Raw Acceleration Data Research Articles

Evaluation of raw acceleration sedentary thresholds in children and adults.

The aim was to develop sedentary (sitting/lying) thresholds from hip and wrist worn raw tri-axial acceleration data from the ActiGraph and GENEActiv, and to examine the agreement between free-living time spent below these thresholds with sedentary time estimated by the activPAL. Sixty children and adults wore an ActiGraph and GENEActiv on the hip and wrist while performing six structured activities, before wearing the monitors, in addition to an activPAL, for 24h. Receiver operating characteristic (ROC) curves were used to determine sedentary thresholds based on activities in the laboratory. Agreement between developed sedentary thresholds during free-living and activPAL were assessed by Bland-Altman plots and by calculating sensitivity and specificity. Using laboratory data and ROC-curves showed similar classification accuracy for wrist and hip thresholds (Areaunder the curve=0.84-0.92). Greatest sensitivity (97-98%) and specificity (74-78%) were observed for thewrist thresholds, with no large differences between brands. During free-living, Bland-Altman plots showed large mean individual biases and 95% limits of agreement compared with activPAL, with smallest difference for the ActiGraph wrist threshold in children (+30min, P=0.3). Sensitivity and specificity for the developed thresholdsduring free-living were low for both age groups and for wrist (Sensitivity, 68-88%, Specificity, 46-59%) and hip placements (Sensitivity, 89-97%, Specificity, 26-34%). Laboratory derived sedentary thresholds generally overestimate free-living sedentary time comparedwith activPAL. Wrist thresholds appear to perform better than hip thresholds for estimating free-living sedentary time in children and adults relative to activPAL, however, specificity for all the developed thresholds are low.

Moving Forward with Backward Compatibility: Translating Wrist Accelerometer Data.

This study aimed to provide a means for calibrating raw acceleration data from wrist-worn accelerometers in relation to past estimates of children's moderate-to-vigorous physical activity (MVPA) from a range of cut points applied to hip-worn ActiGraph data. This is a secondary analysis of three studies with concurrent 7-d accelerometer wear at the wrist (GENEActiv) and hip (ActiGraph) in 238 children age 9-12 yr. The time spent above acceleration (ENMO) thresholds of 100, 150, 200, 250, 300, 350, and 400 mg from wrist acceleration data (≤5-s epoch) was calculated for comparison with MVPA estimated from widely used children's hip-worn ActiGraph MVPA cut points (Freedson/Trost, 1100 counts per minute; Pate, 1680 counts per minute; Evenson, 2296 counts per minute; Puyau, 3200 counts per minute) with epochs of ≤5, 15, and 60 s. The optimal ENMO thresholds for alignment with MVPA estimates from ActiGraph cut points determined from 70% of the sample and cross validated with the remaining 30% were as follows: Freedson/Trost = ENMO 150+ mg, irrespective of ActiGraph epoch (intraclass correlation [ICC] ≥ 0.65); Pate = ENMO 200+ mg, irrespective of ActiGraph epoch (ICC ≥ 0.67); Evenson = ENMO 250+ mg for ≤5- and 15-s epochs (ICC ≥ 0.69) and ENMO 300+ mg for 60-s epochs (ICC = 0.73); Puyau = ENMO 300+ mg for ≤5-s epochs (ICC = 0.73), ENMO 350+ mg for 15-s epochs (ICC = 0.73), and ENMO 400+ mg for 60-s epochs (ICC = 0.65). Agreement was robust with cross-validation ICC = 0.62-0.71 and means within ∣7.8∣% ± 4.9% of MVPA estimates from ActiGraph cut points, except Puyau 60-s epochs (ICC = 0.42). Incremental ENMO thresholds enable children's acceleration data measured at the wrist to be simply and directly compared, at a group level, with past estimates of MVPA from hip-worn ActiGraphs across a range of cut points.

Intensity Thresholds on Raw Acceleration Data: Euclidean Norm Minus One (ENMO) and Mean Amplitude Deviation (MAD) Approaches.

