To ensure that the performance of the system and the Mental Workload MWL of the human operator are maintained at an acceptable level, one possible approach is to introduce real‐time adap
Trang 1Lars Joyce Planke
BEng (Electrical) (Hon 1st Class) (RMIT University)
School of Engineering College of Science, Technology, Engineering and Maths
RMIT University May 2021
Trang 2Declaration
I certify that except where due acknowledgement has been made, this work is that of the author alone; this work has not been submitted previously, in whole or in part, to qualify for any other academic award; the content of the thesis is the result of work, which has been carried out since the official commencement date of the approved research program; any editorial work, paid or unpaid, carried out by a third party is acknowledged; and, ethics procedures and guidelines have been followed.
I acknowledge the support I have received for my research through the provision of an Australian Government Research Training Program Scholarship.
Trang 3Acknowledgements
In conducting my Master of Engineering research program, I would firstly like to express my gratitude to my supervisors Prof. Roberto (Rob) Sabatini and Dr. Alessandro (Alex) Gardi, in which their advice, guidance and aspirations have been of utmost importance in completing
a successful research project. I would like to give my thanks to Rob for identifying a potential
in me, opening up invaluable opportunities and aspiring me to strive for excellence. I would also like to thank Alex for providing time and thoughtful advice throughout the project.
A special thanks goes to my friends and colleagues at RMIT University, including Yixiang Lim, Nichakorn Pongsarkornsathien, Sam Hilton, Rohan Kapoor, Suraj Bijjahalli and Federico Rivalta. Their accompanying presence have made this endeavour highly rewarding, with stimulating interactions that have sparked valuable insights and helped progress my research.
I would also like to give many thanks to Dr. Neta Ezer from Northrop Grumman Corporation for providing helpful feedback and supporting this research project. Also, thanks to Trevor Kistan from Thales Australia for helping me conduct this research and providing feedback. I would also like to acknowledge and express my gratitude to RMIT University for selecting
me for a RMIT Research Stipend Scholarship.
This research project has been conducted amid a historic pandemic, and a particular gratitude goes to my partner Bella Reid and her family for providing extra support through unusual and uncertain times, without which this research would not have been as successful. Finally, I would to like to give my heartfelt appreciation to all my friends and family in Norway and Australia. No matter the physical distance I can always rely on their encouragement, care and constant source of support.
Trang 8
6.6.3 Threshold criterion _ 220
Results 221
6.7.1 Round 1 results 221
6.7.2 Round 2 results 223
Discussion _ 232
6.8.1 Discussion of methodology design and networking considerations _ 232
6.8.2 Discussion of results 234
6.8.3 Discussion of contribution 239
6.8.4 Discussion of multimodal inference of MWL for CHMS _ 241
Conclusion _ 244
References 246
Introduction 249
Synthesis of experimental findings 249
Conclusion _ 254
7.3.1 Objectives achieved _ 256
Recommendations for future research _ 259
Appendix A – Subject Specific Feature Combination 261
Appendix B – Neuropype Pipeline Designer 264
Appendix C – Results from Normality Test 266
Appendix D – List of Publications 268
Trang 9
List of Figures
Figure 1.1 Fundamental concept of the Cognitive Human Machine System (CHMS)………5
Figure 1.2 The general project methodology……….……… 14
Figure 1.3 General methodology for Experimental Activity 1……….…….15
Figure 2.1 Sheridan scale……….……… 26
Figure 2.2 Full CHMS framework ……….……… 31
Figure 2.3 VPA system architecture……….……….33
Figure 2.4 Integrated air‐ground Concepts of Operation for SPO and UAS remote control 35
Figure 2.5 Layers of control for UAS and SPO……….………36
Figure 2.6 Third generation flight deck concept by Thales……….…… 37
Figure 2.7 Evolutionary paths for a multidomain air and space transport operation.….… 38
Figure 2.8 Inverted U model……… …………39
Figure 2.9 Relationship between task load, performance and MWL……….40
Figure 2.10 Conceptual relationship between MWL and SA………41
Figure 2.11 Tasks for the MATB program……….44
Figure 2.12 International 10‐20 system for electrode placement.……… …50
Figure 2.13 (a) Referential montage; (b) Differential amplifier circuit……….………51
Figure 2.14 A simulated head for finding oscillating voltage sources……….… …… 53
Figure 2.15 Signal processing method for a straightforward oscillatory BCI………56
Figure 2.16 EEG signals spatially filtered using CSP algorithms………58
Figure 2.17 Filter Bank Common Spatial Pattern (FBCSP)……….………59
Figure 2.18 A generic signal processing for ERP based BCI……….……….61
Figure 2.19 QRS complex……… 65
Figure 2.20 Example of different modalities that can be used for cognitive load……… 70
Figure 2.21 High level data fusion for multimodal MWL estimation……… 71
Figure 2.22 Example of an architecture of a NFS………76
Figure 3.1 Basic configuration of CHMS……….…………96
Figure 3.2 Fundamental components of sensing and estimation……… ………98
Figure 3.3 The general project methodology……… …….103
Figure 3.4 OTM UAV wildfire detection scenario……… … …….106
Figure 3.5 Tasks for the MATB program……… … 107
Figure 3.6 (a) actiCAP Xpress form BrainProducts; (b) V‐Amp amplifier with USB cable… ……111
Figure 3.7 GP3 eye tracker……… …… 112
Figure 3.8 Zephyr Bioharness 3……….… …….112
Trang 10Figure 3.9 Logitech Extreme 3D Pro……….……….113
Figure 3.10 Example of eye tracker mistakenly detecting features of the EEG electrodes as… 114
Figure 3.11 (a) Cloth that is placed to cover the electrodes of the EEG; (b) Cloth preventing… 114
Figure 3.12 Basic data flow for physiological measurement collection………115
Figure 3.13 General methodology for activity 1………119
Figure 3.14 Overall methodology for Experimental Activity 1……….……… 121
Figure 3.15 Multimodal data fusion with testing of multiple inference models……….123
Figure 3.16 Detailed methodology for offline development and analysis of inference………… 124
Figure 3.17 Offline calibration and validation of ANFIS model……… 125
Figure 3.18 Overall methodology for Experimental Activity 3……… 127
Figure 4.1 Mission concept illustrating the Unmanned Aerial Vehicles (UAVs) bounding…… 134
Figure 4.2 Subjective ratings: (a) subjective workload of all participants, grouped by the…….141
Figure 4.3 (a) Mean task index of all participants (normalized), grouped by the scenario…… 142
Figure 4.4 Physiological measures: (a) mean scan pattern entropy of all participants………….143
Figure 4.5 Mean EEG index of all participants……….…144
Figure 4.6 Comparison for one participant between the task index, SPE and EEG index……… 145
Figure 5.1 (a) Experimental setup while participant performs MATB tasks. A: EEG cap……… 155
