Sage Journals: Discover world-class research

Abstract

Electroencephalogram is the electrical phenomena in the cerebral cortex or the scalp surface due to the electrophysiological activity of brain cells. Electroencephalogram has great theoretical and practical significance in measuring mental workload of people. More precise electroencephalographic is a precondition to study mental workload of miners. In this article, based on the actual situation of the electroencephalographic measurement of miners, the particle swarm optimization is introduced to improve the standard genetic algorithm, and put forward a combined method integrating the genetic algorithm with particle swarm optimization for achieving electroencephalogram-based measures of miners' mental workload. Moreover, the MATLAB simulation platform is used for simulation testing. Testing results prove the effectiveness of the combined method.

Keywords

Genetic algorithm particle swarm optimization improved genetic algorithm mental workload electroencephalogram

Introduction

Mental workload is also known as mental load or stress. Young et al.¹ believed that mental workload is a widely used concept, as well as one of important research objects in ergonomics. Moreover, mental workload is also associated with many measurement concepts. Liao et al.² defined mental workload as “an indicator for measuring the information processing system of people and the extent to which the information process system is occupied in information processing, that is, being inversely proportional to the unused information processing capacity of people.”

The coal industry is a pillar industry for economic development in China. Along with the application of large-scale and complex technology in coal production, the complexity and intelligence level of the man–machine system in coal production has been constantly improved. In the large-scale coal production, coal miners have experienced significant changes in the quality and quantity of their mental workload. Mental workload problems arising from this have become increasingly distinct. Therefore, it is quite essential to monitor and investigate mental workload of miners at intellectual work.

In research on mental workload, electroencephalogram (EEG) is usually used as an important monitoring indicator. Giraudet et al.³ investigated the task of using alerts for monitoring under high mental workload. EEG measurement results showed that the P300 event-related potential declined sharply. According to some studies, EEG power in the frequency bands, such as theta, alpha, and beta and so on are sensitive to changes in mental workload. It is generally believed that the psychological pressure, active thinking, and attention can drive EEG activities to move toward the higher-frequency section and suppress activities of alpha wave.⁴ The energy of the alpha wave and theta wave is negatively correlated to the task difficulty degree and mental workload.^5,6

Through research, Käthner et al.⁷ confirmed that the combination of signals from the time and frequency domains of EEG can be used a valuable method for online testing of workload and fatigue of workers. Cinaz et al.⁸ attempted to use HRV to measure mental workload in the daily working environment, and used the linear analysis method to classify time and frequency-domain features of HRV, to achieve some effects. Under the task of n-back (n = 1, 2, 3), Herff et al.⁹ collected the prefrontal cortex fNIRS signals of subjects and made a linear discrimination analysis for classifications according to change rates of Hb0 and HbR, to get three classifications whose highest accuracy rate was 78%.

Traditionally, the SGA is mainly used to select EEG features. The GA (genetic algorithm) is a global probabilistic research algorithm based on natural selection, genetic mutation, and other biological evolution mechanisms. It can maintain a good balance between the depth and breadth of search space.¹⁰ However, SGA is not quite effective in many cases, because it would easily cause early maturing, poor local search abilities, and other problems.

In this study, on the basis of the actual situation of the EEG measurement of miners, the particle swarm optimization (PSO) is introduced to improve the standard genetic algorithm (SCA), and put forward a combined method integrating the GA with PSO to make EEG-based measures of miners' mental workload. Moreover, the MATLAB simulation platform is used for simulation testing. Testing results prove the effectiveness of the combined method.

EEG-based mental workload evaluations of miners can effectively prevent working pressure of miners, reduce behavioral mistakes of miners, reduce and eliminate unsafe behaviors of minders, guard against the occurrence of accidents in coal mine enterprises, and lower the human factor-provoking accident rate.