Objectives(1) To develop and internally-validate Euclidean Norm Minus One (ENMO) and Mean Amplitude Deviation (MAD) thresholds for separating sedentary behaviours from common light-intensity physical activities using raw acceleration data collected from both hip- and wrist-worn tri-axial accelerometers; and (2) to compare and evaluate the performances between the ENMO and MAD metrics.MethodsThirty-three adults [mean age (standard deviation (SD)) = 27.4 (5.9) years; mean BMI (SD) = 23.9 (3.7) kg/m2; 20 females (60.6%)] wore four accelerometers; an ActiGraph GT3X+ and a GENEActiv on the right hip; and an ActiGraph GT3X+ and a GENEActiv on the non-dominant wrist. Under laboratory-conditions, participants performed 16 different activities (11 sedentary behaviours and 5 light-intensity physical activities) for 5 minutes each. ENMO and MAD were computed from the raw acceleration data, and logistic regression and receiver-operating-characteristic (ROC) analyses were implemented to derive thresholds for activity discrimination. Areas under ROC curves (AUROC) were calculated to summarise performances and thresholds were assessed via executing leave-one-out-cross-validations.ResultsFor both hip and wrist monitor placements, in comparison to the ActiGraph GT3X+ monitors, the ENMO and MAD values derived from the GENEActiv devices were observed to be slightly higher, particularly for the lower-intensity activities. Monitor-specific hip and wrist ENMO and MAD thresholds showed excellent ability for separating sedentary behaviours from motion-based light-intensity physical activities (in general, AUROCs >0.95), with validation indicating robustness. However, poor classification was experienced when attempting to isolate standing still from sedentary behaviours (in general, AUROCs <0.65). The ENMO and MAD metrics tended to perform similarly across activities and accelerometer brands.ConclusionsResearchers can utilise these robust monitor-specific hip and wrist ENMO and MAD thresholds, in order to accurately separate sedentary behaviours from common motion-based light-intensity physical activities. However, caution should be taken if isolating sedentary behaviours from standing is of particular interest.

Wrist-independent energy expenditure prediction models from raw accelerometer data

Purposes: (1) Develop artificial neural network (ANN) models for wrist accelerometer data which can predict energy expenditure (EE) using data collected from either wrist. (2) Develop ANNs for detecting the wrist on which the accelerometer was worn.Forty-four adults wore GENEActiv accelerometers on the left and right wrists and a portable metabolic analyzer while participating in a 90 min semi-structured activity protocol. Participants performed 14 sedentary, lifestyle, exercise, and ambulatory activities and were allowed to choose activity order, duration, and intensity. ANNs were created to predict EE and wrist detection using a leave-one-out cross-validation. In total, 12 combinations of feature sets (mean and variance of raw, vector magnitude, and absolute value data), training methods (left- and right- wrist), and testing methods (left- and right-wrist data) were used to develop EE prediction ANNs. Accuracy of the ANNs was evaluated using correlations, root mean square error (RMSE), and bias, using metabolic analyzer data as the criterion for EE.ANNs using raw data from the same wrist (e.g. EE predicted from right wrist ANNs using accelerometer data from right wrist) had the highest accuracy for EE prediction (r = 0.84, RMSE = 1.25–1.26 METs); conversely, opposite-wrist prediction accuracy (e.g. EE predicted from right wrist ANNs using accelerometer data from left wrist) was lower (r = 0.60–0.64, RMSE = 1.93–2.01 METs). Preprocessing into absolute values prior to ANN development allowed for, high EE prediction accuracy, with no difference in accuracy for same- versus opposite-wrist prediction (r = 0.80–0.83, RMSE = 1.30–1.49 METs). Wrist detection ANNs correctly determined wrist placement 100% of the time.Highly accurate, wrist-independent EE prediction ANNs were developed by computing absolute values of raw acceleration data prior to ANN development. This method provides a potential approach for advancing predictive accuracy of wrist-worn accelerometers.

