Sensor-data Analytics in Cyber Physical Systems

From Husserl to Data Mining

Paul O’Leary, Matthew Harker and Christoph Gugg

Chair of Automation, University of Leoben, Leoben, Austria

Keywords:

Data Mining, Inverse Problems, Data Analytics, Phenomenology, Embedded Simulation.

Abstract:

This position paper proposes a discussion of the need for a solid philosophical basis for mining sensor-data

based on phenomenology. Additionally it is proposed that, when considering cyber physical systems, the

solution of inverse problems is a prerequisite if the results are to have physical meaning. A prototype lexical

analysis tool for sensor-data is presented and its application to knowledge discovery in large mechatronic

systems is demonstrated.

1 INTRODUCTION

There are many different deﬁnitions for what consti-

tutes a cyber physical system (CPS) (Baheti and Gill,

2011; Geisberger and Broy, 2012; Spath et al., 2013a;

NIST, 2013; NIST, 2012; Park et al., 2012; IOSB,

2013; Spath et al., 2013b; Lee, 2008; Tabuada, 2006).

The most succinct and pertinent to this paper is the

deﬁnition given by the IEEE (Baheti and Gill, 2011)

and ACM

: A CPS is a system with a coupling of the

cyber aspects of computing and communications with

the physical aspects of dynamics and engineering that

must abide by the laws of physics. This includes sen-

sor networks, real-time and hybrid systems.

Mining sensor-data from such system and per-

forming knowledge discovery is signiﬁcantly differ-

ent from mining static databases. The fundamental

question in any philosophy is: What is a valid source

of knowledge

and how do we acquire this knowl-

edge?

Consequently, any person developing a data an-

alytics system which includes knowledge discovery

must answer this philosophical question and the phi-

losophy must be implemented within the software.

There is no avoiding the truth of this statement. Most

authors, however, do not explicitly address this issue

in sufﬁcient depth. Fundamentally, there is no infor-

mation is the data stream itself. It is the association of

some form a model with the data and the parameters

ACM/IEEE International Conference on Cyber-

Physical Systems (ICCPS) (iccps.acm.org)

Sometime the word truth is used instead of knowledge.

of this model when applied to the data which reveal

the information. Or more precisely enable an infer-

ence with respect to the cause of the observation

Virtually all literature and standard works on data

mining, e.g., (Han, 2005, Chapters 2 and 3), deal with

data models, predominantly statistical data models,

as a means of revealing correlations. In the case of

mining sensor-data emanating from technical systems

we are observing data related to speciﬁc phenomena

which is inﬂuencing a system. In such cases if we are

to make any statement about the phenomena it is nec-

essary to consider the system model and not just data

models. Unfortunately, inverse problems associated

with such systems and sensor-data are not in general

addressed, see for example (Aggarwal, 2013) where

only data models are addressed.

If we are to perform sensor-data analytics, in a

manner such that the results have physical meaning,

i.e. obeying the laws of physics, then it will be nec-

essary to solve inverse problems in real time. Fur-

thermore, the use of multiple sensors in a large sys-

tem will require compact descriptions of complex

events. Humans use language to express complex

sensory experience and associated perceptions, since

a description of the sensory excitation would be too

longwinded and not understandable. But more im-

portantly, most of the information is to be found in

the common understanding of the word used; and is

If we have no model there can be no information. The

numbers in a computer have no meaning. The attachment

of meaning to a number is fundamentally adding a model.

Many working in sensor networks do this implicitly, wepre-

fer to do it explicitly.

176

O’Leary P., Harker M. and Gugg C..

Sensor-data Analytics in Cyber Physical Systems - From Husserl to Data Mining.

DOI: 10.5220/0005328601760181

In Proceedings of the 4th International Conference on Sensor Networks (SENSORNETS-2015), pages 176-181

ISBN: 978-989-758-086-4

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

not to be found in the sensor-data, but is associated

with the sensor-data through the use of the word.