A series of laboratory experiments carried out by the US Air Force also showed that the EEG activity could be one of sensitive indicators reflecting the mental workload level in pilots and could be used for grading evaluation of mental workload in pilots.^11,12 Käthner et al.⁷ conducted targeted experiments under middle and high mental workload to reveal effects of mental workload and fatigue. Experimental results proved that the combination of signals from the time and frequency domains of EEG can be used a valuable method for online testing of workload and fatigue of workers. Feng used the combination algorithm to launch comparative experiments of Movielens data sets. According to experimental results, compared with the pure collaborative filtering recommendation, a combined GA integrating clustering with collaborative filtering recommendation can achieve more desired results.¹³ To deal with the particle diversity degradation and improve the target tracking accuracy of the wireless sensor network (WSN), Dong and Guo-Wei¹⁴ put forward a target tracking method based on particle filtering, which is optimized by the quantum GA. Kai et al.¹⁵ proposed the GA-BPN algorithm, as a new image filtering algorithm. According to the GA-BPN algorithm, the GA is used to allocate weight in a BPN (back propagation neural network). Compared with traditional optimization algorithms, the GA-BPN algorithm has better performance in global optimization.

Cui et al.¹⁶ used the neural network-based GA to design a wide-band and multimode-based bandpass filter with non-equiripple response, and discussed the accuracy of the optimization method proposed. Merabti and Massicotte¹⁷ designed genetic operators to enhance processing performance and robustness of quantization effects and make it impossible to implement the fixed-point and low bit-wordlength algorithm, thereby saving hardware cost.

In view of the continuity of coal production and high-precision environment required by various EEG measuring equipment and based on comparisons of various methods, this study intends to employ the EEG monitoring physiological measurement method to extract data about mental workload of miners. Specific steps are as follows.

Step 1: A continuous performance test is made of miners, and N-back tasks are divided into two parts: the memory part and the judgment part. The memory part is placed before the CPT task, whereas the judgment part is placed after the CPT task. N-back tasks are taken as a noise variable increasing mental workload to act on miners tested.

Step 2: Fifty-eight channel EEG data are selected from data collected by the Neurone 64-conduction EEG system to analyze EEG characteristics of miners under mental workload.

Step 3: The improved GA-based EEG processing method is used for filtering EEG data collected in Step 2.

Step 4: Filtered EEG data are used to analyze mental workload of miners.

Algorithms

GA

A GA is a random search optimization algorithm, which learns from the theory of natural selection and genetic mechanisms. It is a random global optimization algorithm, which was put forward by Holland in the 1970s.¹⁸ This algorithm is used to generate approximate optimal solutions to objective functions under a particular model, in the case that basic information of EEG is known. The GA is a random search method, which mimics the organic evolution system. This algorithm touches upon following basic concepts and terms in biology.

Chromosome: representation of individuals.

Genetic factors: different characteristics of individuals.

Individual: basic objects and structures, which are processed, that is, solutions.

Population: a collection of individuals.

Fitness: the extent to which individuals adapt to their living environment or the viability of individuals under environment and competition pressures, which is determined by genetic factors.

Selection: exclusive competition in the case of limited resources.

Crossover: interchange of genetic materials between homologous chromosomes.

Mutation: changes in certain genes of an individual.

With the GA, you can know nothing about problems to be solved, but just need to evaluate each chromosome generated by the GA, select according to the fitness value, so that chromosomes with good adaptation have more reproduction opportunities than chromosomes with poor adaptation, and reach convergence after repeated iterations. The GA is especially suitable for processing complex nonlinear problems, which traditional search methods are difficult to solve. At present, the GA has become a commonly used optimization method.

The GA is a process to begin with its iteration searching for a randomly generated population, successively update the superiority of solution to the population, and eventually generate the optimal solution. Supposing that P (t) and C (t) are, respectively, the pair of “parent” solution and the “child” solution of the tth generation, the general structure of the GA is as follows.

Begin t = 0;

Initialization P (t);

Evaluation P (t);

Begin

Recombine P (t), to obtain C (t);

Evaluation C (t);

Choose from P (t) and C (t)

t = t+1;

End

He¹⁹ put forward a channel selection method to solve these problems. An improved GA, which is based on the Rayleigh coefficient feature, is used to identify the optimal subset of channels. According to experimental results of two motor imagery EEG datasets, it is confirmed that our method is effective in the channel selection for the classification of motor imagery EEG signals.

