Available online at www.sciencedirect.com
ScienceDirect
Available online at www.sciencedirect.com Procedia Computer Science 00 (2017) 000–000
ScienceDirect
www.elsevier.com/locate/procedia
Procedia Computer Science 114 (2017) 401–407
Complex Adaptive Systems Conference with Theme: Engineering Cyber Physical Systems, CAS October 30 – November 1, 2017, Chicago, Illinois, USA
Fuzzy Inference System-based Recognition of Slow, Medium and Fast Running Conditions using a Triaxial Accelerometer Nizam Uddin Ahameda,*, Lauren Bensona, Christian Clermonta, Sean T. Osisa,b, Reed Ferber a,b a
Faculty of Kinesiology, University of Calgary, 2500 University Drive N.W. Calgary, Alberta T2N IN4 Running Injury Clinic, University of Calgary, 2500 University Drive N.W. Calgary, Alberta T2N IN4
b
Abstract This paper introduces a fuzzy inference system (FIS)-based model for recognizing running conditions using data collected with a triaxial accelerometer. Specifically, data from three axes of a triaxial accelerometer were used as the input, and various running conditions (slow, medium and fast) were considered the output of the FIS. The MATLAB® fuzzy toolbox, which includes processes such as fuzzification, sets of fuzzy rules, fuzzy inference engine and defuzzification, was used to model the system. Mamdani-type fuzzy modelling was selected for developing the FIS. The structure of the generated fuzzy inference system includes three fuzzy rules (using if-then) and an initial set of membership functions. The performance of the proposed FIS model was assessed using the root mean square error (RMSE), mean absolute error (MAE) and non-dimensional error index (NDEI), which were found to equal 0.059, 0.213 and 0.147, respectively, for the test data. Additionally, the correlation coefficients (r) and coefficient of determination (R2) between the FIS-predicted and the actual values were 0.89 and 0.81, respectively. Finally, the model performance accuracy was measured using Variance-Accounted-For (%VAF), which equaled 96.54%. Thus, the assessment of the overall performance suggests that the proposed FIS model has potential to detect slow, medium and fast running conditions. © 2017 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the Complex Adaptive Systems Conference with Theme: Engineering Cyber Physical Systems. Keywords: Accelerometer; Fuzzy Inference System; Running Condition.
* Corresponding author. Tel.: +1-403.220.8890 E-mail address:
[email protected] 1877-0509 © 2017 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the scientific committee of the Complex Adaptive Systems Conference with Theme: Engineering Cyber Physical Systems. 1877-0509 © 2017 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the Complex Adaptive Systems Conference with Theme: Engineering Cyber Physical Systems. 10.1016/j.procs.2017.09.054
Nizam Uddin Ahamed et al. / Procedia Computer Science 114 (2017) 401–407 Ahamed NU / Procedia Computer Science 00 (2017) 000–000
402
Nomenclature Accl FMS FIS MAE MF NDEI RMSE RMS SISO VAF
accelerometer data fast-medium-slow fuzzy inference system mean absolute error membership function
non-dimensional error index
root mean square error root mean square single-input and single-output variance-accounted-for
1. Introduction The number of studies in the field of human gait recognition using various artificial intelligence techniques has increased over the last two decades [1]. Fuzzy logic is an influential rule-based technique in the artificial intelligence domain that helps predict and identify patterns of human gait movement and running conditions [2, 3]. However, few studies have utilized FIS to predict running conditions (e.g. slow, medium and fast) and patterns. A recent study conducted by Mikołajewska et al. assessed the results of post-stroke gait re-education through the application of traditional and fuzzy-based analyses [4]. Specifically, these researchers assessed spatiotemporal gait parameters before and after therapy and compared the results using a fuzzy-based assessment tools. Xu et al. used fuzzy logic and neural network techniques and four parameters, namely the Staheli index, Chippaux-Smirak index, arch index and modified arch index, as the classification features and collected gait data using various instruments, such as a force platform, stereometric system, accelerometer and pressure platform [5]. Another fuzzy-and-neuralnetwork-based study focused on the gait performance, postural stability, and depression of Parkinson’s patients [6], using the Berg Balance Scale, the Dynamic Gait index, and the Geriatric Depression Scale. In another medicalbased expert system using fuzzy and neural networks for patients with idiopathic scoliosis, researchers extracted kinetic data and used time and walking parameters in the analysis [7]. Several studies have applied fuzzy systems for the development of a mechatronic-based system (such as a robotic haptic or exoskeleton device) that can be used for assessing gait [8-10]. Researchers have selected different input-output parameters related to walking, force and motion velocity, such as ground reaction force, impulse disturbance, and internal biological electromyography noise interference, for evaluating their fuzzy system with the aim of improving the tracking performance and driving the exoskeleton. In contrast, fuzzy logic has been applied for identifying different speed patterns in cases such as vehicles, electronic motors and wind turbines, but not for detecting human running speed [11-13]. While these aforementioned studies have shown the use of fuzzy-based systems, these studies have generally used multiple camera-based gait analysis systems for the input data. However, these gait analysis tools are relatively expensive and restricted to gait labs. In contrast, wireless