In this position paper we wish to initiate a serious

and in depth discussion of the following issues:

1. Can the philosophy of phenomenology (Husserl

and Hardy, 1999; Merleau-Ponty, 2002) be a sup-

port during the design and partitioning of a sensor-

data mining system?

2. Do we need to consider the relevance of the de-

velopment of natural languages (Hmc, 2000, Ap-

pendices I and II) as a possible model for generat-

ing symbolic representations of sensor-data when

considering more complex events?

3. There is a very clear need to implement real-time

solutions of inverse problems if we are to extract

physically relevant information and knowledge.

Statistical data models are not sufﬁcient.

4. There is a need for embedded simulation to im-

prove the overall efﬁciency of real-time sensor-

data mining.

To support our hypothesis we demonstrate results

from a prototype implementation of sensor-data min-

ing systems. The system is motivated by the Indo-

Asian model for phenomenology and implements the

real-time solution of inverse problems and embedded

simulation. The system has been applied to the anal-

ysis of two different large pieces of machinery.

2 THE POSITION

This section presents a discussion of the elements

contained within this position paper. These are the

main topics on which we wish to initiate an in depth

discussion.

2.1 Phenomenology

Phenomenology as a branch of physics is deﬁned as:

a body of knowledge that relates empirical observa-

tions of phenomena to each other, in a way that is

consistent with fundamental theory, but is not directly

derived from theory. This is exactly what we wish to

do with data mining and knowledge discovery. How-

ever, statistical data models alone will not essentially

deliver results which are consistent with the funda-

mental theory of the system, since they make no at-

tempt to model the system being observed. Without

solving the inverse problems associated with the sys-

tem we can not guarantee consistency with the funda-

mental theory.

If we look closer at knowledge discovery we must

also consider phenomenology as a branch of phi-

losophy. Edmund Husserl

is considered to be the

founder of the philosophy of phenomenology which

is normally deﬁned as: the study of the develop-

ment of human consciousness and self-awareness as

a preface to or a part of philosophy. The ori-

gins, however, are much older in the 18th cen-

tury, the Swiss German mathematician and philoso-

pher Johann Heinrich Lambert applied the concept

to that part of his theory of knowledge that distin-

guishes truth from illusion and error. Phenomonology

was later extended by Martin Heidegger who intro-

duced the concepts of present-at-hand and ready-

to-hand. Which proposed that we gain all knowl-

edge only by studying our“average-everyday” un-

derstanding of the world gained from direct inter-

action with objects. Later Maurice Merleau-Ponty

with his best know work “Phenomenology of Percep-

tion” (Merleau-Ponty, 2002) analyzed how we per-

ceive as a result of experiencing phenomena, his con-

cept of embodiment, with a little over simpliﬁcation,

can be summarized as: we perceive phenomena ﬁrst,

then reﬂect on them - quite the opposite to “I think

therefore I am”.

The East-Asian view of phenomenology, as sum-

marized by the ﬁve aggregates (ﬁve Skandhas, see

table 1) (Lusthaus, 2002) is quite different. It de-

scribes the path from sensory excitation to discursive

thought in ﬁve distinct steps, i.e. the ﬁve aggregates.

The aggregates are only recently ﬁnding reception in

the western cognitive sciences (Davis and Thompson,

2013)

. The ﬁve steps are interesting since the assen-

Table 1: Abbreviated presentation of the ﬁve aggregates

(skandhas) and their possible technical interpretation.

tation is that we are never in direct contact with ob-

jects in the world, but always in contact with a model

for the world; with which we are back with the inverse

problems and models for the system being observed.

It is beyond the scope of this text to enter the de-

Edmund Husserl (1859-1938) graduated from the Uni-

versity of Vienna with a doctoral dissertation in theoretical

mathematics on the calculus of variations.

Some care must be taken when considering this text

since it is a western interpretation of an East-Asian philoso-

phy which is based on contemplative practice. The authors

are scholars and not contemplatives.

