Tri-axial Accelerometer Signals Research Articles

Parkinsonian Tremor Detection with Compact Convolutional Transformer from Bispectrum Representation of tri-Axial Accelerometer Signals.

After the breakthroughs of Transformer networks in Natural Language Processing (NLP) tasks, they have led to exciting progress in visual tasks as well. Nonetheless, there has been a parallel growth in the number of parameters and the amount of training data, which led to the conclusion that Transformers are not suited for small datasets. This paper is the first to convey the feasibility of Compact Convolutional Transformers (CCT) for the prediction of Parkinsonian postural tremor based on the Bispectrum (BS) representation of IMU accelerometer time series. The dataset includes tri-axial accelerometer signals collected unobtrusively in-the-wild while subjects are on a phone call, and labelled by neurologists and signal processing experts. The BS is a noise-immune, higher-order representation that reflects a signal's deviation from Gaussianity and measures quadratic phase coupling. We performed comparative classification experiments using the CCT, pre-trained CNNs such as VGG-16 and ResNet-50, and the conventional Vision Transformer (ViT). Our model achieves competitive prediction accuracy and F1 score of 96% with only 1.016 M trainable parameters, compared to the ViT with 21.659 M trainable parameters, in a five-fold cross-validation scheme. Our model also outperforms pre-trained CNNs such as VGG-16 and ResNet-50. Furthermore, we show that the performance gains are maintained when training on a larger dataset of BS images. Our effort here is motivated by the hypothesis that data-efficient transformers outperform transfer learning using pre-trained CNNs, paving the way for promising deep learning architecture for small-scale, novel and noisy medical imaging datasets.Clinical relevance- Novel deep learning model for unobtrusive prediction of Parkinsonian Postural Tremor from Bispectrum image representation of tri-axial accelerometer signals collected in-the-wild.

Machine learning-based classification of circadian rhythm characteristics for mild cognitive impairment in the elderly.

Using wrist-wearable sensors to ecological transient assessment may provide a more valid assessment of physical activity, sedentary time, sleep and circadian rhythm than self-reported questionnaires, but has not been used widely to study the association with mild cognitive impairment and their characteristics. 31 normal cognitive ability participants and 68 MCI participants were monitored with tri-axial accelerometer and nocturnal photo volumetric pulse wave signals for 14 days. Two machine learning algorithms: gradient boosting decision tree and eXtreme gradient boosting were constructed using data on daytime physical activity, sedentary time and nighttime physiological functions, including heart rate, heart rate variability, respiratory rate and oxygen saturation, combined with subjective scale features. The accuracy, precision, recall, F1 value, and AUC of the different models are compared, and the training and model effectiveness are validated by the subject-based leave-one-out method. The low physical activity state was higher in the MCI group than in the cognitively normal group between 8:00 and 11:00 (P < 0.05), the daily rhythm trend of the high physical activity state was generally lower in the MCI group than in the cognitively normal group (P < 0.05). The peak rhythms in the sedentary state appeared at 12:00-15:00 and 20:00. The peak rhythms of rMSSD, HRV high frequency output power, and HRV low frequency output power in the 6h HRV parameters at night in the MCI group disappeared at 3:00 a.m., and the amplitude of fluctuations decreased; the amplitude of fluctuations of LHratio nocturnal rhythm increased and the phase was disturbed; the oxygen saturation was between 90 and 95% and less than 90% were increased in all time periods (P < 0.05). The F1 value of the two machine learning algorithms for MCI classification of multi-feature data combined with subjective scales were XGBoost (78.02) and GBDT (84.04). By collecting PSQI Scale data combined with circadian rhythm characteristics monitored by wrist-wearable sensors, we are able to construct XGBoost and GBDT machine learning models with good discrimination, thus providing an early warning solution for identifying family and community members with high risk of MCI.

A Wearable System Based on Multiple Magnetic and Inertial Measurement Units for Spine Mobility Assessment: A Reliability Study for the Evaluation of Ankylosing Spondylitis.