Comparison Of Outcomes Between Raw Acceleration And Counts-based Methods For Processing Wrist-worn Accelerometers

Evolution of accelerometry-based activity monitors has led to novel processing techniques of raw acceleration data collected at the wrist as opposed to the waist. However, few analyses have compared accelerometer-based activity estimates between emerging wrist thresholds. PURPOSE: The purpose was to compare moderate to vigorous physical activity (MVPA) time and prevalence of meeting youth physical activity guidelines (PAG) when using intensity thresholds for raw acceleration versus activity counts data. METHODS: Data are from the National Cancer Institute’s FLASHE study: a cross-sectional, surveillance study of youth/family behaviors related to cancer prevention. A sub-sample of 409 adolescents (11-17yrs; 50% female) wore an Actigraph GT3X+ monitor on the wrist over 7 days to provide objective data on physical activity. ActiLife software (version 6.10.2) was used to download all data. Raw data were processed with the R package GGIR and associated Hildebrand millig (mg) cut points. Counts data were processed using both Crouter and Chandler cut-points for vertical axis (VA) and vector magnitude (VM). MVPA was defined as METs ≥4. To compare the different processing techniques, differences in MVPA time and adherence to the youth PAG were evaluated using repeated ANOVAs with Bonferroni adjustments and Kappa statistics, respectively. RESULTS: The Hildebrand cut points yielded significantly fewer minutes of MVPA (M±SD: 10.6 ±14.7 min/d) compared to estimates from the Crouter cutpoints (107.3±56.4min/d for VA; 126.4±63.4min/d for VM) and Chandler cutpoints (52.3±34.3min/d for VA; 67.3±40.8min/d for VM): p-values <0.0001 for all comparisons. The prevalence of children meeting the PAG was 0% with Hilldebrand, but ranged substantially from 7% with Chandler VA to 64% with Crouter VM with the counts data. The mg method and counts methods showed very low classification agreement for the PAG (i.e. Kappa ranges from 0.01 to 0.06). Similar patterns were found when analyses were stratified by sex and weight status. CONCLUSIONS: Large discrepancies in MVPA estimates and PAG adherence rates were observed between the raw and counts data processing techniques. Using these different threshold-based methods may lead to disparate conclusions about youth behavior patterns and their relationships with health indicators.

Effect of Raw Acceleration Filtering Methods on the Relationship Between Accelerometer Outputs and Energy Expenditure

Despite the ubiquity of accelerometers in physical activity research, there remains no consensus method for processing raw acceleration data into end-user appropriate metrics. One aspect of accelerometer data processing that has received limited attention is the manner in which raw acceleration data are filtered to remove non-physical noise prior to computing summary outputs. PURPOSE: To evaluate how raw acceleration filtering methods affect the relationship between accelerometer outputs and measured energy expenditure in children and adolescents. METHODS: 122 children and adolescents (61 boys, 13.0 ± 4.3 years, 22.4 ± 6.2 kg/m2) wore an ActiGraph GT3X+ triaxial accelerometer (80 Hz) on their non-dominant wrist while performing jumping jacks (JJ), dribbling a basketball (DB), and ambulating on a treadmill at 2.4 (T1), 4.8 (T2), and 7.2 km/h (T3). Raw data for each axis were filtered nine times using 4th order Butterworth bandpass filters defined by all combinations of three lower (0.15, 0.25, and 0.35 Hz) and upper (2.5, 4.5, and 8.5 Hz) cutoff frequencies. Average signal vector magnitude (SVM) was calculated each second for all nine filters and the average 1 s value for the last 2 min of each 5 min activity trial was retained. Oxygen uptake (VO2) was measured by the COSMED K4b2 system concurrent with accelerometry measures. METs were derived by dividing mass relative VO2 by resting energy expenditure (Schofield equation estimate). Pearson correlations were used to quantify associations between SVM and METs for each activity/filter combination. Meng’s Z-test was used to evaluate differences between dependent correlations within each activity. RESULTS: Correlations between SVM and METs across the evaluated filters for JJ, DB, T1, T2, and T3 ranged from 0.43–0.55, 0.10–0.25, 0.11–0.15, 0.19–0.32, and 0.11–0.22, respectively. Significant between-filter differences (p < 0.05) in the magnitude of association between SVM and METs were noted for JJ (n = 20), DB (n = 20), T2 (n = 9), and T3 (n = 8). CONCLUSION: Filter choice can significantly influence the relationship between an accelerometer’s output and measured energy expenditure. Further study is needed to identify filtering methods which produce accelerometer outputs that best reflect the energy cost of human movement. Support: NIH NICHD 1R21HD073807-01A1

From Raw Acceleration Data to Activity Count

From Raw Acceleration Data to Activity Count

Performance of Activity Classification Algorithms in Free-Living Older Adults.