According to Rejer and Lorenz,²⁰ one feature selection approach is often used in research on brain–computer interface. This approach is a GA, which can code all features of an individual. Rejer figured out that although this approach can help to obtain features with high classification precession, it would also lead to a highly redundant set of features. Karakis et al.²¹ used a GA-based ANN model to identify sources of brain activities, on the basis of an overview of recent patents and literature. Preliminary results showed that the GA-based ANN model proposed in his research could be used to find sources of the equivalent current dipoles. Magee and Givigi²² used a GA combined with both linear discrimination analysis classifiers and a neural network to improve the single-trial P300 detection. The GA perfected feature selection to train the classifiers. Moreover, Magee also investigated results of features, which are found to have influence on P300 classification, and suggested the future development direction for single-trial P300 detection research.

As can be seen from above analyses, the GA has been widely used in the EEG signal analysis. However, in the development process, it is not quite effective to purely use the SGA in the EEG signal analysis. The SGA would easily cause pre-maturing and show poor performance local optimization searching. The SGA usually relies on experience to determine the crossover and mutation probability, which would affect the global optimality and convergence.

PSO

PSO is also known as the particle swarm algorithm. It is a group collaboration-based random search algorithm, which is developed by simulating foraging behaviors of a bird flock. The PSO is generally considered to be a kind of swarm intelligence (SI), so it can be incorporated into the Multiagent Optimization System (MAOS). The PSO was first put forward by Dr. Eberhart and Dr. Kennedy.

PSO simulates food-finding behaviors of bird flocks. A flock of birds randomly search food in the search space with only one piece of food. All the birds do not know where the food is, but they know how far the food is away from their current locations. What is the best strategy to find the food? The most simple and effective strategy is to search surrounding areas of the bird which is nearest to the food. PSO is developed by learning from this model and used to solve optimization problems. In the PSO, the solution to each optimization problem is a bird in the search space. The solution is called “particle.” All particles get their respective optimized functions which can determine their fitness values. Each particle has a velocity determining its direction and distance of flying. Particles would follow the current optimal particles to search in the solution space. Particles get their candidate solutions through PSO initialization and find the optimal solution through iterations. In each iteration, particles follow the “two extreme values” to update their position. One is its own best known position, which is called “pBest.” The other is the best known position of the entire swarm, which is called “gBest.” Alternatively, the best known position of an entire swarm can be replaced by the best known position of a sub-swarm around. In this case, the extreme value of all particles in this sub-swarm is the best known position of this sub-swarm. PSO can effectively optimize various functions. In the past few decades, PSO has been largely developed, applied, and popularized in many fields.

In his article, Krzeszowski et al.²³ put forward that a particle filter embedded with PSO could be used as an effective method to deal with 3D model-based tracking of human body. Su et al.²⁴ took advantage of the improved PSO to adjust gains of PID controller, iterative learning controller and the zero-phase Butterworth filter bandwidth of ILC. According to simulation results, the controller can remove errors along with repetition. Zhou et al.²⁵ introduced a novel strategy with polynomial mutation and variable random functions into the PSO, that is, the PSO-RM, which can fulfill variable random functions and mutation. Experimental results proved the effectiveness of this strategy.

In the continuous space coordinate system, the mathematical description of the PSO is as follows. Let N be the particle swarm size. The coordinate position vector of each particle in D-dimensional space is ${X_{i}}^{\to} = (x_{i 1}, x_{i 2}, \dots, x_{id}, \dots, x_{iD})$ ; the speed vector is ${V_{i}}^{\to} = (v_{i 1}, v_{i 2}, \dots, v_{id}, \dots, v_{iD})$ ; the optimal position of an individual particle (the optimal location of this particle) is ${P_{i}}^{\to} = (p_{i 1}, p_{i 2}, \dots, p_{id}, \dots, p_{iD})$ ; the optimal position of a particle swarm (the optimal location of any individual particle in this swarm) is ${P_{g}}^{\to} = (p_{g 1}, p_{g 2}, \dots, p_{gd}, \dots, p_{gD})$ . It does not lose generality. In terms of minimization, in the initial version of PSO, the iterative formula for the optimal location of individual particles is:

The optimal position of a swarm is the best position among optimal positions of individual particles in this swarm. The speed and position iterative formulas are:

v_{i, t + 1}^{d} = v_{i, t}^{d} + c_{1} * rand * (p_{i, t}^{d} - x_{i, t}^{d}) + c_{2} * Rand * (p_{g, t}^{d} - x_{i, t}^{d})

(2)

x_{i, t + 1}^{d} = x_{i, t}^{d} + v_{i, t + 1}^{d}

(3)

As the initial version of PSO does not have good application effects on optimization, an improved algorithm was put forward soon after the initial version of PSO was presented. The inertia weight ω is added to the speed iterative formula, so the speed iterative formula becomes:

v_{i, t + 1}^{d} = ω v_{i, t}^{d} + c_{1} * rand * (p_{i, t}^{d} - x_{i, t}^{d}) + c_{2} * Rand * (p_{g, t}^{d} - x_{i, t}^{d})

(4)

Although compared with the initial version, the complexity of the improved version is not greatly improved, the performance of the improved version is greatly improved. Therefore, the improved algorithm has been widely applied. In general, this improved algorithm is called standardized PSO, while the initial algorithm is called the original PSO.

By analyzing the convergence behavior of PSO, Clerc introduced a PSO variation with constriction factors. Constriction factors ensure convergence and improve the convergence speed. In this case, the speed iterative formula is:

v_{i, t + 1}^{d} = χ (v_{i, t}^{d} + φ_{1} * rand * (p_{i, t}^{d} - x_{i, t}^{d}) + φ_{2} * Rand * (p_{g, t}^{d} - x_{i, t}^{d}))

(5)

Obviously, iterative formulas (3.4) and (3.5) have no essential difference. As long as proper parameters are selected, the four formulas are identical.

The PSO has two versions: the global version and local version. In the global version, the two extreme values that a particle traces are, respectively, the optimal position of its own $p_{Best}$ and the optimal position of its swarm $g_{Best}$ . In the local version, besides the optimal position of its own $p_{Best}$ , a particle traces optimal positions $n_{Best}$ of all particles in the neighboring area, instead of the optimal position of its swarm $g_{Best}$ . In this case, the speed iterative formula changes to be:

v_{i, t + 1}^{d} = ω v_{i, t}^{d} + c_{1} * rand * (p_{i, t}^{d} - x_{i, t}^{d}) + c_{2} * rand * (p_{l, t}^{d} - x_{i, t}^{d})

(6) where

{P^{\to}}_{l}

is the optimal location of a local area.

Compared with GA, in most cases, the PSO can achieve faster convergence of all particles to the optimal solution. In the computation process with the PSO algorithm, when a particle finds the current best known position, other particles will quickly move closer to it. However, if the current best known position is the best known position of a sub-swarm, particles will not be able to detect the best known position of another sub-swarm. Therefore, the PSO would face pre-maturing convergence. The biggest drawback of PSO is that particles would be trapped in local optimal solutions, while searching the optimal solution of the entire swarm.

In this article, an improved algorithm is developed by combining the GA with PSO, with the aim to better improve the performance of the algorithm.

Strategy for improving the PSO algorithm

The GA is method for random global search and optimization, which is developed by mimicking the biological evolution mechanism in nature. It learns from Darwin's theory of evolution and Mendel's theory of heredity. The solution-seeking process of the GA is essentially to randomly search optimization, so it suffers no local convergence limitation. However, the GA is the optimization-seeking operation based on probabilities. Therefore, in most cases, the GA can only get the suboptimal solution of the entire swarm, but hardly obtain the optimal solution, because its search is random and blind.