Spinal mobility assessment is essential for the diagnostic of patients with ankylosing spondylitis. BASMI is a routine clinical evaluation of the spine; its measurements are made with goniometers and tape measures, implying systematic errors, subjectivity, and low sensitivity. Therefore, it is crucial to develop better mobility assessment methods. The design, implementation, and evaluation of a novel system for assessing the entire spine’s motion are presented. It consists of 16 magnetic and inertial measurement units (MIMUs) communicated wirelessly with a computer. The system evaluates the patient’s movements by implementing a sensor fusion of the triaxial gyroscope, accelerometer, and magnetometer signals using a Kalman filter. Fifteen healthy participants were assessed with the system through six movements involving the entire spine to calculate continuous kinematics and maximum range of motion (RoM). The intrarater reliability was computed over the observed RoM, showing excellent reliability levels (intraclass correlation >0.9) in five of the six movements. The results demonstrate the feasibility of the system for further clinical studies with patients. The system has the potential to improve the BASMI method. To the best of our knowledge, our system involves the highest number of sensors, thus providing more objective information than current similar systems.

Classification of the Korean Sign Language Alphabet Using an Accelerometer with a Support Vector Machine

Recognition and understanding of sign language can aid communication between nondeaf and deaf people. Recently, research groups have developed sign language recognition algorithms using multiple sensors. However, in everyday life, minimizing the number of sensors would still require the use of a sign language interpreter. In this study, a sign language classification method was developed using an accelerometer to recognize the Korean sign language alphabet. The accelerometer is worn on the proximal phalanx of the index finger of the dominant hand. Triaxial accelerometer signals were used to segment the sign gesture (i.e., the time period when a user is performing a sign) and recognize the 31 Korean sign language letters (producing a chance level of 3.2%). The vector sum of the accelerometer signals was used to segment the sign gesture with 98.9% segmentation accuracy, which is comparable to that of previous multisensor systems (99.49%). The system was able to classify the Korean sign language alphabet with 92.2% accuracy. The recognition accuracy of this approach was found to be higher than that of a previous work in the same sign language alphabet classification task. The findings demonstrate that a single-sensor accelerometer with simple features can be reliably used for Korean sign language alphabet recognition in everyday life.

Wearable Armband Device for Daily Life Electrocardiogram Monitoring.

A wearable armband electrocardiogram (ECG) monitor has been used for daily life monitoring. The armband records three ECG channels, one electromyogram (EMG) channel, and tri-axial accelerometer signals. Contrary to conventional Holter monitors, the armband-based ECG device is convenient for long-term daily life monitoring because it uses no obstructive leads and has dry electrodes (no hydrogels), which do not cause skin irritation even after a few days. Principal component analysis (PCA) and normalized least mean squares (NLMS) adaptive filtering were used to reduce the EMG noise from the ECG channels. An artifact detector and an optimal channel selector were developed based on a support vector machine (SVM) classifier with a radial basis function (RBF) kernel using features that are related to the ECG signal quality. Mean HR was estimated from the 24-hour armband recordings from 16 volunteers in segments of 10 seconds each. In addition, four classical HR variability (HRV) parameters (SDNN, RMSSD, and powers at low and high frequency bands) were computed. For comparison purposes, the same parameters were estimated also for data from a commercial Holter monitor. The armband provided usable data (difference less than 10% from Holter-estimated mean HR) during 75.25%/11.02% (inter-subject median/interquartile range) of segments when the user was not in bed, and during 98.49%/0.79% of the bed segments. The automatic artifact detector found 53.85%/17.09% of the data to be usable during the non-bed time, and 95.00%/2.35% to be usable during the time in bed. The HRV analysis obtained a relative error with respect to the Holter data not higher than 1.37% (inter-subject median/interquartile range). Although further studies have to be conducted for specific applications, results suggest that the armband device has a good potential for daily life HR monitoring, especially for applications such as arrhythmia or seizure detection, stress assessment, or sleep studies.

CMFALL: A Cascade and Parallel Multi-State Fall Detection Algorithm Using Waist-Mounted Tri-Axial Accelerometer Signals

As one of the main threats to people's health, especially for the elderly, falls have caused a large number of accidents. Detecting falls in time can minimize the severity of the injury and save lives. Therefore, designing fall detection algorithms to assist people's daily life has been a hot research topic. In the last decade, different fall detection approaches based on wearable sensors have been proposed. However, since the hardware resources of wearable sensors are very limited, designing accurate but energy-efficient fall detection algorithms remains an open challenge. To deal with this, an accurate but low-cost fall detection algorithm is proposed in this paper. Particularly, a novel cascade and parallel method that efficiently employs the characteristics of human falls and the advanced modeling ability of the neural network is proposed. Also, a novel design of a lightweight convolutional neural network with self-attention is proposed to achieve the best recognition/numerical complexity tradeoff. The proposed method, namely CMFALL, is evaluated together with a multitude of state-of-the-art models on a large dataset, where it performs the best with an F1-score exceeding 99%. Meanwhile, its computational cost and model size are only a few thousandths of other models. Such low computational cost and small size not only enable to embed it in a wearable sensor but also make the system power requirements quite low, which can enhance the autonomy of the wearable fall detector.