Figure 5.2 MATB task load profile for the experiment………156
Figure 5.3 Diagram showing all the categories and subcategories of data collection……….157
Figure 5.4 Feature extraction using FBSPoC and continuous classification using ridge………….159
Figure 5.5 Pipeline for offline calibration and validation of EEG model……… 160
Figure 5.6 Four features extracted from the eye tracking data……… 162
Figure 5.7 ROI for calculating simple (a) and complex; (b) scan pattern entropy……….… 163
Figure 5.8 Network diagram for processing and collection of data……….169
Figure 5.9 Protocol for synchronising measurements………170
Figure 5.10 Result for participant 11 after calibrating the EEG model……… 172
Figure 5.11 Results from the EEG pipeline for participant six……… 174
Figure 5.12 (a) EEG measure for participant 4; (b) scan pattern entropy for participant 2………176
Figure 5.13 (a) Blinks per minute for participant 16; (b) pupil diameter for participant 2……… 176
Figure 5.14 (a) Dwell time for participant 5; (b) Heart rate for participant 3………177
Figure 5.15 Control inputs for participant 14……….177
Figure 5.16 Calibration and validation on separate halves of the data set……….182
Figure 5.17 Results after calibrating and validating the ANFIS model for participant 4……….… 189
Figure 6.1 Overall methodology……… 214
Figure 6.2 Task profile for Round 1 of the MATB scenario……… …… 215
Figure 6.3 Task profile for the online validation round in Round 2………216
Trang 11Figure 6.4 Online validation network for Experimental Activity 3……….….… 217
Figure 6.5 Round 1 network for Experimental Activity 3……… 218
Figure 6.6 Round 2 network for Experimental Activity 3……… 219
Figure 6.7 ANFIS model 2 from participant 5 with a MAE = 0.37………222
Figure 6.8 Task level (blue) and ANFIS output (red): (a) ANFIS output 4; (b) ANFIS…… ………….225
Figure 6.9 Task level (blue) and ANFIS output (red): (a) ANFIS output 7; (b) ANFIS………… 225
Figure 6.10 Task level (blue) and ANFIS output (red): (a) ANFIS output 11; (b) ANFIS.………225
Figure 6.11 EEG model result for participant 5……… 230
Figure 6.12 EEG model result for participant 6.……….230
Figure 6.13 EEG model result for participant 7.……….230
Figure 6.14 EEG model result for participant 4.……….230
Trang 12
List of Tables
Table 2.1 Considerations for HMI2 evolution………29
Table 3.1 Primary and secondary mission objectives………105
Table 3.2 Defining levels of difficulty for MATB scenario……… 109
Table 3.3 All noteworthy software’s used as part of this research……….116
Table 4.1 Task index calculation for each UAV……….………137
Table 4.2 All results from the ANOVA analysis……… …… 140
Table 4.3 Notable correlation coefficients for each participant……….…… 145
Table 4.4 All feature comparison. Mean and respective standard deviation across all………146
Table 5.1 Detailed task load profile during the experiment……… …… 157
Table 5.2 EEG model results for all participants……… 173
Table 5.3 Results from calibrating on all the data and validating with five‐fold validation…… 175
Table 5.4 Correlation between the task analysis (levels) and all other features for all……….179
Table 5.5 Correlation between features averaged for each participant……….180
Table 5.6 All noteworthy feature combinations used to calibrate and validate the ANFIS………184
Table 5.7 Eye feature combinations only used for training ANFIS………185
Table 5.8 Results for all participants with best feature combinations and all seven features… 186
Table 5.9 Best feature combination without control inputs………187
Table 5.10 Final ANFIS models based on generic combinations and subject specific………188
Table 5.11 Calibrate and validate on all data with the feature combinations of interest………….190
Table 6.1 Requirements for task profile for Round 1……….… 215
Table 6.2 Requirements for task profile for online validation in Round 2……… …… ….216
Table 6.3 MAE from online inference of MWL using a cross‐session ANFIS models 1‐5………… 222
Table 6.4 EEG result from offline calibration with MAE from five‐fold validation………… ……….223
Table 6.5 CC across all subjects and for all features……….…………223
Table 6.6 MAE and CC for Round 2 between task level and the ANFIS outputs……….227
Table 6.7 MAE and CC between online EEG inference and task profile in Round 2……….229
Table 6.8 CC between all subject……… 232
Trang 13
List of Acronyms and Abbreviations
ASSAP Airborne Surveillance and
Separation Assurance Processing AOR Area of Responsibility
ANOVA Analysis of Variance
aCAMS automated Cabin Air
Management System BCI Brain Computer Interface
DE Differential Evolution DEACS Differential Evolution with
Ant Colony Search DFHM Dual Frequency Head Maps DWELL Proportional Dwell Time DIA Pupil Diameter ECG Electrocardiogram EEG Electroencephalogram EMG Electromyogram EMI Electro Magnetic Interference EOG Electrooculogram ERP Event Related Potential FBCSP Filter Bank CSP FBCSP Filter Bank SPoC FCM Fuzzy C‐Means fMRI Functional Magnetic
Resonance Imaging fNIR Functional Near Infrared
Spectroscopy FFT Fast Fourier Transform FIS Fuzzy Inference System
FO First Officer GCS Ground Control Station
Trang 14
HbR Deoxygenated Hemoglobin
HMI Human Machine Interface
HMI2 Human Machine Interfaces
and Interactions HRV Heart Rate Variability
NN Normal to Normal OTM One‐to‐Many OFS Operator Functional State pBCI passive BCI
PY Python PBO Performance Based Optimisation PCA Principal Component Analysis PSD Power Spectral Density PET Position Emission Tomography
PF Pilot Flying PNF Pilot Not Flying
PM Pilot Managing RDA Remote Data Access RESMAN Resource Management RMSE Root Mean Square Error RPAS Remote Piloted Aircraft System
RR R‐to‐R RMIT Royal Melbourne Institute
of Technology
RQ Research Question ROI Region of Interest SCHED Schedule
Trang 15
Trang 16
is imperative that human operators maintain the required level of Situational Awareness (SA) in order to contribute effectively to both strategic and tactical decision‐making processes. On the other hand, it is also essential that the cognitive ability of human operators is constantly monitored to assess their ability to perform effectively in a closed loop human‐machine system environment. To ensure that the performance of the system and the Mental Workload (MWL) of the human operator are maintained at an acceptable level, one possible approach is to introduce real‐time adaptation in the Human‐Machine Interfaces and Interactions (HMI2). While current aerospace systems and interfaces are limited in adaptability, a Cognitive Human Machine System (CHMS) addresses these issues with a cyber‐physical human design that provides dynamic real‐time system adaptation. Nevertheless, to reliably drive adaptation of current and emerging aerospace systems there
is a need to accurately and repeatably estimate cognitive states, in particular MWL, in real‐time. Henceforth, this research has studied methods for sensing physiological and behavioural responses associated with MWL and have used the corresponding measures to provide a real‐time multimodal inference of MWL.
