System Identification Techniques within Revealing Human-Operator Skills

A new approach to abnormal situations with regard for the heuristic regularities of human-operator thinking process is proposed. The regularities are revealed on basis of recording the motions of the humanoperator eyes over the information field of the control board and processing the experimental data obtained. For data processing, a probability theoretical approach is utilized. Such an approach based on involving the notion of consistency of measures of dependence of random variables. Within the approach, a set of the so called information correlations has been proposed to serve as a quantitative performance index of humanoperator skills. 1. State of the Art and Problem Description Numerous papers are devoted to investigation of various issues of design and performance of humanmachine systems with regard for the human factor [BAB95], [BLA00], [BUS93], [CAL00], [HUN93], [IIG98], [LEE00], [MCC97], [MEN01], [PAN91], [PAN96], [RAM90], [RED01], [ROU89], [SMI98], [SRI99], [STE93], {STO88], [TSV99], [XU 02], [YAN92], paying special attention to revealing/identification of humanoperator skills and experience. Human-operator decision making under abnormal situations is the most main and the most difficult task of their professional activity, with the task being considerably complicated by incomplete or false information, stress, or lack of time, as well as aggravation by the huge cost of the human-operator errors. The abnormal situations imply the human-operator to involve intelligent actions and make decisions which are not envisioned by nominal operator manuals. The decision making process is of a heuristic nature, and analysis and localization by skilled operators of unexpected abnormal situations which never occurred or were envisioned theoretically fall into the category of creative nature processes which are accompanied by a specific state of emotional stress [TSV99]. Within the above mentioned branch, regarding for human skills modeling, the fundamental paper [ROU89] plays so far a basic role, and only few papers may be related to such a “narrowed” branch of humanmachine researches. Determination and study of the actual algorithm of a human-operator performance involve some methodological difficulties. Within these, a new approach to the problem of human-operator describing/modeling which is based on studying the regularities of actions of skilled human-operators, whose decision making can be regarded as a reference both in terms of its structure and characteristics, should be considered as the most vital one. A unique measure of human attention behavior may be implemented by use of eye tracking. This is particularly important in evaluating present and future environments in which humans do and will work. The examination of human interaction and behavior with their environments, particularly ones in which humans often perform functions critical to safety, is one topic studied by human factors and industrial engineers. Traditional measurement methods of human performance often include measures of reaction time and accuracy, e.g., how fast a person completes a task and how well this task is performed. These are generally measures associated with performance. To study the steps taken to perform the tasks requires analysis of the individual procedures performed. For this analysis, process measures are often needed. Eye movements are particularly interesting in this latter context since the present measures which can provide insight into the visual, cognitive, and attentional aspects of human performance [DUC03]. At present, there is available a wide variety of eye movement recording techniques involving the measurement of distinguishable features of the eyes under rotation/translation, e.g., the apparent shape of the pupil, the position of the limbus (the iris-sclera boundary), and corneal reflections of a closely situated directed light source (often infra-red). In order to disambiguate head movement from eye rotation, multiple ocular features must be measured. Two such features are the corneal reflection (of a light source, usually infra-red) and the pupil center. The corneal reflection of the light source is measured relative to the location of the pupil center. Corneal reflections are known as the Purkinje reflections, or Purkinje images. Due to the construction of the eye, four Purkinje reflections are formed: 1, reflection from front surface of the cornea; 2, reflection from rear surface of the cornea; 3, reflection from front surface of the lens; 4, reflection from rear surface of the lens being almost the same size and formed in the same plane as the first Purkinje image, but due to change in index of refraction at rear of lens, intensity is less 1% of that of the first Purkinje image. Video-based eye trackers typically locate the first Purkinje image. With appropriate calibration procedures, these eye trackers are capable of measuring a viewer’s Point Of Regard (POR) on a suitable positioned surface on which calibration points are displayed [DUC03]. Within the problem outlined, one may state a problem of assessment/identification of industrial plant human-operator skills. Then, within the framework, skills of experienced human-operators will be considered as a basic, i.e. “ideal” with respect to the given plant control task, knowledge. Corresponding investigation is implemented basing on analysis of the time which is required by the human-operator to make a decision on the plant process behavior using the information provided by detectors distributed over an information/control board. The detectors on the information board are associated in groups in accordance with types of information they provide. These groups are considered as zones. To make a total decision with respect to the plant process behavior, the human-operator has to take certain time to work with a zone, where “time to work with a zone” means the time which the human-operator spends during looking on the zone. In turn, such a time is fixed on basis of the human-operator eye motions by use of the corresponding hardware, the eye trackers. Availability of cheap, fast, accurate and usable eye trackers provides wide rage of applications [CLA98] (and references therein)), [DUC02a], [DUC02b), [GRA99], [TSV99], [VOR02]. Analysis of the actual human-operator performance algorithm and its deviations