diagnosisframework for AALBayesian network for eachscenario that is supposed to befulfilled by the AAL system toassist the userbayesian networkconstruction algorithmscenario of inhabitant in thepresence of sensor failure fail-stop ability to detect system faultSensors 2018, 18, 1991 16 of 19As illustrated by the number and importance of the limitations of the existing works,fault-tolerance in AAL is still in its early phase. Thus, intensive research work is still neededto tackle them. The research topics to be addressed can be grouped in the three followingresearch questions:• Can novel machine learning techniques tackle the problem of sensor failure detection in AALwithout the need for expert knowledge?• Should the research priority be directed towards enhancing the accuracy of binary sensors orinstead towards dealing with the faulty sensors data through fault-tolerant systems?• Would differentiating between behaviour anomalies of residents and sensor anomaliesbe possible?As a conclusion, as highlighted by this systematic literature review, methods for fault-tolerantAmbient Assisted Living are still in their infancy stage. Also, intensive research works would beneeded to ensure the development and implementation of a robust sensor fault detection and diagnosissystem for Ambient Assisted Living in a near future.Conflicts of Interest: The authors declare no conflict of interest.AbbreviationsThe following abbreviations are used in this manuscript:

AAL Ambient assisted livingICT Information and communication technologiesADL Activities of daily livingGMM Gaussian mixture modelEM Expectation maximizationNB Naive BayesHMM Hidden Markov modelPCA Principle component analysisCCA DBSCAN LCS CBLOF SMART IHL Canonical correlation analysisDensity-based spatial clustering of applications with noiseLeast common subsumerCluster-based local outlier factorSimultaneous multi-classifier activity recognition techniqueIndoor human localizationPIR Passive infraredFDI Fault detection and isolationIR InfraredTBM Transferable belief modelMCM HSMM Markov chain modelHidden semi-Markov model

References1. World Health Organization. Global Strategy and Action Plan on Ageing and Health; World Health Organization:Geneva, Switzerland, 2016.2. Al-Shaqi, R.; Mourshed, M.; Rezgui, Y. Progress in ambient assisted systems for independent living bythe elderly. SpringerPlus 2016, 5, 624. [CrossRef] [PubMed]3. Monekosso, D.; Florez-Revuelta, F. ; Remagnino, P. Ambient Assisted Living [Guest editors’ introduction]IEEE Intell. Syst. 2015, 30, 2–6. [CrossRef]4. Memon, M.; Wagner, S.R.; Pedersen, C.F.; Beevi, F.H.A.; Hansen, F.O. Ambient assisted living healthcareframeworks, platforms, standards, and quality attributes. Sensors 2014, 14, 4312–4341. [CrossRef] [PubMed]5. Steinke, F. Influence of Trust in Ambient Assisted Living Technologies. Ph.D. Thesis, LebenswissenschaftlicheFakultät, Berlin, Germany, 2015.Sensors 2018, 18, 1991 17 of 196. Rashidi, P.; Mihailidis, A. A survey on ambient-assisted living tools for older adults. IEEE J. Biomed.Health Inf. 2013, 17, 579–590. [CrossRef]7. Debes, C.; Merentitis, A.; Sukhanov, S.; Niessen, M.; Frangiadakis, N.; Bauer, A. Monitoring activities ofdaily living in smart homes: Understanding human behavior. IEEE Signal Process. Mag. 2016, 33, 81–94.[CrossRef]8. Nazerfard, E.; Cook, D.J. CRAFFT: An activity prediction model based on Bayesian networks. J. AmbientIntell. Humaniz. Comput. 2015, 6, 193–205. [CrossRef] [PubMed]9. Viard, K.; Fanti, M.P.; Faraut, G.; Lesage, J.J. An Event-Based Approach for Discovering Activities of DailyLiving by Hidden Markov Models. In Proceedings of the International Conference on Ubiquitous Computingand Communications and 2016 International Symposium on Cyberspace and Security (IUCC-CSS),Granada, Spain, 14–16 December 2016; pp. 85–92.10. Danancher, M. A Discrete Event Approach for Model-Based Location Tracking of Inhabitants in SmartHomes. Ph.D. Thesis, École normale supérieure de Cachan-ENS Cachan, Cachan, France, 2013.11. International Organization for Standardization (ISO). 