As part of this research, three experimental activities have been conducted. Experimental Activity 1 included an exploratory study that implemented and analysed an Electroencephalogram (EEG) index as well as a straightforward data fusion method during a complex One‐to‐Many (OTM) Unmanned Aerial Vehicle (UAV) wildfire detection scenario. The EEG index included previously proven features of MWL and showed to be sensitive to changes in MWL in the complex task scenario using the Correlation Coefficient (CC). Moreover, a straightforward data fusion approach showed that fusing the EEG index and an eye activity feature gave the highest correlation with a secondary task performance measure (CC = 0.73 ± 0.14).
The following Experimental Activities 2 and 3 were more comprehensive activities and involved offline and online testing of a multimodal inference model of MWL that was tested during the Multi‐Attribute Task Battery (MATB) scenario. This involved a rigorous analysis of
Trang 17an average MAE = 0.36.
The final Experimental Activity 3 included the online validation (during two rounds) of 11 selected ANFIS models as determined in the previous activity. The results from online validation of the first five ANFIS models (containing different feature combinations of eye activity and control input features) all demonstrated good performance with a MAE around the 0.68 mark, with the best performing model showing an average MAE = 0.67 and CC = 0.71. This was similarly reflected in the results from performing cross‐session validation. The remaining multimodal models of MWL showed a larger error as the online inference from the EEG model had an arbitrary offset resulting in an equivalent offset in the output of the multimodal ANFIS model. The efficacy of the model could however be seen with the normalised pairwise correlation with the target value and showed good results, with ANFIS model 11 demonstrating the highest average correlation across the models tested (CC = 0.77). Henceforth this study has demonstrated the ability for multimodal data fusion from features extracted from EEG, eye activity and control inputs to produce an accurate and repeatable inference of MWL. The investigation of multimodal fusion for MWL inference has assisted in corroborating the viability of real‐time system adaptation in future aerospace Cyber‐Physical‐Human Systems (CPHS) architectures
Trang 18it has been noted in the literature that automation has its consequences [1‐3], as some of the main contributors to aviation accidents are attributed to human errors as a result of automation bias, complacency [4] and a deterioration of manual flying skills [5]. Among some of the cases, it has been identified that low Mental Workload (MWL) environments are a source of boredom and cognitive underload as a result of inactivity [2]. In certain periods of flight like cruise, the MWL is commonly low. This under‐load can lead to a negative effect on information processing and Situational Awareness (SA). Then if an abnormal situation appears while airborne, the pilot will quickly be exposed to cognitive over‐load and will experience a total loss of SA [2]. With these concerns regarding the integration of automation there is a need for a new interaction between human operators and machines that departs from independent automation and human activities and moves towards Cyber‐Physical‐Human Systems (CPHS).
Trang 19These human‐centric systems are the cooperation of humans with machines, and are designed such that the skills and abilities of humans are not replaced, but rather co‐exist with and assist humans in performing more efficiently and effectively [6]. Such a human centric design represents an improved design and engineering philosophy, where automation is treated as a further expansion of the humans physical, sensorial and cognitive capabilities. As such a human centric system is designed to be engineered with computational and communication techniques with adaptive control systems (i.e. adaptive automation) that maintain human‐in‐the‐loop. Some of the current Human Machine Interfaces and Interactions (HMI2) of aerospace systems are capable of integrating and fusing information from various sources to perform a variety of different functions. However, current systems and interfaces are limited in adaptability where the authority for reconfiguration and task allocation in current adaptable HMI2 are manually controlled by the human operator [7]. Henceforth, a Cognitive Human Machine System (CHMS) addresses these issues with a CPHS design that provides the necessary real‐time system adaptation [8, 9]. A CHMS provides real‐time system adaptation by collecting physiological, behavioural as well as mission, environmental and operational data (i.e. performance measures) which are then fused in real‐time to provide a final estimation of the operator’s cognitive states (i.e. mental workload, mental fatigue, attention, etc.). This estimation will then dynamically change system functions and HMI2 formats. The configuration for this can be seen in Figure 1.1 and illustrates a closed loop system as described above. The CHMS consists of four fundamental components which include the human operator, sensing module, estimation module and lastly the adaptation module. The sensors used for collecting the physiological and behavioural measurements can e.g. include data collected from an Electroencephalogram (EEG), Functional Near Infrared Spectroscopy (fNIR), eye activity tracker, cardiorespiratory sensor and control input sensors, while the performance measures can among other be deduced from the primary and/or secondary mission objectives. These measures are then fused in the estimation module to produce estimations
of the cognitive states that then drives the system adaptation. This then produces a new task load presented to the human operator, which thus results in new cognitive states of the human as well as new system conditions and the cycle then continues.
Trang 20Estimation Sensing
Actual system conditions
Actual cognitive states
⋮
Mission performance Environmental conditions Operational conditions
+
Task 1 Task n
Adaptation
HUMAN
Actual external conditions
Figure 1.1. Fundamental concept of the Cognitive Human Machine System (CHMS).