accelerometers are inexpensive and portable. Unfortunately, few studies have incorporated FIS-analysis techniques using accelerometer gait data [14, 15]. Thus, the FIS-based recognition of running conditions using a triaxial accelerometer is a novel idea in gait analysis. Therefore, this study investigated whether the developed Single-Input and Single-Output (SISO)-based fuzzy model was able to accurately identify different running conditions. Specifically, we proposed the following hypotheses: (i) if the triaxial accelerometer yields low, average and high RMS values, the output will be slow, medium and fast running condition, respectively, and (ii) the calculated RMSE, MSE, VAF, NDEI and correlation results from the actual and predicted values are acceptable in terms of evaluating performance and error evaluation. To the best of our knowledge, this study provides the first FIS-based running conditions identification approach based on signals generated by a single triaxial accelerometer. 2. Methodology Three runners participated in the study and provided written informed consent. The runners did not have any history of disorder or pain in their lower limbs. This protocol was approved by the University of Calgary Conjoint
Nizam Uddin Ahamed et al. / Procedia Computer Science 114 (2017) 401–407 Ahamed NU / Procedia Computer Science 00 (2017) 000–000
403
Health Research Ethics Board and the runners were treated in accordance with the ethical standards of the Declaration of Helsinki. A 3D accelerometer (Shimmer3 GSR+®±8 g, Shimmer Inc., Dublin IE) was placed over top the sacrum of each runner to record center of mass acceleration at a sampling frequency of 201.03 Hz. The runner was then instructed to run on a treadmill for approximately 60 s (about 12,000 data points) for each of three running trials with a maximum 2-min rest interval between the trials. Accelerometer data were collected for the three axes of the accelerometer (X, Y and Z). Ten thousand data points from the middle of each trial were considered which was divided into 10 approximately 5-s data epoch windows (1,000 data points in each window) with no overlapping. The acceleration signal from each window was Band-pass filtered at 20-100 Hz using a thirdorder Butterworth filter. A feature extraction metric, namely the root mean square (RMS), was calculated from the X, Y and Z axes. To consider the three-dimensional acceleration uniformly, the resultant of the three axes was considered, and the corresponding RMS was ca lculated in a manner similar to that proposed by Purwar et al.[16]. RMS =
1 X2 Y 2 Z2 N
(1)
The calculated RMS value was adjusted automatically by the developed FIS system to obtain the exact connection between the input and the output. Statistical analysis was performed using MedCalc for Windows, version 17.6 (MedCalc Software Ostend, Belgium). 2.1. Fuzzy Inference System A fuzzy inference system (FIS) operates on knowledge stated in terms of ‘if-then’ rules and can be applied to predict the behavior of many undefined systems and data-driven decision-making process [17]. The main advantage of FIS is that it does not require any knowledge of the underlying physical process as a prerequisite for its application. Fig. 1 illustrated the simple block diagram of the developed Single-Input and Single-Output (SISO)based FIS system. As shown in the figure, the accelerometer data from three axes (X, Y and Z) were served as the input variables to the system, and the running conditions, which were classified as ‘slow’, ‘medium’ or ‘fast’, was the output of the system. The accelerometer data were first used as the input variables for a fuzzification process, wherein fuzzy arithmetic and criteria were applied based on the input variables with certain rules, and the final results were defuzzified to yield the output. A simple MATLAB code to evaluate the Mamdani FIS (predicted value) with actual data is given below. Here, ‘FMS’ was the name which stored actual data, ‘mamdani’ was the name of the developed FIS model, ‘evalfis’ was the built-in keyword which performed the evaluation process between actual (‘FMS’) and the generated data (‘fis’), and then, the predicted data from the model was stored in ‘output’. Finally, the actual data (‘FMS’) and the predicted data (‘output’) from the developed model were used to evaluate the performance and error. load FMS; fis = readfis(‘mamdani’); output = evalfis (FMS, fis); In this study, a single-input and single-output (SISO)-based Mamdani fuzzy model with the following three rules was developed: i) If Accl is Low, then Running Condition is Slow ii) If Accl is Average, then Running Condition is Medium iii) If Accl is High, then Running Condition is Fast Note that Accl represents the RMS value of the triaxial accelerometer data and that value was the classification of the running condition. Table 1 presents the overall parameters used to develop the FIS model. The first column presents the different conditions of the accelerometer data, which correspond to changes in the predicted running condition. Accordingly, the rules were set to change the output parameters according to the different conditions of the input parameters. Table 2 represents the fuzzy input and output ranges as well as fuzzy sets.
404
Ahamed NU / Procedia Computer Science 00 (2017) 000–000 Nizam Uddin Ahamed et al. / Procedia Computer Science 114 (2017) 401–407
iz
Fig. 1 Block Diagram of FIS Table 1 Different parameters of FIS Parameter type
Name/value
Input
01
Output
01
Rules
03
MF Type
Triangular-shaped
Defuzzification
Centroid
MF: Membership Function Table 2: Fuzzy sets for the input and output variables. Input Ranges Fuzzy Sets Results(RMS); Accelerometer 0.35-1.5 Low Running (RMS) 1.5-1.8 Average Conditions 1.7-2.0 High
Output Ranges