Driver Identification and Verification From Smartphone Accelerometers Using Deep Neural Networks

This paper addresses driver identification and verification using Deep Learning (DL) on tri-axial accelerometer signals from drivers’ smartphones. The proposed driver identification architecture includes ResNet-50 followed by two Stacked Gated Recurrent Units (SGRUs). ResNet provides a deep layer model, thanks to shortcut connections, is able to extract rich features from accelerometers, and GRU layers model the dynamics of drivers’ behavior. ResNet-50 pre-trained on image classification has been evaluated testing two approaches to map 1D accelerometer signals into 2D images. Siamese Neural Networks and Triplet Loss Training have been proposed for driver verification. The Siamese architecture is built on the same ResNet-50 + GRU model of driver identification, while the Triplet loss has required obtaining embeddings at journey level. Experimental results have been obtained for a dataset of 25 drivers, performing 20,025 daily life journeys with more than 800 per driver. Driver identification achieved top-1 and top-5 accuracies of 71.89% and 92.02%, respectively, and driver verification a F1 score of 74.09%. These results are competitive with state-of-the-art research that have generally tested smaller databases (in many cases based only on predefined routes), and have relied on information sources other than accelerometers, such as gyroscopes, magnetometers and GPS. Therefore, we believe that the proposed DL architectures are suitable for developing efficient driver monitoring applications based on only energy-efficient smartphone accelerometer signals.

Using Domain Knowledge for Interpretable and Competitive Multi-Class Human Activity Recognition

Human activity recognition (HAR) has become an increasingly popular application of machine learning across a range of domains. Typically the HAR task that a machine learning algorithm is trained for requires separating multiple activities such as walking, running, sitting, and falling from each other. Despite a large body of work on multi-class HAR, and the well-known fact that the performance on a multi-class problem can be significantly affected by how it is decomposed into a set of binary problems, there has been little research into how the choice of multi-class decomposition method affects the performance of HAR systems. This paper presents the first empirical comparison of multi-class decomposition methods in a HAR context by estimating the performance of five machine learning algorithms when used in their multi-class formulation, with four popular multi-class decomposition methods, five expert hierarchies—nested dichotomies constructed from domain knowledge—or an ensemble of expert hierarchies on a 17-class HAR data-set which consists of features extracted from tri-axial accelerometer and gyroscope signals. We further compare performance on two binary classification problems, each based on the topmost dichotomy of an expert hierarchy. The results show that expert hierarchies can indeed compete with one-vs-all, both on the original multi-class problem and on a more general binary classification problem, such as that induced by an expert hierarchy’s topmost dichotomy. Finally, we show that an ensemble of expert hierarchies performs better than one-vs-all and comparably to one-vs-one, despite being of lower time and space complexity, on the multi-class problem, and outperforms all other multi-class decomposition methods on the two dichotomous problems.

Physiological complexity of gait between regular and non-exercisers with Parkinson's disease

BackgroundPhysiological complexity represents overall health of a system and its underlying capacity to adapt to stresses. The primary purpose of this study was to determine if physiological complexity of gait both ON and OFF anti-Parkinson medication differed between regular and non-exercisers with Parkinson's disease. MethodsTwenty participants with idiopathic Parkinson's disease were enrolled in this cross-sectional study (regular exercisers n = 10, non-exercisers n = 10). Two data collection sessions were completed during a single visit, first after a 12-hour overnight withdrawal from anti-Parkinson medications (OFF), and again one-hour after taking anti-Parkinson medications (ON). During each session participants completed a 2-minute walking task at their preferred pace while wearing wireless inertial measurement units on each lower extremity segment (thigh, shank, foot). Multivariate multiscale entropy was calculated from the tri-axial accelerometer signals and converted to a complexity index for analysis. FindingsRegular exercisers demonstrated significantly higher complexity indices ON and OFF anti-Parkinson medications compared to non-exercisers (ON F = 3.84 P = 0.02; OFF F = 3.61, P < 0.03). Regular exercisers did not significantly differ in complexity between OFF and ON states (most affected leg F = 0.15 P = 0.71; least affected leg F = 0.30 P = 0.60), but non-exercisers demonstrated significantly decreased complexity in the least affected leg OFF anti-Parkinson medications (F = 5.17 P < 0.04). InterpretationEnhanced gait complexity in the regular exercisers may indicate that ongoing exercise is a key ingredient contributing to health in persons with Parkinson's disease. Exercising on a regular basis with Parkinson's disease may augment one's ability to adapt to barriers encountered during gait regardless of medication state.