In terms of avionics used for aerospace systems there are three main aspects including Communication, Navigation and Surveillance (CNS) [10]. The operation of e.g. Air Traffic Management (ATM) is highly dependent on CNS systems to achieve the objectives of preventing collisions and ensuring a high flow of traffic [11]. The collective term for the technologies used for fulfilling many of the vital functions of aerospace systems is CNS/ATM and Avionics (CNS+A). The aforementioned considerations for a human centric design in the form of a CHMS is important as the advancements in CNS+A concepts progress [11‐13]. With the introduction of these technologies there is a need to process and present an increasing amount of information. This will in turn result in aerospace Cyber‐Physical Systems (CPS) which incorporate higher levels of automation and improved information sharing [14]. Additionally, for some of the proposed CNS+A concepts there is in fact a need for automated computation, as the function cannot be completed without computational support, such as
4‐Dimensional Trajectory Based Optimisation (4DT) where humans are not able to process
4‐dimensional space [11]. Among some of the significant advantages of the CNS+A advancements and the introduction of CPHS are the capabilities of de‐crewing of current
Trang 21to Single Pilot Operation (SPO) in commercial aircrafts [8, 15], One‐to‐Many (OTM) Unmanned Aerial Vehicle (UAV) operation [16], evolution of ATM [17] and UAS traffic management (UTM) [18]. The implementation of the CHMS in all these applications will support these aerospace systems to operate at higher levels of automation while ensuring that the human operator maintains a central role in the system and that the degree of trust with the system is maintained.
Research gaps
To reliably drive system adaptation in an operational CHMS there is a need to provide accurate and repeatable estimations of cognitive states in real‐time. Among the cognitive states, MWL is of particular importance as it directly affects the system performance [19]. However, to allow for aerospace CPHS architectures that perform dynamic real‐time system adaptation there is a need to further develop suitable models and algorithms that can infer MWL based on multiple real‐time physiological and behavioural sensor measurements. Such measurements have been extensively researched for detecting physiological and behavioural responses associated to MWL, including the use of sensors such as EEG [20‐23], fNIR [24, 25], eye activity tracking [26, 27], Electrocardiogram (ECG) [28‐30], Galvanic Skin Response (GSR) [31] and control inputs [32]. Here corresponding features associated with MWL include among others: changes in alpha and theta power‐bands with the EEG; changes
in blood oxygenation and blood volume with the fNIR; changes in Blinks Per Minute (BPM), proportion dwell time, pupil diameter and scan pattern behaviour with eye activity measures; changes in Heart Rate (HR), Heart Rate Variability (HRV) and respiration with cardiorespiratory sensors; changes in skin conductance with GSR; and lastly, changes in accuracy, response time as well as mouse movements with a computer mouse used for control inputs. Nonetheless, extensive reviews have identified that the measures of MWL are not universally valid for all task scenarios [33, 34]. The reviews both identified multiple studies that reported statistical significance in physiological and behavioural measures that are able to measure MWL for different task scenarios. However, the reviews noted that the
Trang 22measures are sensitive depending on the task type and difficulty, and that a single measure was not identified that is generalised across various tasks. The inconsistency in the sensitivity of physiological and behavioural measures can for example be seen for HR. Here studies have reported a sensitivity of HR to changes in MWL [29, 30], while others found no sensitivity [35]. With this variability in the efficacy of sensor measurements it is important
to extract measurements that are sensitive to the specific task scenario.
The task loads implemented for provoking MWL have included using many different task scenarios. Among these include using simpler task scenarios in controlled laboratory environments while performing tasks such as the n‐back task, arithmetic task, Hampshire tree task, Sternberg task and other equivalent tasks [36, 37]. Other task scenarios include the use of somewhat more complex task loads that require multi‐tasking such as Multi‐Attribute Task Battery (MATB) [38] or the automated Cabin Air Management System (aCAMS) [39]. Task scenarios that generate more complex task loads include studies that implement ATM simulations, driving simulations or flight simulation [28, 40‐45]. Lastly, only limited studies have performed physiological and behavioural measurements of MWL while performing real operational conditions, such as during actual flight or actual driving conditions [46, 47].
Studies have implemented the use of supervised Machine Learning (ML) techniques for estimating MWL during a specific task scenario by either implementing classification models for classifying MWL between discrete states, or the use of regression approaches for continuously inferring MWL. This has been used for fusing the respective features from within a modality, generally used with EEG only [48‐53] as well as fNIR only [43], or fusing features across modalities [54‐58]. Here commonly used classification approaches have included the use of Artificial Neural Networks (ANN) [59, 60], Support Vector Machines (SVM) [37, 56, 61, 62] or Linear Discriminant Analysis (LDA) [40, 43] (or extended variants), with some studies comparing multiple classification models [55, 61, 63] when classifying between discrete MWL states (i.e. resting, low and high). The use of regression models is less reported but includes the use of Neuro Fuzzy System (NFS) [16, 58, 64] and Gaussian Process Regression (GPR) [52], which provides a continuous estimation of MWL.
Trang 23The type of task scenario, and task loads presented to the subject can differ quite substantially between the various studies. Moreover, many of the studies performing multimodal data fusion include ML techniques that perform the calibration and validation
in offline processing. Nonetheless, real‐time/ online data processing often shows variable and inconsistent results. The following studies by Hogervorst et al. and Wilson and Russell have performed online validation of multimodal data fusion models [57, 60]. In the study conducted by Wilson and Russell [60], an ANN was used to fuse data from an EEG (band‐power from delta, theta, alpha, beta and gamma from six electrodes), eye activity measures (blinks and interblink intervals) and cardiorespiratory measures (respiration, HR and HRV) in
a MATB task scenario using resting, low and high task load conditions. Here the ANN would perform online classification of the respective task load at 5 second intervals and demonstrated a classification accuracy of 84.3%. In the other study by Hogervorst et al. [57], the online validation was simulated after collecting the data. Here data from EEG, GSR, cardiorespiratory (respiration, HR, HRV) and eye activity measures (pupil size and eye blink) were collected from the participant performing an n‐back task, between high and low task loads. A classification accuracy of 91% was achieved from using 2‐minute intervals and using
an elastic net calibrated on features from an EEG and eye activity measures. However, it was concluded that fusing the measure did not notably improve the results. Both results showed
a good accuracy, however, for both studies a classification model was used which only discriminates between discrete classes (i.e. low and high). Moreover, the latter study does not perform a true real‐time processing, as it only implemented a simulated online validation, and it additionally used a relatively large interval of 2 minutes.
Some studies have in fact performed elements of real‐time system adaptation by using measures of MWL [24, 40, 64, 65]. However, generally these studies have used models that implement one modality, such as only EEG [40], only fNIR [24] or only task performance [65]. The study by Ting et al. [64], is one of the few studies that performed real‐time system adaptation by implementing a multimodal data fusion model. Here EEG features, HRV and
a task performance measure was used to drive system adaptation in an aCAMS simulation, although the physiological measures were taken at quite large intervals at 7.5 minutes. However, it was concluded that the physiological and performance measures improved the system adaptation over using only system error for driving the adaptation.