The objective of this study is to compare activity type classification rates of machine learning algorithms trained on laboratory versus free-living accelerometer data in older adults. Thirty-five older adults (21 females and 14 males, 70.8 ± 4.9 yr) performed selected activities in the laboratory while wearing three ActiGraph GT3X+ activity monitors (in the dominant hip, wrist, and ankle; ActiGraph, LLC, Pensacola, FL). Monitors were initialized to collect raw acceleration data at a sampling rate of 80 Hz. Fifteen of the participants also wore GT3X+ in free-living settings and were directly observed for 2-3 h. Time- and frequency-domain features from acceleration signals of each monitor were used to train random forest (RF) and support vector machine (SVM) models to classify five activity types: sedentary, standing, household, locomotion, and recreational activities. All algorithms were trained on laboratory data (RFLab and SVMLab) and free-living data (RFFL and SVMFL) using 20-s signal sampling windows. Classification accuracy rates of both types of algorithms were tested on free-living data using a leave-one-out technique. Overall classification accuracy rates for the algorithms developed from laboratory data were between 49% (wrist) and 55% (ankle) for the SVMLab algorithms and 49% (wrist) to 54% (ankle) for the RFLab algorithms. The classification accuracy rates for SVMFL and RFFL algorithms ranged from 58% (wrist) to 69% (ankle) and from 61% (wrist) to 67% (ankle), respectively. Our algorithms developed on free-living accelerometer data were more accurate in classifying the activity type in free-living older adults than those on our algorithms developed on laboratory accelerometer data. Future studies should consider using free-living accelerometer data to train machine learning algorithms in older adults.

Sedentary Sphere

Access to raw acceleration data should facilitate comparisons between accelerometer outputs regardless of monitor brand. To evaluate the accuracy of posture classification using the Sedentary Sphere in data from two widely used wrist-worn triaxial accelerometers. Laboratory: Thirty-four adults wore a GENEActiv and an ActiGraph GT3X+ on their nondominant wrist while performing four lying, seven sitting, and five upright activities. Free-living: The same participants wore both accelerometers on their nondominant wrist and an activPAL3 on their right thigh during waking hours for 2 d. Laboratory: Using the Sedentary Sphere with 15-s epoch GENEActiv data, sedentary and upright postures were correctly identified 74% and 91% of the time, respectively. Corresponding values for the ActiGraph data were 75% and 90%. Free-living: Total sedentary time was estimated at 534 ± 144, 523 ± 143, and 528 ± 137 min by the activPAL, the Sedentary Sphere with GENEActiv data and with ActiGraph data, respectively. The mean bias, relative to the activPAL, was small with moderate limits of agreement (LoA) for both the GENEActiv (mean bias = -12.5 min, LoA = -117 to 92 min) and ActiGraph (mean bias = -8 min, LoA = -103 to 88 min). Strong intraclass correlations (ICC) were evident for the activPAL with the GENEActiv (0.93, 0.84-0.97 (95% confidence interval) and the ActiGraph (0.94, 0.86-0.97). Agreement between the GENEActiv and ActiGraph posture classifications was very high (ICC = 0.98 (0.94-0.99), mean bias = +3 min, LoA = -58 to 63 min). These data support the efficacy of the Sedentary Sphere for classification of posture from a wrist-worn accelerometer in adults. The approach is equally valid with data from both the GENEActiv and ActiGraph accelerometers.

Working postures and physical activity among registered nurses

Nurses report a high prevalence of musculoskeletal discomfort, particularly of the low back and neck/shoulder. This study characterized the full-shift upper arm and trunk postures and movement velocities of registered nurses using inertial measurement units (IMUs). Intensity of occupational physical activity (PA) was also ascertained using a waist-worn PA monitor and using the raw acceleration data from each IMU. Results indicated that nurses spent a relatively small proportion of their work time with the arms or trunk in extreme postures, but had few opportunities for rest and recovery in comparison to several other occupational groups. Comparisons between nurses in different PA groups suggested that using a combination of accelerometers secured to several body locations may provide more representative estimates of physical demands than a single, waist-worn PA monitor. The findings indicate a need for continued field-based research with larger sample sizes to facilitate the development of maximally effective intervention strategies.