PPG Derived Heart Rate Estimation During Intensive Physical Exercise

Accurate and reliable estimation of heart rate (HR) from photoplethysmographic (PPG) signals during moderate and vigorous physical activities is a challenging task, since intense motion artifacts can easily disguise the true HR. A novel method for estimating HR from PPG signal, during intensive physical exercise, is presented in this paper. The proposed method employs a recursive Wiener filtering technique for HR estimation from motion artifacts-corrupted PPG signal and simultaneously recorded the triaxial accelerometer signal. The experimental results demonstrated that the average relative error and the average absolute error of the proposed method on a public dataset (IEEE 2015 Signal Processing Cup Database) of 23 PPG recordings were 1.73 and 1.85 beats per minute, respectively. Our proposed approach is faster and more accurate than the existing proposals. Therefore, the proposed algorithm can be a reliable solution for HR estimation from noisy PPG signal.

Quantitative Assessment of Balance Impairment for Fall-Risk Estimation Using Wearable Triaxial Accelerometer

Falls are a major cause of morbidity and long-term hospitalization among growing older population. An automated and accurate fall-risk assessment system is vital to identify high fall-risk population and to prevent falls by early intervention. Therefore, this paper provides an objective, cost-effective, and unsupervised method to obtain functional balance and mobility assessment-based fall-risk of community-dwelling older adults. More specifically, waist-mounted triaxial accelerometer signals acquired from directed routine (supervised control movements) are used to estimate the well-known clinical assessment score-Berg balance scale (BBS). The trunk acceleration signals are used to extract features and to find the optimal subset of features for each training data during repeated tenfold cross validation of the BBS estimation model. The average of two BBS estimates based on test and retest yielded a strong correlation $\rho = 0.86$ with the standard BBS score. Also, high correlation ( $\rho = 0.90$ ) and low root-mean-square error (1.66) was observed between the two estimates of each subject. The proposed method is well suited for the assessment of balance impairment and pre-screening of quantitative fall-risk in an unsupervised setting. It has the potential to act as a surrogate of the standard clinical assessment-BBS.

A Novel Kalman Filter for Human Motion Tracking With an Inertial-Based Dynamic Inclinometer.

Design and development of a linear Kalman filter to create an inertial-based inclinometer targeted to dynamic conditions of motion. The estimation of the body attitude (i.e., the inclination with respect to the vertical) was treated as a source separation problem to discriminate the gravity and the body acceleration from the specific force measured by a triaxial accelerometer. The sensor fusion between triaxial gyroscope and triaxial accelerometer data was performed using a linear Kalman filter. Wrist-worn inertial measurement unit data from ten participants were acquired while performing two dynamic tasks: 60-s sequence of seven manual activities and 90s of walking at natural speed. Stereophotogrammetric data were used as a reference. A statistical analysis was performed to assess the significance of the accuracy improvement over state-of-the-art approaches. The proposed method achieved, on an average, a root mean square attitude error of 3.6° and 1.8° in manual activities and locomotion tasks (respectively). The statistical analysis showed that, when compared to few competing methods, the proposed method improved the attitude estimation accuracy. A novel Kalman filter for inertial-based attitude estimation was presented in this study. A significant accuracy improvement was achieved over state-of-the-art approaches, due to a filter design that better matched the basic optimality assumptions of Kalman filtering. Human motion tracking is the main application field of the proposed method. Accurately discriminating the two components present in the triaxial accelerometer signal is well suited for studying both the rotational and the linear body kinematics.