Trang 24of particular interest as it is well suited for fusing data from multiple modalities and is more transparent than other ML methods thus overcoming some of the “black box” problem that are faced by many of the other ML methods. An NFS can optimise the parameters of a Fuzzy Inference System (FIS) based on calibration data, and one notable method is the Adaptive Neuro Fuzzy Inference System (ANFIS) [66]. There have been limited uses of NFS for MWL estimation, however the studies that have used NFS include the following references [16,
58, 64, 67]. The methods used to optimise the FIS parameters include the use of a Genetic Algorithm (GA) based Mamdani fuzzy model [64, 67], where 7.5 minute intervals were used for the EEG and HRV measures. An extension of that work was presented by Wang et al. [58], with a 2 minute interval and using Differential Evolution (DE) and Differential Evolution Algorithm with Ant Colony Search (DEACS), as means to optimise the ANFIS parameters. The preceding studies demonstrated good results in an offline calibration and validation with the respective models, but they also implemented a general ANFIS model for comparison that showed poor results. With large intervals used, this could result in a limited amount of calibration data for the models, particularly for the general ANFIS model. A study by Lim et
al. also implemented an ANFIS and was calibrated on data from cardiac features, eye activity features and features from an fNIR [16]. The offline validation showed good results from calibrating the model on the normalised features but failed to demonstrate good results in the online validation. These aforementioned studies thus lack in using a general ANFIS paradigm to produce accurate results in inferring MWL. Some of the studies demonstrated quite large intervals used, while the latter study lacked to perform an inference of MWL in
an online validation. Moreover, other more recent studies have outlined the importance of investigating the respective features contributing to the performance of the respective models used [55, 63]. This is also an area that has not been investigated for NFS in the previous studies and is an area that can assist in the explainability of the model.
As mentioned before recent extensive reviews have identified that there is not a current measure of MWL that is universal for all task scenario [33, 34]. While many of the ML classification/ inference models have demonstrated the ability of increasing the accuracy of the estimation of MWL, there are still significant limitations in ML models’ ability for cross‐
Trang 25task, cross‐level, cross‐session and cross‐subject classification/ inference [48‐50, 62, 68]. For example, the study conducted by Zhao et al. demonstrated a within‐classification accuracy of 95.3%, while the online validation of cross‐task classification dropped to 53.8% (with 33.33% as random), and cross‐level had an accuracy of 72.2% [62]. Henceforth, the capability of accurate cross‐task and at minimum cross‐level is vital for the estimation of MWL to be reliable enough for a CHMS to be used in operational aerospace systems. Moreover, for driving real‐time system adaptation there is a need for defining reference values that serve as thresholds for executing the system adaptation. Defining such reference values is nevertheless a challenging area given that there is no established ground truth for what is considered overload and underload of MWL. In addition, the aforementioned limitations of current classification/ inference models (i.e. cross‐task and cross‐session inference) and the human operators subjective experience of a given task load also provides challenges for defining accurate and reliable threshold values. Henceforth, before proceeding with the further implementation of real‐time system adaptation and the effects
of this on aerospace systems and the human operator, there is a need to thoroughly study methods for sensing physiological, behavioural and performance measures associated with MWL. This includes further researching different methods for devising multimodal data
fusion of MWL that can accurately and repeatably infer MWL in real‐time.
1.2.1 Scope
To clearly define the scope of this research the following aspects were considered. Firstly, regarding the CHMS, this project was primarily focused on the sensing and estimation module of the system. This meant that the adaptation module, and its integration for various aerospace systems was kept outside of the scope. This additionally meant that defining threshold values for driving system adaptation was not included in the scope. Nonetheless, the consideration of how the estimation of MWL would affect the adaptation module was briefly considered as part of the review and would also determine requirements for the sensing and estimation modules. As such, the scope of this research was kept on the various physiological, behavioural and performance measures, including the applicable sensors used, the extraction of features and the multimodal fusion to extract one final accurate and repeatable measure.
Trang 26In terms of cognitive states, this research was limited to the study of MWL, although this is arguably also interchangeably named in the literature as cognitive load, mental load, Operator Functional State (OFS) and cognitive workload. The theoretical background for some of the other cognitive states would however be briefly included in the review and involved fatigue, SA, working memory and attention.
The sensor suite used for experimental testing and analysis would be limited to an EEG, ECG, eye activity tracker and computer mouse. This included both the offline processing as well
as real‐time processing of the data from the respective sensors. Other sensors would however be included as part of the scope for review. Elements of artifact rejection was included as part of the review and the experimental activities would include some methods
of artifact rejections as allowed in real‐time processing.
The scope was limited to the testing of a regression type ML model for real‐time multimodal data fusion. Other methods of data fusion, such as other ML classification techniques, was included as part of the scope for review and discussion. As part of the analysis from the experimental testing, some elements of explainability of the inference of MWL from the respective multimodal model was included as part of the scope. Lastly, cross‐task, cross‐level and cross‐subject capability of ML models would not be considered as part of the scope, apart from review and for comparison in the discussion. Elements of cross‐session inference was however included as part of the scope and was experimentally tested.
Research aim and question
The primary aim of this research was to further develop the sensing and estimation module
of the CHMS framework with an improved multimodal inference model of MWL as needed for the application in aerospace CPHS. Here the focus of this research was on the implementation of a multimodal inference model of MWL that would be developed and experimentally tested. This was done with the purpose of studying the viability of real‐time system adaptation in aerospace systems. As such the main aim of this research was:
Trang 27OBJ. 5 Perform a thorough analysis from offline calibration and validation of a
multimodal inference model of MWL and analyse performance from using different generic and subject specific feature combinations;
Trang 28Overview of research methodology
This section outlines an overview of the overall methodology and the approach taken for the experimental activities conducted, which is further elaborated on in Chapter 3. The first part of this research is a comprehensive literature review conducted with the focus on the applicable body of work. The second part of this research includes experimentally testing and studying the functionalities of a multimodal sensing and estimation module as needed for a CHMS. The general approach for this includes three experimental activities and are outlined in the following sub sections.
