Machine adaptive methods allow the analysis of various GPSparameters (tuple) in order to track the movements of human beings and to distinguish their positions from critical to uncritical working environments inside a petrochemical plant. A programmed learning capability has been implemented into the prototypic Z-Tracker-Personal-Emergency-System (PES), based on the Naive Bayes Classification of GPS-related signal pattern. In this context each pattern can be taken as typical for a covered or non covered working environment. Step by step the PES gains a deeper knowledge regarding to typical, whereabouts within the infrastructure of an industrial plant depending on specific GPS signal pattern and their related probable type of working environment. The System is able to decide whether a new tuple pattern represents an uncritical, tractable movement (class) of a person or a highly critical steady state position (class) after a vital accident due to an emergency. As a direct consequence of the learning process mathematical regularities can be deduced for each possible location scenario. Field studies in the Leuna Refinery (Germany) revealed a high reliability of the ZTracker-Personal-Emergency-System.
In an industrial complex like a petrochemical refinery a high potential of hazards does
exists concerning the personal safety of the service crew. The periodical maintenance
cycles of such plants are given by German law (Behälterverordnung) [
The Z-Tracker-PES offers a possibility to locate moving objects in the plant and to
classify their positions regarding to emergency situations. It consists of the
Z-TrackerGPS-handheld device (to track and locate a person) and the Z-Tracker-Background
Analytical System (to evaluate and visualize the received position statements).The
position of a person is calculated by the integrated GPS-chip [
The characteristic type of environment within the refinery, i.e. piping,
constructions, containers or bearing structures, which possibly surrounds a candidate
person affects its positioning in the way that the calculation of the GPS data exhibits a
typical location failure (hereby assigned to covered and not covered positions inside the
refinery, Figure 2). In order to detect and to classify such failures in the positioning of
personnel the adaptive machine learning concept of the Naive Bayes Classifier has been
applied to the PES [
The basis of the Naive Bayes Classifier (NBC) is the Bayes Theorem (1) which allows calculations considering conditional probabilities. (1)
The NBC is a supervised learning method and is well proven for other purposes
which classify instances by a tuple of attributes to predefined classes. It has to be stated
that the results of the NBC are as good as other comparable algorithms, like neural
networks or decision trees [
Applying the NBC means to describe every instance X by a tuple of attributes ai. In a target function these attributes are to be classified representing the instance X and is therefore able to represent every possible class.
The attributes of an instance should be similar and conditional independent to each
other. In reality this norm is not often realized because most attributes affect each other
in various mutual ways. Anyhow, even if the naive conditional independence
assumption is not fulfilled totally in the model the NBC shows a quite good
performance [
SNR SNR PDOP HDOP VDOP Number of Satellites
Position-Fix frequency of the test volume allows the calculation of probabilities for each type of tuple.
At this point the membership of each new, unknown class to every possible class
will be calculated and as a result the instance will assigned to the class with the highest
likelihood of membership. The function of the NBC in this case is as follows:
vNB = arg max P(v j )∏ P(ai | v j )
v j∈V i
(2)
Where P(ai | v j ) represents the single probability of the attribute to be assigned to the
class v j . The product of all ai is the probability of each class v j , where arg max
represents an unknown instance and its association to the most likely classification [
The classification of covered and uncovered locations is represented by the same binary classes and an unknown instance is classified into these classes by the target function.
As mentioned above an attribute tuple characterizes an instance and with regard to the
calculation of the correlation coefficient and the rank correlation coefficient
respectively [
The NBC implemented in the PES is trained by known sample GPS data representing specific test scenarios. (thereby the probability for all attributes will be assigned; Table 2 is an example of the calculation of the probabilities for the SNR). This is achieved by counting the single parameter values in the distribution from each attribute (NBh = single probability). Afterwards the training will be improved by another sample of the validation data (test scenarios) being discrete to the training data collected by the ZTracker (Figure 3). 2.4. Evaluation of the adaptive learning method The performance of the machine learning is estimated by a confusion matrix. The matrix is with regard to the two possible classes covered and uncovered situations divided in two columns and two rows. The expected value is signed into the columns and the estimated value into the rows. The main diagonal represents the correct matches while the mismatching values are located aside of them. By counting the right and false matches one is able to validate the performance of the learning. Therefore the expected and the estimated value of the covered and not covered classes are set into relation. The best combination of attributes to classify an instance was the SNR and the Number of used satellites, in this particular case containing only one misclassified instance (Table 3).
It became clear that other combination of attributes (tuple), for instance the performance of PDOP and Number of used satellites (Table 4) leads to a less reliable result.
303 140 1 350 45 210 99,86 73,50
This study underlines that machine learning methods, if implemented into GPS movement monitoring systems like the Z-Tracker PES, are able to distinguish between uncritical not covered and critical covered positions of personnel within an industrial plant like a refinery. In this context the NBC has been proved as a reliable method to evaluate GPS measurements in order to gain information about the potential spatial hazards for employees within their working environment.
One of the future requirements which can be deduced from this prototypic model is the applicability of the PES regarding to other types of petrochemical plants or any other industrial complex with an ‘endemic’ local primary spatial location systems: In that case: Is it possible to apply the same kind of NBC-knowledge base or should the system be trained by additional data (satellite or terrestrial signals) to achieve a more differentiated knowledge about an instance or location? Could this additional information be measured by the Z-Tracker or other handheld devices (like a GPSphone or a data-logger)?
Also new developments in the (D) GPS/Galileo Systems might increase the capability of the Z-Tracker-PES, like the exploitation of differentiated frequencies or a faster and higher mobile data communication. Further applications of a NBC-trained location system are likely in the entertainment industries or special health services.