Improving the Energy Efficiency of Wearable Computing Units Using on Sensor Fifo Memory

Proliferation of wearable devices with wide spectrum of sensing capabilities together with commercial availability has increased the applicability of ambient intelligence concepts in practical system designs. Being wearable enforces extra constraints in terms of form factor and weight that limit the computational properties and the battery lifetime. There has been increasingly many number of studies for the energy efficiency of embedded and mobile hardware platforms. Due to the known techniques, increasing the energy consumption of an embedded system inherently requires some components to go into the low energy modes with a certain pattern, which in turn entails performance penalties at the application level. Existing solutions for increasing energy efficiency mainly focus only on a certain component of the system, such as hardware, networking firmware and try to achieve energy efficiency without considering the state the application is dynamically in. In this study, the critical balance between energy efficiency and application performance is handled. Application feedback is merged with energy efficiency and according to the application performance, duty cycle mechanism can be configured dynamically. A memory unit (FIFO) of the sensing component is also involved into the dynamic sleep scheduling mechanism in order to process latest sampled data while microprocessor and radio module of the sensor devices are in sleep mode. In this context, one of the fundamental implementations of ambient application which is based on triaxial accelerometer signal, pedometer is performed. Experiments realized on the dataset proved that it exists an interval where energy efficiency is obtained without degrading application performance under critical level and also usage of FIFO showed a significant impact on application performance and energy gain.

Human Activity Recognition Based on the Hierarchical Feature Selection and Classification Framework

Human activity recognition via triaxial accelerometers can provide valuable information for evaluating functional abilities. In this paper, we present an accelerometer sensor-based approach for human activity recognition. Our proposed recognition method used a hierarchical scheme, where the recognition of ten activity classes was divided into five distinct classification problems. Every classifier used the Least Squares Support Vector Machine (LS-SVM) and Naive Bayes (NB) algorithm to distinguish different activity classes. The activity class was recognized based on the mean, variance, entropy of magnitude, and angle of triaxial accelerometer signal features. Our proposed activity recognition method recognized ten activities with an average accuracy of 95.6% using only a single triaxial accelerometer.

Adaptive motion artefact reduction in respiration and ECG signals for wearable healthcare monitoring systems.

Wearable healthcare monitoring systems (WHMSs) have received significant interest from both academia and industry with the advantage of non-intrusive and ambulatory monitoring. The aim of this paper is to investigate the use of an adaptive filter to reduce motion artefact (MA) in physiological signals acquired by WHMSs. In our study, a WHMS is used to acquire ECG, respiration and triaxial accelerometer (ACC) signals during incremental treadmill and cycle ergometry exercises. With these signals, performances of adaptive MA cancellation are evaluated in both respiration and ECG signals. To achieve effective and robust MA cancellation, three axial outputs of the ACC are employed to estimate the MA by a bank of gradient adaptive Laguerre lattice (GALL) filter, and the outputs of the GALL filters are further combined with time-varying weights determined by a Kalman filter. The results show that for the respiratory signals, MA component can be reduced and signal quality can be improved effectively (the power ratio between the MA-corrupted respiratory signal and the adaptive filtered signal was 1.31 in running condition, and the corresponding signal quality was improved from 0.77 to 0.96). Combination of the GALL and Kalman filters can achieve robust MA cancellation without supervised selection of the reference axis from the ACC. For ECG, the MA component can also be reduced by adaptive filtering. The signal quality, however, could not be improved substantially just by the adaptive filter with the ACC outputs as the reference signals.

Assessment of the Cutting Tooling Effect on Turning Chatter

Machining performance is often limited by chatter vibration at the tool-workpiece interface. Chatter is a type of machining self-excited vibration which originates from the variation in cutting forces and the flexibility of the machine tool structure. Machining chatter is an inherently nonlinear phenomenon that is affected by many parameters such as cutting conditions, tool geometry, cutting speed, feed rate, depth of cut, overhang length of tool, clamping condition of workpiece. This study presents experimental approach for investigation of effects of various cutting tool geometry on the onset of chatter. In turning process, measured cutting force signal and triaxial accelerometer signal was used to know the characteristics of chatter vibration. The static and dynamic component of cutting forces reflect onset of chatter vibration. Proper selection of tooling is an important parameter in terms of chatter elimination in machining.