1.5.1 General project methodology
This research aims to develop an accurate and repeatable multimodal inference model for estimating MWL as needed for a CHMS system. In working towards this aim, the following general project methodology was implemented as part of this research project. The full general project methodology can be seen in Figure 1.2 and elements of this was used for all three experimental activities completed. The full approach consisted of firstly using a task scenario to controllably generate various task loads that were presented to the human subject. The human then performed the presented task load, which in turn resulted in an
“actual MWL” perceived by the human. The “observe” block was then where various wearable and non‐wearable sensors measured various physiological and behavioural responses that originated from the human performing the given task. This included the use
of an EEG, eye activity tracker, ECG and monitoring control inputs to measure the raw data and then further extract features such as HR from the ECG signal. Then in the “orient” module the various extracted features were collected, and the physiological and behavioural features were oriented using the Pearson Correlation Coefficient (CC), to assess the pairwise linear relationship with other objective measures of MWL. This could either be the correlation between the feature and a pre‐determined task level, a pairwise correlation with a secondary task performance measure, or reserving another feature for comparison. Subsequently, in the “decide” module an assessment could be made that decided which physiological/ behavioural feature were best suited for MWL estimation based on the pairwise correlation results. This assessment could finally be used for the last module, where
Trang 29
Figure 1.2. The general project methodology.
1.5.2 Experimental activities
To prove the capabilities of a sensing and estimation module’s ability to infer MWL in real‐time, three experimental activities were conducted using various steps of the general project methodology presented above. Firstly, this included an exploratory experimental activity that assessed some of the physiological, behavioural and performance measurements associated with MWL during a complex OTM UAV wildfire detection scenario. The following two experimental activities were more comprehensive and included the offline and online calibration and validation of a multimodal data fusion model developed for real‐time inference of MWL. These experimental activities (Experimental Activity 1, 2 and 3) were conducted to complete the research objectives and answer the research question associated to this research project.
In the first exploratory Experimental Activity 1 the use of physiological, behavioural and performance measures of MWL were implemented and analysed in a complex OTM UAV wildfire detection scenario. Overlapping with the general project methodology presented above, this experimental activity only implemented the observe and orient sections as further seen in Figure 1.3. In this activity the task load presented to the human operator was generated from a wildfire detection scenario, where the task complexity was controlled in three incremental levels. The measures collected as a part of this activity included physiological, behavioural, performance and subjective measures. This exploratory activity was conducted in collaboration with another project [16], and expanded on the work by further developing an EEG index and implementing a data fusion approach with the EEG
Trang 30index and an eye activity measure. The EEG index and data fusion approach were analysed using the Analysis of Variance (ANOVA) analysis and the CC, and additionally included analysing the other features for comparison. The pairwise correlation was conducted with
an objective secondary task performance index for assessing how sensitive the features were to the complex OTM UAV task scenario.
Figure 1.3 General methodology for Experimental Activity 1
In the following Experimental Activity 2, the emphasis was on the inference of MWL using a supervised ML model for multimodal data fusion. This activity included experimentally collecting physiological and behavioural data, and then performing a data fusion of these multimodal features in an offline calibration and validation procedure. Here multiple models were tested that contained different feature combinations. Compared with the general methodology, the full methodology was implemented with the observation of the features, the orientation of the features, decision of which features to use and lastly the implementation of those features for offline calibration and validation of the developed and tested inference models.
Lastly, Experimental Activity 3 investigated the online validation of the inference models examined in the previous experimental activity. While the previous activity investigated the offline calibration and validation of multimodal inference models, this activity took the optimal models from the previous experimental activity and applied it in an online validation. This would verify the inference model’s capability for real‐time estimation of MWL. This used the same approach as in Experimental Activity 2 with the full implementation of the general approach.
Trang 31Thesis outline
This section will briefly outline the remaining chapters of this thesis and its contents. Chapter 2 of the thesis includes a literature review of the relevant literature for this research. This includes a detailed review on the body of work within MWL measurements, EEG, ECG, eye activity tracking, control inputs, human factors, adaptive automation, closed loop system adaptation, pBCI, HMI2, CNS+A, multimodal data fusion, machine learning and ANFIS.
Chapter 3 will then outline some of the design requirements and functionalities for a sensing and estimation module. This will guide the design of the experiments as these dictates what
is needed for the sensing and estimation module of a CHMS. This chapter will additionally include outlining the methodology as part of this research and will detail the hardware and software infrastructure used.
In Chapter 4, 5 and 6 the work from Experimental Activities 1, 2 and 3 will be presented, where all three activities involve humans. These chapters will also present the methods implemented for the respective activity as well as the results from each experiment and lastly a discussion. Chapter 4 is an exploratory experimental activity that presents the results from the development and testing of an EEG index, and the fusion between the EEG index and an eye activity measure in a complex OTM UAV wildfire detection scenario. Chapter 5 and 6 are closely related, where Chapter 5 will include the presentation of the development and experimental results from offline calibration and validation of various inference models using EEG, eye activity tracking, cardiac and control input measures of MWL. While Chapter
6 is then the online validation of the previously analysed models, using an optimal calibration protocol. Lastly Chapter 7 provides a synthesis of the discussions from the respective experimental activities, a conclusion and recommendations for future research.
Trang 322006
[8] J. Liu, A. Gardi, S. Ramasamy, Y. Lim, and R. Sabatini, "Cognitive pilot‐aircraft interface for single‐pilot operations," Knowledge‐Based Systems, vol. 112, pp. 37‐53, November 2016.
[12] T. Kistan, A. Gardi, R. Sabatini, S. Ramasamy, and E. Batuwangala, "An evolutionary outlook of air traffic flow management techniques," Progress in Aerospace Sciences, vol. 88, pp. 15‐42, October 2016.
[13] E. Batuwangala, S. Ramasamy, L. Bogoda, and R. Sabatini, "An interoperability assessment model for CNS/ATM systems," in ATRF, Melbourne, Australia, 2016, pp. 1‐8.
[14] Y. Lim et al., "Avionics Human‐Machine Interfaces and Interactions for Manned and Unmanned Aircraft,"
Progress in Aerospace Sciences, vol. 102, pp. 1‐46, August 2018.
[15] Y. Lim, V. Bassien‐Capsa, S. Ramasamy, J. Liu, and R. Sabatini, "Commercial airline single‐pilot operations: System design and pathways to certification," IEEE Aerospace and Electronic Systems Magazine, vol. 32,
no 7, pp. 4‐21, July 2017.
[16] Y. Lim et al., "Adaptive Human‐Robot Interactions for Multiple Unmanned Aerial Vehicles," Robotics, vol.
10, p. 12, January 2021.
Trang 33[17] A. Gardi, N. Pongsakornsathien, L. Planke, Y. Lim, T. Kistan, and R. Sabatini, "Development of a Cognitive HMI for Air Traffic Management Systems – Report No. 4: CHMI2 System‐Level Implementation and Verification in a Representative Simulation Environment," Bundoora, VIC, Australia, RMIT/SENG/CPS/003 ‐2019, 2019.