Towards unsupervised physical activity recognition using smartphone accelerometers

The development of smartphones equipped with accelerometers gives a promising way for researchers to accurately recognize an individual's physical activity in order to better understand the relationship between physical activity and health. However, a huge challenge for such sensor-based activity recognition task is the collection of annotated or labelled training data. In this work, we employ an unsupervised method for recognizing physical activities using smartphone accelerometers. Features are extracted from the raw acceleration data collected by smartphones, then an unsupervised classification method called MCODE is used for activity recognition. We evaluate the effectiveness of our method on three real-world datasets, i.e., a public dataset of daily living activities and two datasets of sports activities of race walking and basketball playing collected by ourselves, and we find our method outperforms other existing methods. The results show that our method is viable to recognize physical activities using smartphone accelerometers.

Sampling frequency affects the processing of Actigraph raw acceleration data to activity counts.

ActiGraph acceleration data are processed through several steps (including band-pass filtering to attenuate unwanted signal frequencies) to generate the activity counts commonly used in physical activity research. We performed three experiments to investigate the effect of sampling frequency on the generation of activity counts. Ideal acceleration signals were produced in the MATLAB software. Thereafter, ActiGraph GT3X+ monitors were spun in a mechanical setup. Finally, 20 subjects performed walking and running wearing GT3X+ monitors. Acceleration data from all experiments were collected with different sampling frequencies, and activity counts were generated with the ActiLife software. With the default 30-Hz (or 60-Hz, 90-Hz) sampling frequency, the generation of activity counts was performed as intended with 50% attenuation of acceleration signals with a frequency of 2.5 Hz by the signal frequency band-pass filter. Frequencies above 5 Hz were eliminated totally. However, with other sampling frequencies, acceleration signals above 5 Hz escaped the band-pass filter to a varied degree and contributed to additional activity counts. Similar results were found for the spinning of the GT3X+ monitors, although the amount of activity counts generated was less, indicating that raw data stored in the GT3X+ monitor is processed. Between 600 and 1,600 more counts per minute were generated with the sampling frequencies 40 and 100 Hz compared with 30 Hz during running. Sampling frequency affects the processing of ActiGraph acceleration data to activity counts. Researchers need to be aware of this error when selecting sampling frequencies other than the default 30 Hz.

Validation of Cut-Points for Evaluating the Intensity of Physical Activity with Accelerometry-Based Mean Amplitude Deviation (MAD).

PurposeOur recent study of three accelerometer brands in various ambulatory activities showed that the mean amplitude deviation (MAD) of the resultant acceleration signal performed best in separating different intensity levels and provided excellent agreement between the three devices. The objective of this study was to derive a regression model that estimates oxygen consumption (VO2) from MAD values and validate the MAD-based cut-points for light, moderate and vigorous locomotion against VO2 within a wide range of speeds.Methods29 participants performed a pace-conducted non-stop test on a 200 m long indoor track. The initial speed was 0.6 m/s and it was increased by 0.4 m/s every 2.5 minutes until volitional exhaustion. The participants could freely decide whether they preferred to walk or run. During the test they carried a hip-mounted tri-axial accelerometer and mobile metabolic analyzer. The MAD was calculated from the raw acceleration data and compared to directly measured incident VO2. Cut-point between light and moderate activity was set to 3.0 metabolic equivalent (MET, 1 MET = 3.5 ml · kg-1 · min-1) and between moderate and vigorous activity to 6.0 MET as per standard use.ResultsThe MAD and VO2 showed a very strong association. Within individuals, the range of r values was from 0.927 to 0.991 providing the mean r = 0.969. The optimal MAD cut-point for 3.0 MET was 91 mg (milligravity) and 414 mg for 6.0 MET.ConclusionThe present study showed that the MAD is a valid method in terms of the VO2 within a wide range of ambulatory activities from slow walking to fast running. Being a device-independent trait, the MAD facilitates directly comparable, accurate results on the intensity of physical activity with all accelerometers providing tri-axial raw data.