Analysis of gait and balance through a single triaxial accelerometer in presymptomatic and symptomatic Huntington's disease

PurposeTo investigate the capacity of a single triaxial accelerometer sensor in detecting gait and balance impairments in pre-manifest and manifest Huntington's disease (HD) subjects. MethodsFourteen manifest HD (MHD) (age: 51.83±14.8), ten pre-manifest HD (PHD) (age: 44.8±11.7) and ten healthy subjects (HLY) (age: 56.4±10.9) were recruited. The sensor was attached to the upper sternum as subjects completed gait and Romberg balance tests. An inverted pendulum model of the body's centre of mass and an unbiased autocorrelation procedure were employed to derive gait parameters from the triaxial accelerometer signal. The accuracy of the gait measurements was compared to those recorded by a computerized walkway. ResultsStrong agreement was seen between the sensor and the walkway; cadence (ICC=0.95, CI=[0.75, 0.97]), velocity (ICC=0.94, CI=[0.75, 0.97]) and step length (ICC=0.89, CI=[0.77, 0.95]). Sensor derived velocity was significantly higher in HLY (p<0.001) and PHD (p<0.005) when compared to MHD. Step and stride length was significantly longer in HLY (p<0.05) and PHD (p<0.001) when compared to MHD. Significant differences between subject groups across all four balance tasks (p<0.001) were found. ConclusionAn accelerometer based sensor may be an effective means of differentiating between pre-manifest and manifest Huntington's disease subjects.

A Real-Time and Self-Calibrating Algorithm Based on Triaxial Accelerometer Signals for the Detection of Human Posture and Activity

Assessment of human activity and posture with triaxial accelerometers provides insightful information about the functional ability: classification of human activities in rehabilitation and elderly surveillance contexts has been already proposed in the literature. In the meanwhile, recent technological advances allow developing miniaturized wearable devices, integrated within garments, which may extend this assessment to novel tasks, such as real-time remote surveillance of workers and emergency operators intervening in harsh environments. We present an algorithm for human posture and activity-level detection, based on the real-time processing of the signals produced by one wearable triaxial accelerometer. The algorithm is independent of the sensor orientation with respect to the body. Furthermore, it associates to its outputs a "reliability" value, representing the classification quality, in order to launch reliable alarms only when effective dangerous conditions are detected. The system was tested on a customized device to estimate the computational resources needed for real-time functioning. Results exhibit an overall 96.2% accuracy when classifying both static and dynamic activities.

Directionally Independent Failure Prediction of End-Milling Tools During Pocketing Maneuvers

Tracking the health of cutting tools under typical wear conditions is advantageous to the speed and efficiency of manufacturing processes. Existing techniques monitor tool performance through analyzing force or acceleration signals whereby prognoses are made from a single sensor type. This work proposes to enhance the spectral output of autoregressive (AR) models by combining triaxial accelerometer and triaxial dynamometer signals. Through parallel processing of force and acceleration signals using single six degree of freedom modeling, greater spectral resolution is achieved. Two entirely independent methods of tracking the tool wear are developed and contrasted. First, using the discrete cosine transform, primary component analysis will be applied to the spectral output of each AR auto and cross spectrum (Method 1). Each discrete cosine transform of the six-dimensional spectral data is analyzed to determine the magnitude of the critical (primary) variance energy component of the respective spectrum. The eigenvalues of these selected spectral energies are then observed for trends toward failure. The second method involves monitoring the eigenvalues of the spectral matrices centered at the toothpass frequency (Method 2). The results of the two methodologies are compared. Through the use of the eigenvalue method, it is shown that, for straight and pocketing maneuvers, both methods successfully track the condition of the tool using statistical thresholding. The techniques developed in this work are shown to be robust by multiple life tests conducted on different machine platforms with different operating conditions. Both techniques successfully identify impending fracture or meltdown due to wear, providing sufficient time to remove the tools before failure is realized.

Development of a Cutting Direction and Sensor Orientation Independent Monitoring Technique for End-Milling

In the authors’ previous work, univariate models were fit to acceleration data to predict impending tool failure. Numerous end-milling life tests, conducted under a wide variety of cutting conditions, demonstrated that the method could consistently warn of impending failure between 6 inches (15 cm) and 8 inches (20 cm) prior to the actual event. This paper presents an improved method that increases the warning time and allows the technique to function independent of the cutting direction or sensor orientation. Using multivariate autoregressive models fit to tri-axial accelerometer signals, monitoring indices are developed, verified and the results are compared with those from the univariate models. The multivariate models detected impending failure 30 inches (76 cm) prior to its occurrence, 23.5 inches (60 cm) earlier than with the univariate models. Furthermore, the multivariate models are able to monitor the condition of the tool, regardless of the cutting direction or sensor orientation. [S1087-1357(00)01003-0]