Given that the GA and PSO have limitations, we combine the GA with PSO to seek optimal solutions to the set of evacuation instruction sequences. Specific calculation steps are as follows:

Step 1: work is done to randomly generate 10 particle positions and obtain corresponding 10 sets of evacuation instruction sequences based on the above-mentioned algorithm as the initial populations (particle swarms). The initialized extreme values of the entire swarm and individual particle are the evacuation fitness function value of the first individual particle.

Step 2: The above-mentioned algorithm is used to generate the evacuation fitness function value of the instruction sequence described by each individual particle in the current particle swarm.

Step 3: The current particle swarm is sorted according to the evacuation fitness function value to find the extreme values of individual particles and the entire particle swarm.

Step 4 (Selection): Five new randomly chosen particles are used to replace five particles with the lowest evacuation fitness function values in the current particle swarm.

Step 5 (Mutation): The roulette gambling method is used to mutate some individual particles in the current particle swarm.

Step 6 (Crossover): The updating rules of PSO are followed to update all individual particles in the current particle swarm.

Step 7: Check whether the maximum number of iterations has been reached. If yes, we can go to Step 8. If not, we should go back to Step 2.

Step 8: Simulation is finished. The specific computation process of the PSO can be seen in Figure 2. A detailed process is shown in Figure 1.

Figure 1.

The improved genetic algorithm.

Figure 2.

Diagrammatical drawing of the original α waveform.

Simulation results

A piece of program code is written for the ordinary filter algorithm, the GA and the improved PSO on the Matlab 2010a GUI platform, to solve EEG problems discussed in this article.

Specific simulation results of the program code are shown as follows:

α wave is collected, without any processing. The α waveform is shown in Figure 2.

The traditional method is used to filter α wave and get the waveform as shown in Figure 3.

The GA is used to filter α wave and get the waveform as shown in Figure 4.

The improved GA is used to filter α wave and get the waveform as shown in Figure 5.

The iteration–convergence curve of α waveform filtered by the GA is shown in Figure 6.

The iteration–convergence curve of α waveform filtered by the improved GA is shown in Figure 7.

Figure 3.

Diagrammatical drawing of the α waveform filtered by the traditional method.

Figure 4.

Diagrammatical drawing of the α waveform filtered by the GA.

Figure 5.

Diagrammatical drawing of the α waveform filtered by the improved GA.

Figure 6.

Iteration–convergence curve of α waveform filtered by the GA.

Figure 7.

Iteration–convergence curve of α waveform filtered by the improved GA.

According to a comparative analysis of α wave filtering effects of the three algorithms shown, respectively, in Figures 2, 3, and 4, it can be seen that the three filtering algorithms cause different energy losses. When the improved GA is used for filtering, the energy loss is the least, indicating that the improved GA shows best performance in preserving EEG information and produces best filtering effects. When the GA is used for filtering, the energy loss is less, indicating that the GA shows better performance in preserving EEG information and produces better filtering effects. When the traditional method is used for filtering, the energy loss is the highest, indicating that the traditional filtering method shows poor performance in preserving EEG information and has poor filtering effects. As can be seen from iteration–convergence curves in Figures 5 and 6, compared with the GA, the improved GA proposed in this study can search a more optimal solution, which can further prove that the improved GA proposed in this article is more suitable for filtering EEG.

Conclusion

This article mainly discusses EEG optimization in the mental workload measurement of miners. The PSO is combined with the GA to develop an improved GA for EEG filtering. To test the validity of this improved GA, a comparative analysis is made of the SGA and the traditional GA. Analytical results show that when the improved GA is used for filtering, it brings the least energy loss and shows best filtering effects. Compared with the traditional GA, the improved GA proposed in this study can seek a more optimal solution. This indicates that the improved GA, which is developed based on the PSO in this article, is more suitable for EEG filtering.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: National Natural Science Foundation of China (71271169, 71273208) and The Doctoral Program Foundation (20126121110004, 20116121110002).

References

Young

Brookhuis

Wickens

. State of science: mental workload in ergonomics. Ergonomics 2015; 58: 1–17.

Liao

. Mental workload and its measurement. J Syst Eng 1995; 10: 119–123.

Giraudet

St-Louis

Scannella

. P300 event-related potential as an indicator of inattentional deafness? PLoS One 2015; 10: e0118556–e0118556.