from the nominal one enables one to elucidate the general structural features of the decision making process and prove that they are repeatable and stable to various operating conditions. Such an analysis allows one to establish how the structure of decision making varies vs. the complexity of the abnormal situation, its novelty, and the human-operator experience, as well as to reveal hidden human-operator errors made during information collection. In turn, the fact that the structure of the actual human-operator performance algorithm involves characteristics which are independent of the professional and individual distinctions of the human-operator, the complexity and novelty of the abnormal tasks, and the quality of their execution enables one to identify the human factors which must be taken into account to improve the human-machine systems. Studies have demonstrated that the deviations of the human-operator activity from the nominal logical schemes of execution are not occasional or due to their insufficient training, but rather reflect the basic human, heuristic aspects of operational thinking. Along with identifying abnormal disturbances, the activity of the human-operators involves analysis of the state of the entire plant and control system; analysis of the reserves that would assume the load in case of possible disabling of faulty units; prediction of possible variants of accident progression; modeling of the actions to eliminate the given accident and of the variants of its progression; analysis of possible – even improbable – causes of the accident [TSV99]. The structure of decision making is characterized by concurrent execution of the abnormal situation task, checking that no other accidents occurred at the plant, and clarifying the possibility of driving the plant to the optimal state in the presence of this accident. The all three tasks are executed against the background of repeated checking of the data selected for decision making, the time required by the last two tasks ranging from 30% to 80% of the total time of execution. During the execution, the nominal human-operator performance algorithm is entwined with the actual one; its sequence is violated, and returns are added to it. More importantly, there are no accepted psychological phases of decision making which correlate with the duration of the fixing motions of eyes, such as familiarization, analysis, variant choice, search for a new variant, etc. Upon receiving information about an abnormal situation, the operator immediately hypothesizes about the nature of the accident and checks it [TSV99]. A subtle investigation of human oculomotor system under a challenging search task is presented in [AKS02]. 2. Analytical Tools and Model Constructing In accordance with reasoning of the above section, the problem of assessment/identification of humanoperator skills is to be considered, from a system theory point of view, as a problem of deriving a model of a mildly formalized system. Consequently, the human-operator model may be considered in terms of system input/output description, with available for observation input and output variables reflecting significant features of the model. Even if no exact analytical model of the input/output relationship between the variables is stated, obviously, there always exists an inherent link which reflects dependence of the output variables from the input ones. Thus, the above described time which the human-operator spends during looking on each zone i is considered as a realization of the input variable ) (i zone τ , while the time required to make a final decision with respect to the plant process behavior is considered as a realization of the output variable dec Τ . Within the modeling scheme, for a false decision making, the value of the corresponding decision making time is imposed to be equal to infinity. Finally, dec Τ should be expressed as a function of ) (i zone τ , N i , , 1K = , where N stands for the number of zones. For a stochastic case, a natural way to establish an approximate empirical input/output relationship is using measures of stochastic dependence of random variables. The entity of the approach proposed in the paper is, thus, eliciting the above mentioned inherent link between the (output) variable dec Τ and (input) variables ) (i zone τ , N i , , 1K = . Such a link, when expressed quantitatively, is to serve to reflect the human-operator skills on the plant process control. Hence, a significant feature of such an approach to human-operator skills identification is just the choice of an appropriate measure of stochastic dependence between random variables. Among various measures of dependence, the product correlation is well known and commonly used. However, the ordinary product correlation of random variables Y and X may vanish even provided that the random variables are completely dependent, i.e. if there exists a deterministic function ) (• f such that ) (X f Y = with probability 1 [RAJ81], [REN59]. Within the paper approach, the basic requirement imposed on a measure of dependence to be used is the measure to be consistent, i.e. (following to Kolmogorov’ terminology) vanishing if and only if the random variables are stochastically independent. In sequel, throughout the paper, the symbols ) (• M , ) (• D , { } • • M , { } • • D , and ) , ( • • cov will respectively stand for the mathematical expectation, variance, conditional expectation, conditional variance, and covariance. In [REN9] seven axioms which are seemed to be the most natural for a measure of dependence ( ) Y X , μ between two random variables X and Y has been presented. These are: A) ( ) Y X , μ is defined for any pair of random variables X and Y, neither of them being constant with probability 1. B) ( ) Y X , μ = ( ) X Y , μ . C) ( ) 1 , 0 ≤ ≤ Y X μ . D) ( ) 0 , = Y X μ if and only if X and Y are independent. E) ( ) 1 , = Y X μ if there is a strict dependence between X and Y, i.e. either ) (X Y φ = or ) (Y X ψ = where φ and ψ are Borel-measurable functions. F) If a Borel-measurable functions φ and ψ map the real axis in a one-to-one way onto itself, ( ) ( ) Y X Y X , ) ( ), ( μ ψ φ μ = . G) If the joint distribution of X and Y is normal, then ( ) ( ) Y X r Y X , , = μ , where ( ) Y X r , is the ordinary correlation coefficient of X and Y. Commonly used measures of dependence are the ordinary correlation coefficient, the correlation ratio