Sensor-dataAnalyticsinCyberPhysicalSystems-FromHusserltoDataMining

177

tails of the ﬁve Skandhas and how there are being re-

ceived within the cognitive sciences. It sufﬁces to say,

that independent of their ﬁnal truth or not, they do of-

fer a good basis for discussion on the partitioning of

a system and the suitability of which techniques at

which level in a cognitive system.

To summarize the issue of phenomenology:

1. system models are required and the correspond-

ing inverse problems must be solved if we are

to extract conclusions from sensor-data which are

based on physical systems.

2. there is much discussion on cognitive systems.

We believe it would be more appropriate to use

the term perceptive systems. Both the western and

eastern view is that perception is the product of

sensors activity and context, but not the product

of thinking. Thinking is the process which leads

to an understanding of the perceived.

The data-mining system presented later in this po-

sition paper implements the real-time solution of an

inverse problem for each and every sensor channel

and the possibility of an embedded simulation for ev-

ery actuator channel.

2.2 Natural Language

The Yogachara (Lusthaus, 2002)

school asserts that

language is generated when repetition is cognised in

sensory data. It is the repetition which is considered to

be characteristic and not the thinking about the repe-

tition, this process leads to a naming. At this point we

may not understand but we do perceive. In this man-

ner the use of a word is considered to be a metaphor

which points at a sensory experience rather than de-

scribing the experience directly. Monosyllabic words

are considered to represent simpler sensory experi-

ences, while polysyllabic words tend to describe more

abstract experience which are commonly the result of

complex multi-sensory experience.

When designing symbolic representations for

sensor-data which should enable complex symbolic

queries, see for example (Lin et al., 2003), we should

consider the metaphoric nature of the symbols and as-

sociate human readable text with the symbols. This

will facilitate greatly the process of knowledge dis-

covery and support the automatic generation of de-

rived polysyllabic symbols by combination of events,

Yogachara is a philosophical school founded by Va-

subandhu and Asanga in the 4

century in India. It was

carried to China where it was adopted and known under the

name Chan. A part of the philosophy is associated with the

Tridhatu, i.e., the relationship between the realms of body,

language and mind.

each of which was associated with a single channel

sensor-data.

As will be seen later we propose a method of

symbolic representation which at the scanner level

consists of tokens and predicates extracted from the

sensor data. The tokens and predicates can then be

combined to deﬁne syntax. This opens the door to

using concepts such as those embodied in Lex and

Yacc (Johnson, 1975) to enable BNF deﬁnitions for

the sensor-data analysis

and queries. With this the

step has been made from natural language to com-

puter implementable language.

2.3 Inverse Problems and Embedded

Simulation in Real-time

The details of a method and framework for the imple-

menting real-time solutions to inverse problems and

embedded simulation can be found in (O’Leary and

Harker, 2012) and the required automatic code gener-

ation in (Gugg et al., 2014). For this reason we shall

not go into more detail on the method at this point.

The solution approach requires the application of a

linear differential operator to the data stream.

3 THE PROTOTYPE

IMPLEMENTATION

In this section we present the initial design and imple-

mentation of a prototype sensor-data mining system

motivated by the position presented in this paper.

3.1 Single Channel Lexical Analyzer

The schematic diagram for the single channel lexical

analyzer (SCLA) is shown in Figure 1. It consists of

a linear differential operator (LDO) which is applied

to the sensor data stream. Each sensor channel has

its own linear operator associated with it. The LDO

can be used either to implement the solution to an in-

verse problem or to an embedded simulation, i.e. it

can solve ODEs in both a forward and an inverse man-

ner. The lexical analysis element corresponds to sym-

bol aggregate approximation (SAX) (Lin et al., 2003).

However, it is not applied directly to the sensor data

but to the processed sensor data. As an extension of

the SAX approach k-means clustering is applied to

the processed signal to identify repetitive clusterings

It is important to differentiate between markup lan-

guages for sensor data, e.g. sensorML, and markup for the

sensor-data analysis and queries.