26262–1: 2011 Road Vehicles Functional Safety;International Organization for Standardization: Geneva, Switzerland, 2011.12. Ni, K.; Ramanathan, N.; Chehade, M.N.H.; Balzano, L.; Nair, S.; Zahedi, S.; Kohler, E.; Pottie, G.; Hansen, M.;Srivastava, M. Sensor network data fault types. ACM Trans. Sens. Netw. 2009, 5, 25. [CrossRef]13. Mahapatro, A.; Khilar, P.M. Fault diagnosis in wireless sensor networks: A survey. IEEE Commun.Surv. Tutor.2013, 15, 2000–2026. [CrossRef]14. Kapitanova, K.; Hoque, E.; Stankovic, J.A.; Whitehouse, K.; Son, S.H. Being SMART about failures:Assessing repairs in SMART homes. In Proceedings of the 2012 ACM Conference on Ubiquitous Computing,Pittsburgh, PA, USA, 5–8 September 2012; pp. 51–60.15. Flöck, M. Activity Monitoring and Automatic Alarm Generation in AAL-Enabled Homes; Logos VerlagBerlin GmbH: Berlin, Germany, 2010.16. Rahal, Y.; Pigot, H.; Mabilleau, P. Location estimation in a smart home: System implementation andevaluation using experimental data. Int. J. Telemed. Appl. 2008, 2008, 4. [CrossRef] [PubMed]17. Munir, S.; Stankovic, J.A. Failuresense: Detecting sensor failure using electrical appliances in the home.In Proceedings of the 2014 IEEE 11th International Conference on Mobile Ad Hoc and Sensor Systems(MASS), Philadelphia, PA, USA, 28–30 October 2014; pp. 73–81.18. Ye, J.; Stevenson, G.; Dobson, S. Fault detection for binary sensors in smart home environments.In Proceedings of the 2015 IEEE International Conference on Pervasive Computing and Communications(PerCom), St. Louis, MO, USA, 23–27 March 2015; pp. 20–28.19. Hnat, T.W.; Srinivasan, V.; Lu, J.; Sookoor, T.I.; Dawson, R.; Stankovic, J.; Whitehouse, K. The hitchhiker’sguide to successful residential sensing deployments. In Proceedings of the 9th ACM Conference onEmbedded Networked Sensor Systems, lSeattle, WA, USA, 1–4 November 2011; pp. 232–245.20. Kasteren Datasets. Available online: https://sites.google.com/site/tim0306/datasets (accessed on 22 March 2018).21. van Kasteren, T.L.; Englebienne, G.; Kröse, B.J. Human activity recognition from wireless sensor networkdata: Benchmark and software. In Activity Recognition in Pervasive Intelligent Environments; Springer:Berlin, Germany, 2011; pp. 165–186.22. CASAS Datasets. Available online: http://ailab.wsu.edu/casas/datasets/ (accessed on 9 March 2018).23. PlaceLab Datasets. Available online: http://web.mit.edu/cron/group/house_n/data/PlaceLab/PlaceLab.htm (accessed on 22 March 2018)24. Logan, B.; Healey, J.; Philipose, M.; Tapia, E.M.; Intille, S. A long-term evaluation of sensing modalitiesfor activity recognition. In Proceedings of the International Conference on Ubiquitous Computing,Innsbruck, Austria, 16–19 September 2007; Springer: Berlin, Germany, 2007; pp. 483–500.25. Tapia Datasets. Available online: http://courses.media.mit.edu/2004fall/mas622j/04.projects/home/(accessed on 12 May 2018)26. Ye, J.; Stevenson, G.; Dobson, S. Using temporal correlation and time series to detect missing activity-drivensensor events. In Proceedings of the 2015 IEEE International Conference on Pervasive Computing andCommunication Workshops (PerCom Workshops), St. Louis, MO, USA, 23–27 March 2015; pp. 44–49.27. Kodeswaran, P.A.; Kokku, R.; Sen, S.; Srivatsa, M. Idea: A system for efficient failure management insmart iot environments. In Proceedings of the 14th Annual International Conference on Mobile Systems,Applications, and Services, Singapore, 26–30 June 2016; pp. 43–56.Sensors 2018, 18, 1991 18 of 1928. Monekosso, D.N.; Remagnino, P. Data reconciliation in a smart home sensor network. Expert Syst. Appl.2013, 40, 3248–3255. [CrossRef]29. Ye, J.; Stevenson, G.; Dobson, S. Detecting abnormal events on binary sensors in smart home environments.Pervasive Mob. Comput. 2016, 33, 32–49. [CrossRef]30. Kapitanova, K.; Hoque, E.; Stankovic, J.A.; Son, S.H.; Whitehouse, K.; Alessandrelli, D. Being SMARTAbout Failures: Assessing Repairs in Activity Detection. In Proceedings of the 2012 ACM Conference onUbiquitous Computing, Pittsburgh, PA, USA, 5–8 September 2012.31. Veronese, F.; Pour, D.S.; Comai, S.; Matteucci, M.; Salice, F. Method, Design and Implementation of aSelf-checking Indoor Localization System. In Proceedings of the International Workshop on AmbientAssisted Living, Belfast, UK, 2–5 December 2014; pp. 187–194.32. Veronese, F.; Comai, S.; Matteucci, M.; Salice, F. Method, design and implementation of a multiuser indoorlocalization system with concurrent fault detection. In Proceedings of the 11th International Conference onMobile and Ubiquitous Systems: Computing, Networking and Services, London, UK, 2–4 December 2014;pp. 100–109.33. Mohamed, A.; Jacquet, C.; Bellik, Y. A