Mean amplitude deviation calculated from raw acceleration data: a novel method for classifying the intensity of adolescents' physical activity irrespective of accelerometer brand.

BackgroundUsing raw acceleration data to assess the intensity of physical activity enables direct comparisons between studies using different accelerometer brands. Mean amplitude deviation (MAD in mg) calculated from resultant tri-axial raw acceleration signal was recently shown to perform best in classifying the intensity of physical activity in adults irrespective of the accelerometer brand. This study compared MAD values and cut-points between two different accelerometers in adolescents.MethodsTwenty voluntary participants (10 girls and 10 boys) of average age of 14 wore two accelerometers (Actigraph GTX3, Pensacola FL, USA and Hookie AM13, Espoo, Finland) and heart rate monitors (M61, Polar Electro Oy, Kempele, Finland) while completing ten 2-min patterns of typical activities ranging from sedentary behaviour to light, moderate and vigorous-intensity locomotion. Bland-Altman method examined the agreement of MAD values between the accelerometers. Correlation coefficient between individual heart rates and MAD values indicated the validity of pattern-based intensity classification. Generalized ordinal logistic regression determined the intensity-specific MAD cut-points for both accelerometers.ResultsMAD values varied from 3 mg (lying supine) to 1609 mg (running). Hookie gave higher values than Actigraph in accelerations exceeding 700 mg. The correlation coefficient between MAD values and heart rates was 0.96 for Hookie and 0.97 for Actigraph. Respectively, the MAD cut-points were 29 and 27 (light), 338 and 330 (moderate), and 604 and 558 (vigorous).ConclusionsMAD values and cut-points of Hookie and Actigraph showed excellent agreement. Analysing raw accelerometer data with MAD values may enable the comparison of accelerometer results between different studies also in adolescents.

Features and models for human activity recognition

Human Activity Recognition (HAR) is aimed at identifying current subject task performed by a person as a result of analyzing data from wearable sensors. HAR is a very challenging task that has been applied in different areas such as rehabilitation and localization. During the past ten years, plenty of models, number of sensors and sensor placements, and feature transformations have been reported for this task. From this bunch of previous ideas, what seems to be clear is that the very specific applications drive to the selection of the best choices for each case.Present research is focused on early diagnosis of stroke, what involves reducing the feature space of gathered data and subsequent HAR, among other tasks. In this study, an Information Correlation Coefficient (ICC) analysis was carried out followed by a wrapper Feature Selection (FS) method on the reduced input space. Additionally, a novel HAR method is proposed for this specific problem of stroke early diagnosing, comprising an adaptation of the well-known Genetic Fuzzy Finite State Machine (GFFSM) method.To the best of the author׳s knowledge, this is the very first analysis of the feature space concerning all the previously published feature transformations on raw acceleration data. The main contributions of this study are the optimization of the sample rate, selection of the best feature subset, and learning of a suitable HAR method based on GFFSM to be applied to the HAR problem.

Exploring the ActiLife® filtration algorithm: converting raw acceleration data to counts

Though portable accelerometers are ubiquitous in physiology and public health studies, their accuracy as objective measures of physical activity is still being examined. This paper enumerates and analyzes the various biases of the widely used ActiLife® software in reporting activity counts from ActiGraph® accelerometers. In particular, we focus on the two-stage proprietary filtration algorithm used to convert raw acceleration data, for a sampling rate of 30 Hz, to compressed 1 Hz signals; we develop simple novel methods to analyze the action of the software filter on the raw data in the frequency domain.

Love thy neighbour: automatic animal behavioural classification of acceleration data using the K-nearest neighbour algorithm.