[18] N. Pongsakornsathien, A. Gardi, R. Sabatini, T. Kistan, and N. Ezer, "Human‐Machine Interactions in Very‐ Low‐Level UAS Operations and Traffic Management," in IEEE/AIAA 39th Digital Avionics Systems
Conference, DASC2020, Piscataway, NJ, USA, 2020.
[19] M. S. Young, K. A. Brookhuis, C. D. Wickens, and P. A. Hancock, "State of science: mental workload in ergonomics," Ergonomics, vol. 58, no. 1, pp. 1‐17, 2015.
[20] S. Puma, N. Matton, P.‐V. Paubel, É. Raufaste, and R. El‐Yagoubi, "Using theta and alpha band power to assess cognitive workload in multitasking environments," International Journal of Psychophysiology, vol.
123, pp. 111‐120, October 2018.
[21] A. Gevins et al., "Monitoring Working Memory Load during Computer‐Based Tasks with EEG Pattern
Recognition Methods," Human Factors, vol. 40, no. 1, pp. 79‐91, March 1998.
[22] W. Klimesch, "EEG alpha and theta oscillations reflect cognitive and memory performance: a review and analysis," Brain Research Reviews, vol. 29, no. 2‐3, pp. 169‐195, November 1999.
[27] J. Gilland, "Driving, eye‐tracking and visual entropy: Exploration of age and task effects," PhD, Department of Psychology, University of South Dakota, 2008.
[28] H. Mansikka, P. Simola, K. Virtanen, D. Harris, and L. Oksama, "Fighter pilots’ heart rate, heart rate variation and performance during instrument approaches," Ergonomics, vol. 59, no. 10, pp. 1344‐1352,
March 2016.
[29] A. M. Hughes, G. M. Hancock, S. L. Marlow, K. Stowers, and E. Salas, "Cardiac Measures of Cognitive Workload: A Meta‐Analysis," Human Factors, vol. 61, no. 3, pp. 393‐414, May 2019.
[30] T. Heine, G. Lenis, P. Reichensperger, T. Beran, O. Doessel, and B. Deml, "Electrocardiographic features for the measurement of drivers’ mental workload," Applied Ergonomics, vol. 61, pp. 31‐43, January 2017.
[31] Y. Shi, N. Ruiz, R. Taib, E. Choi, and F. Chen, "Galvanic skin response (GSR) as an index of cognitive load,"
in CHI Conference on Human Factors in Computing Systems, United States, 2007, pp. 2651‐2656.
[32] A. Khawaji, F. Chen, J. Zhou, and N. Marcus, "Trust and cognitive load in the text‐chat environment," in
26th Australian Computer‐Human Interaction Conference, 2014, pp. 324–327.
Trang 341070 ‐1077, March 2014.
[36] G. N. Dimitrakopoulos, Y. Sun, K. Ardian, N. V. Thakor, and A. Bezerianos, "A method for cross‐task mental workload classification based on brain connectivity," Frontiers in Human Neuroscience, vol. 10, pp. 1940‐
1949, November 2017.
[37] T. Radüntz, "Dual Frequency Head Maps: A New Method for Indexing Mental Workload Continuously during Execution of Cognitive Tasks," Frontiers in Physiology, vol. 8, p. 1019, December 2017.
[38] J. R. Comstock Jr and R. J. Arnegard, "The multi‐attribute task battery for human operator workload and strategic behavior research," Langley Research Center, Hampton, Virginia, 1992.
[39] B. Lorenz, F. Di Nocera, S. Röttger, and R. Parasuraman, "Automated fault‐management in a simulated spaceflight micro‐world," Aviation, Space and Environmental Medicine, vol. 73, no. 9, pp. 886‐897,
April 2020.
[42] Y. Lu, J. Wu, and S. Fu, "A Study of the Relationship between Novice Pilots’ Performance and Multi‐ Physiology Signals," in Engineering Psychology and Cognitive Ergonomics, Crete, Greece, 2014, pp. 344‐
Trang 35[49] R. Hefron, B. Borghetti, C. S. Kabban, J. Christensen, and J. Estepp, "Cross‐Participant EEG‐Based Assessment of Cognitive Workload Using Multi‐Path Convolutional Recurrent Neural Networks,"
Sensors, vol. 18, no. 5, p. 1339, April 2018.
[50] Z. Yin and J. Zhang, "Cross‐session classification of mental workload levels using EEG and an adaptive deep learning model," Biomedical Signal Processing and Control, vol. 33, pp. 30‐47, November 2017.
[51] Z. Yin and J. Zhang, "Cross‐subject recognition of operator functional states via EEG and switching deep belief networks with adaptive weights," Neurocomputing, vol. 260, pp. 349‐366, May 2017.
[52] M. S. Caywood, D. M. Roberts, J. B. Colombe, H. S. Greenwald, and M. Z. Weiland, "Gaussian Process Regression for Predictive But Interpretable Machine Learning Models: An Example of Predicting Mental Workload across Tasks," Frontiers in Human Neuroscience, vol. 10, p. 647, January 2017.
[53] P. Zhang, X. Wang, W. Zhang, and J. Chen, "Learning Spatial–Spectral–Temporal EEG Features With Recurrent 3D Convolutional Neural Networks for Cross‐Task Mental Workload Assessment," IEEE
Transactions on Neural Systems and Rehabilitation Engineering, vol. 27, no. 1, pp. 31‐42, January 2019.
[54] Y. Liu, H. Ayaz, and P. A. Shewokis, "Multisubject “Learning” for Mental Workload Classification Using Concurrent EEG, fNIRS, and Physiological Measures," Frontiers in Human Neuroscience, vol. 11, p. 389,
July 2017.
[55] Y. Ding, Y. Cao, V. G. Duffy, Y. Wang, and X. Zhang, "Measurement and identification of mental workload during simulated computer tasks with multimodal methods and machine learning," Ergonomics, vol. 63,
no 7, pp. 896‐908, April 2020.
[56] J. Zhang, Z. Yin, and R. Wang, "Recognition of Mental Workload Levels Under Complex Human Machine Collaboration by Using Physiological Features and Adaptive Support Vector Machines," IEEE Transactions
Control, vol. 7, no. 5, pp. 490‐498, September 2012.