fault detection and diagnosis framework for ambient intelligentsystems. In Proceedings of the 2012 9th International Conference on Ubiquitous Intelligence & Computingand 9th International Conference on Autonomic & Trusted Computing (UIC/ATC), Fukuoka, Japan,4–7 September 2012; pp. 394–401.34. Jacquet, C.; Mohamed, A.; Bellik, Y. An ambient assisted living framework with automatic self-diagnosis.Int. J. Adv. Life Sci. 2013, 5, 10.35. Oliveira, C.H.S.; Giroux, S.; Ngankam, H.; Pigot, H. Generating Bayesian Network Structures forSelf-diagnosis of Sensor Networks in the Context of Ambient Assisted Living for Aging Well. In Proceedingsof the International Conference on Smart Homes and Health Telematics, Paris, France, 29–31 August 2017;pp. 198–210.36. Amri, M.H.; Aubry, D.; Becis, Y.; Ramdani, N. Robust fault detection and isolation applied to indoorlocalization. IFAC PapersOnLine 2015, 48, 440–445. [CrossRef]37. Marhic, B.; Delahoche, L.; Solau, C.; Jolly-Desodt, A.M.; Ricquebourg, V. An evidential approach for detectionof abnormal behaviour in the presence of unreliable sensors. Inf. Fusion 2012, 13, 146–160. [CrossRef]38. Hong, X.; Nugent, C.; Mulvenna, M.; McClean, S.; Scotney, B.; Devlin, S. Assessment of the impact of sensorfailure in the recognition of activities of daily living. In Proceedings of the International Conference onSmart Homes and Health Telematics, Ames, IA, USA, 28 June–2 July 2008; pp. 136–144.39. Hong, X.; Nugent, C.; Mulvenna, M.; McClean, S.; Scotney, B.; Devlin, S. Evidential fusion of sensor data foractivity recognition in smart homes. Pervasive Mob. Comput. 2009, 5, 236–252. [CrossRef]40. Hong, X.; Nugent, C.; Liu, W.; Ma, J.; McClean, S.; Scotney, B.; Mulvenna, M. Uncertain informationmanagement for adl monitoring in smart homes. Intell. Patient Manag. 2009, pp. 315–332.41. Mckeever, S.; Ye, J.; Coyle, L.; Bleakley, C.; Dobson, S. Activity recognition using temporal evidence theory.J. Ambient Intell. Smart Environ. 2010, 2, 253–269.42. Liao, J.; Bi, Y.; Nugent, C. Evidence fusion for activity recognition using the Dempster-Shafer theory ofevidence. In Proceedings of the 9th International Conference on Information Technology and Applicationsin Biomedicine, Larnaca, Cyprus, 4–7 November 2009; pp. 1–4.43. Liao, J.; Bi, Y.; Nugent, C. Reasoning activity for smart homes using a lattice-based evidential structure.In Proceedings of the International Conference on Knowledge Science, Engineering and Management, Belfast,UK, 1–3 September 2010; pp. 186–197.44. Liao, J.; Bi, Y.; Nugent, C. Using the Dempster–Shafer theory of evidence with a revised lattice structure foractivity recognition. IEEE Trans. Inf. Technol. Biomed. 2011, 15, 74–82. [CrossRef] [PubMed]45. Javadi, E.; Moshiri, B.; Yazdi, H.S. Activity Recognition In Smart Home Using Weighted Dempster-ShaferTheory. Int. J. Smart Home 2013, 7, 23–34. [CrossRef]46. Ballardini, A.L.; Ferretti, L.; Fontana, S.; Furlan, A.; Sorrenti, D.G. An indoor localization system fortelehomecare applications. IEEE Trans. Syst. Man Cybern. Syst. 2016, 46, 1445–1455. [CrossRef]47. Ahvar, E.; Lee, G.M.; Han, S.N.; Crespi, N.; Khan, I. Sensor network-based and user-friendly user locationdiscovery for future smart homes. Sensors 2016, 16, 969. [CrossRef] [PubMed]48. Isermann, R. Model-based fault-detection and diagnosis–status and applications. Annu. Rev. Control 2005,29, 71–85. [CrossRef]Sensors 2018, 18, 1991 19 of 1949. Wu, J.; Xia, M.; Shu, S. Location Tracking of a Single Inhabitant in Smart Home: A Discrete Event SystemApproach. In Proceedings of the 2016 IEEE International Conference on Internet of Things (iThings) andIEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing(CPSCom) and IEEE Smart Data (SmartData), Chengdu, China, 15–18 December 2016; pp. 615–620.50. Bruckner, D.; Yin, G.Q.; Faltinger, A. Relieved commissioning and human behavior detection in AmbientAssisted Living Systems. Elektrotech. Inftech. 2012, 129, 293. [CrossRef]© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open accessarticle distributed under the terms and conditions of the Creative Commons Attribution(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Let’s block ads! (Why?)

Do you need any assistance with this question?
Send us your paper details now
We’ll find the best professional writer for you!


error: Content is protected !!