Zhao

Dong

Guo

. Hybrid brain-computer interface system based on αwave and motor imagery. J Electron Measure Instrum 2014; 28: 625–629.

Wang

. Driving fatigue detection based on EEG recognition and vehicle handling characteristics. Chinese J Sci Instrum 2014; 35: 398–404.

Wang

. EEG characteristic analysis of coach bus drivers in fatigue state. Chinese J Sci Instrum 2013; 34: 1146–1152.

Käthner

Wriessnegger

Müller-Putz

. Effects of mental workload and fatigue on the P300, alpha and theta band power during operation of an ERP (P300) brain–computer interface. Biol Psychol 2014; 102: 118–129.

Cinaz

Arnrich

La Marca

. Monitoring of mental workload levels during an everyday life office-work scenario. Person Ubiquitous Comput 2013; 17: 229–239.

Herff

Heger

Fortmann

. Mental workload during n-back task-quantified in the prefrontal corex using fNIRS. Front Hum Neurosci 2014; 7: 935.

10.

Ebrahimi

Vesin

Garcia

. Brain-computer interface in multimedia communication. IEEE Signal Process Mag 2003; 20: 14–24.

11.

Evstigneev

Filipenkov

Klochkov

. The EEG indicators of mental workload in pilots during the dynamic flight simulation on the human centrifuge. Int J Psychophysiol 2008; 69: 276–316.

12.

Noel

Bauer

Lanning

. Improving pilot mental workload classification through feature exploitation and combination: A feasibility study. Comput Oper Res 2005; 32: 2713–2730.

13.

Feng

Yi-Dan

Qin

. Recommendation algorithm of combining clustering with collaborative filtering based on genetic algorithm. Comput Technol Dev 2014; 24: 35–38.

14.

Dong

Guo-Wei

. Wireless sensor networks target tracking based on particle filtering optimized by quantum genetic algorithm. Sci Technol Eng 2013; 13: 3305–3309.

15.

Kai

Aiguo

Xia

. An adaptive filtering algorithm based on genetic algorithm-backpropagation network. Math Prob Eng 2013; 2013: 289–307.

16.

Cui S, Sun S, Gao SS, et al. Design of wideband non-equiripple filtering response using genetic algorithm based neural network. In: Piers Proceedings, August 2014, p.1364.

17.

Merabti

Massicotte

. Hardware implementation of a real-time genetic algorithm for adaptive filtering applications. In: 2014 IEEE 27th Canadian conference on electrical and computer engineering (CCECE). 2014: 1–5.

18.

Wang XP. Genetic algorithm: theory, application and software implementation. Xi'an: Xi'an Jiaotong University Press, 2002.

19.

. Channel selection by Rayleigh coefficient maximization based genetic algorithm for classifying single-trial motor imagery EEG. Neurocomputing 2013; 121: 423–433.

20.

Rejer

Lorenz

. Genetic algorithm and forward method for feature selection in EEG feature space. J Theor Appl Comput Sci 2013; 7: 72–82.

21.

Karakis

Capraz

Bilir

. EEG source localization using a genetic algorithm-based artificial neural network. Recent Patents Biomed Eng 2013; 6: 188–194.

22.

Magee

Givigi

. A genetic algorithm for single-trial P300 detection with a low-cost EEG headset. In: S2015 9th annual IEEE international,. 2015.

23.

Krzeszowski

Kwolek

Wojciechowski

. Presentation - Articulated body motion tracking by combined particle swarm optimization and particle filtering. Lecture Notes Comput Sci 2013, pp. 147–154.

24.

Chao

Huang

. Adaptive zero-phase filtering bandwidth of iterative learning control by particle swarm optimization. In: Proceedings of the 2nd international conference on intelligent technologies and engineering systems (ICITES2013), Springer International Publishing, 2014, pp. 1031–1037.

25.

Zhou

Yang

Gui

. A particle swarm optimization algorithm with variable random functions and mutation. Acta Automat Sin 2014; 40: 1339–1347.