[59] A. Craik, Y. He, and J. L. Contreras‐Vidal, "Deep learning for electroencephalogram (EEG) classification tasks: a review," Journal of Neural Engineering, vol. 16, no. 3, p. 031001, April 2019.
[60] G. F. Wilson and C. A. Russell, "Real‐Time Assessment of Mental Workload Using Psychophysiological Measures and Artificial Neural Networks," Human Factors, vol. 45, no. 4, pp. 635‐643, 2003.
[61] T. K. K. Ho, J. Gwak, C. M. Park, and J. I. Song, "Discrimination of Mental Workload Levels from Multi‐ Channel fNIRS Using Deep Leaning‐Based Approaches," IEEE Access, vol. 7, pp. 24392‐24403, January
Trang 36[64] C.‐H. Ting et al., "Real‐Time Adaptive Automation System Based on Identification of Operator Functional
State in Simulated Process Control Operations," IEEE Transactions on Systems, Man, and Cybernetics,
vol 40, no. 2, pp. 251‐262, March 2010.
[65] D. B. Kaber, M. C. Wright, L. J. Prinzel, and M. P. Clamann, "Adaptive Automation of Human‐Machine System Information‐Processing Functions," Human Factors: The Journal of the Human Factors and
Ergonomics Society, vol. 47, no. 4, pp. 730‐741, 2005.
[66] J. S. R. Jang, "ANFIS: adaptive‐network‐based fuzzy inference system," IEEE Transactions on Systems,
Man, and Cybernetics, vol. 23, no. 3, pp. 665‐685, June 1993.
[67] J.‐H. Zhang, X.‐Y. Wang, M. Mahfouf, and D. A. Linkens, "Fuzzy Logic based Identification of Operator Functional States Using Multiple Physiological and Performance Measures," in International Conference
on BioMedical Engineering and Informatics, Sanya, 2008, pp. 570‐574.
[68] Y. Ke et al., "Towards an effective cross‐task mental workload recognition model using
electroencephalography based on feature selection and support vector machine regression,"
Trang 37This page is intentionally left blank to support presswork
Trang 38of areas but is predominantly centred around MWL, its theoretical background as well as means of sensing, estimating and analysing real‐time measurements during different task load conditions. Thus, the literature search is conducted using MWL and its variants such as cognitive load, cognitive workload and operator functional state. Here its theoretical background is reviewed in relation to human factors and its relationship with other cognitive states such as situational awareness, working memory, attention and fatigue. The review also considers an overview of methods for sensing and estimating MWL using various measurement methods for real‐time estimation. This includes reviewing physiological and behavioural sensors such as EEG, ECG, eye activity tracking and control input devices, where the sensor technology is outlined, and the features extracted are considered.
As the aim of measuring MWL is driving real‐time system adaptation for aerospace operations, a review of human centric systems such as adaptive automation/ adaptive systems/ closed loop system adaptation/ CHMS/ pBCI is considered as well as other aerospace considerations such as the HMI2 and CNS+A. This is further reviewed for its
Trang 39of ML, with a further emphasis on the ANFIS. Generalised inference models with the capability for cross‐task, cross‐subject and cross‐session will also be briefly reviewed. The databases used to conduct the literature search include Google Scholar, Scopus, Web of Science, RMIT Library database, Science Direct and JSTOR.
Human centred systems
The dynamic complex systems in modern work is becoming problematic when looking at for example the human response to automation. A number of complex tasks would not be achievable without assistance from automation. As an example, within the current cockpit
of commercial aircraft, automation has a vital role for maximising the safety, efficiency and sustainability of an aircraft. Moreover, some of the most advanced automated flight decks
in commercial aircraft can function with high levels of automation, from gear up following take‐off and all the way to the landing performed by category III instrument landing system approach. During flight the human operator has the role of maintaining complete oversight
of the systems, with the capability to override and operate manually at any given point in the unlikely event of a failure. Within the flight deck the tasks that are automated include flight director, auto throttle, autopilot, flight management system as well as centralised warning and alerting systems [1]. Nonetheless, it has been noted in the literature that the automation has its consequences [1‐3], where some of the main contributors to aviation accidents are attributed to human errors as a result of automation bias, complacency [4] and a deterioration of manual flying skills [5]. As mentioned above, it has been identified that low workload environments are a source of boredom and mental underload as a result
of inactivity [3]. In certain periods of flight, like cruise, the workload is commonly low, where the pilots only undertake monitoring roles that include observing the performance parameters of the systems, navigation as well as Air Traffic Controller (ATC) reporting. This under‐load can lead to a negative effect on information processing and SA. In an event were the automation does not function as it should and an abnormal situation appears while
Trang 40airborne, the pilot will quickly be exposed to mental overload and will experience a total loss of SA [3]. However, when operations are under normal automation conditions, human performance is improved, the SA is increased and MWL is significantly reduced. On the other side, if the there is an event where the automation fails, the performance can be seen to decrease, despite the capabilities of the machine or the human [6]. One of the factors of this is out of the loop syndrome where the operator loses mode awareness of the system. The technological developments across industries has introduced what some refer to as an Operator 4.0 [7]. This new interaction between operators and machines departs from independent automation and human activities and evolves more towards CPHS systems. These human‐centric systems are the cooperation of humans with machines, and are designed such that the skills and abilities of the human are not replaced, but rather they co‐exist with and assist humans in performing more efficiently and effectively [7]. This Operator 4.0 represents an improved design and engineering philosophy, where automation is treated as a further expansion of the humans physical, sensorial and cognitive capabilities.
An important aspect of this is the implementation of neuroergonomics, which is the study
of the brains function to work. Here it takes the best from subjective, behaviour and physiological measurements to get an overall view of mental capacity [8]. As such, a human centric system is designed to be engineered with computational and communication techniques with adaptive control systems (adaptive automation) that maintain human‐in‐the‐loop. Henceforth, CPHS systems, as identified in the literature, aim to:
1 improve human abilities to dynamically interact with machines in the cyber‐ and physical‐ worlds by means of ‘intelligent’ human‐machine interfaces, using human‐computer interaction techniques designed to fit the operator;
2 improve human physical‐, sensing‐ and cognitive capabilities, by means of various enriched and enhanced technologies (e.g. using wearable devices) [7].
Traditionally, function allocation in systems has been categorised into comparison allocation, such as MABA‐MABA (Machines Are Best At – Men Are Best At), leftover allocation and economic allocation [9]. However, humans and machines are complementary and while the traditional approach of function allocation considers “who does what”, the successful implementation needs to consider “who does what and when”. As such there is