Researchers hoping to elucidate the behaviour of species that aren’t readily observed are able to do so using biotelemetry methods. Accelerometers in particular are proving particularly effective and have been used on terrestrial, aquatic and volant species with success. In the past, behavioural modes were detected in accelerometer data through manual inspection, but with developments in technology, modern accelerometers now record at frequencies that make this impractical. In light of this, some researchers have suggested the use of various machine learning approaches as a means to classify accelerometer data automatically. We feel uptake of this approach by the scientific community is inhibited for two reasons; 1) Most machine learning algorithms require selection of summary statistics which obscure the decision mechanisms by which classifications are arrived, and 2) they are difficult to implement without appreciable computational skill. We present a method which allows researchers to classify accelerometer data into behavioural classes automatically using a primitive machine learning algorithm, k-nearest neighbour (KNN). Raw acceleration data may be used in KNN without selection of summary statistics, and it is easily implemented using the freeware program R. The method is evaluated by detecting 5 behavioural modes in 8 species, with examples of quadrupedal, bipedal and volant species. Accuracy and Precision were found to be comparable with other, more complex methods. In order to assist in the application of this method, the script required to run KNN analysis in R is provided. We envisage that the KNN method may be coupled with methods for investigating animal position, such as GPS telemetry or dead-reckoning, in order to implement an integrated approach to movement ecology research.

An Unsupervised Approach for Automatic Activity Recognition Based on Hidden Markov Model Regression

Using supervised machine learning approaches to recognize human activities from on-body wearable accelerometers generally requires a large amount of labeled data. When ground truth information is not available, too expensive, time consuming or difficult to collect, one has to rely on unsupervised approaches. This paper presents a new unsupervised approach for human activity recognition from raw acceleration data measured using inertial wearable sensors. The proposed method is based upon joint segmentation of multidimensional time series using a Hidden Markov Model (HMM) in a multiple regression context. The model is learned in an unsupervised framework using the Expectation-Maximization (EM) algorithm where no activity labels are needed. The proposed method takes into account the sequential appearance of the data. It is therefore adapted for the temporal acceleration data to accurately detect the activities. It allows both segmentation and classification of the human activities. Experimental results are provided to demonstrate the efficiency of the proposed approach with respect to standard supervised and unsupervised classification approaches.

Wireless system for the continuous observation of whole-body vibration in heavy machinery

Several occupational groups are exposed to substantial amounts of vibration daily due to normal working routines, like operating heavy equipment or driving vehicles on uneven surfaces. Vibration that affects the whole human body and is transferred to the body through supporting systems such as the seat or floor is defined as whole-body vibration (WBV). Studies have shown that longterm exposure to WBV creates adverse health effects, of which lower back pain is the most common. Such symptoms may affect a person's quality of life or even cause incapacitation. Even though a correlation between health problems and long-term WBV exposure is evident, mechanisms behind the symptoms have remained unexplained. This paper presents a novel approach to the continuous observation of WBV in heavy machinery. The key elements of the approach are the incorporation of a measurement system into heavy machinery, enhanced comfort, ease of use, and cost-effectiveness. Small alternative sensor devices that improve comfort with respect to the standardized method were investigated in wired and wireless form. Wireless data transmission was then utilized to provide effortless and cost-effective data acquisition from the machinery to the remote server. The system was able to execute the calculations, transfer the resulting parameters to a remote server where information about predominant vibration conditions could be accessed instantaneously, and acquire and store raw acceleration data for further research.

Abnormal state detection of building structures based on symbolic time series analysis and negative selection

A symbolization-based negative selection (SNS) algorithm that combines the advantages of symbolic time series analysis (STSA) and negative selection (NS) is proposed for detecting the abnormal states of building structures. In SNS, no prior knowledge of a structure's abnormal state is needed. Only the response of the structure in a current state is used as input data. In addition, this approach can work even with one sensor, so it is highly practical and flexible. A state sequence histogram (SSH) transformed from raw acceleration data by using STSA can capture the main features of structure dynamics and alleviate the effects of harmful noise. SSHs of the normal and abnormal states of a structure are defined as self and non-self elements, respectively. A new detector generation strategy and matching mechanism is proposed that makes the procedure more effective, along with guidelines for appropriate parameters in SNS. Numerical simulations and experimental verifications for different abnormal state cases were conducted to demonstrate the feasibility of the proposed method. Copyright © 2013 John Wiley & Sons, Ltd.