SENSORNETS2015-4thInternationalConferenceonSensorNetworks

178

single channel lexical analyzer

linear differential

operation

symbolization

lexical

compression

processed data

symbols

tokens

lexical deﬁnition

data stream

Figure 1: Schematic of the single channel lexical analyzer.

of values

. Each cluster is assigned a symbol and a

human readable text describing the meaning of this

cluster. Finally the lexical compression generate the

tokens and predicates, i.e. each persistent symbol is

replaced by a single token with a length predicate and

pointer which link the token to its start and end po-

sitions in the real data. This enables backtracking of

the symbolic queries to real sensor data values.

The SCLA can be deﬁned in a simple structure

and an array of such structures is used to manage the

deﬁnitions of all signals and lexical analyzers. In this

manner the implementing is not in the form of ded-

icated code, but as a generic processing engine. Its

input is the set of sensor-data and the lexical deﬁni-

tions and it returns the token-tables for each SCLA

and their combinations. In this manner applying the

system to a new machine only required the deﬁnition

of a new set of SCLA parameters and combinations.

3.2 Multi Channel Lexical Analyzer

The schematic for the muiti-channel lexical analyzer

is shown in Figure 2. The outputs of the SCLAs are

grouped according to their relevance to speciﬁc op-

erations. The token tables are merged automatically,

and the statistics for each token combination are com-

puted during this task. This enables the identiﬁcation

of repetitive events, events which are seldom and of

events which do not occur. The output is itself once

again a token table. This is the ﬁrst level of automatic

knowledge discovers, since it characterizes the opera-

tions of the machines.

Symbolic queries are implemented as regular ex-

pressions which are applied to the token tables and

their predicates. In this manner speciﬁc sequences of

events can be located. The lexical compression im-

plicitly implements dynamic time warping (DTW),

and there is a positive semi deﬁnite distance metric

associated with the tokens and their predicates. This

ensures that it is possible to do sequence similarity

measurements.

The ﬁnal stage of the MCLA is the generation of

a dictionary. Since all symbols are associated with

K-means clustering was chosen because the histogram

exhibits clear peaks and the data is well modelled by a sum

of gaussian distributions.

multi channel lexical analyzer

SCLA 1

SCLA 2

SCLA

...

token

merge

statistical

properties

text

merged

tokens

MCLA

lexical

deﬁnition

textual

mapping

(FIFO)

token

permutations

stat.

eval.

Figure 2: Schematic of the multi channel lexical analyzer.

human readable text it is possible to generate dictio-

nary entries automatically which explain the mean-

ing of tokens derived from combining single token

tables. The combination of multiple tokens is con-

sidered akin to the formation of polysyllabic words.

4 DEMONSTRATION

The method has been applied to a number of different

machines to evaluate the feasibility. Here we present

the results of the application of the system to a large

scale reclaimer, see Figure 3.

4.1 SCLA Demonstration

The slewing data from the machine is used to demon-

strate a simple single channel lexical analyser. The

data channel consists of n = 840000 samples. We

are interested in determining the rate of change of

slewing, i.e., the speed and direction of slewing. For

this reason the linear differential operator implements

a regularizing differentiation (O’Leary and Harker,

2010). The slew data and the result of the compu-

tation are shown in Figure 4.

Figure 3: The large scale reclaimer used to demonstrate the

concepts presented in this paper.

Sensor-dataAnalyticsinCyberPhysicalSystems-FromHusserltoDataMining

179

1 2 3 4 5 6 7 8

x 10

100

150

200

250

Slew

1 2 3 4 5 6 7 8

x 10

−40

−20

Figure 4: Top: the slew data and bottom: the rate of change

of slew. The data consists of n = 86225 samples, a day’s

production sampled at t

= 1s.

−50 −40 −30 −20 −10 0 10 20 30 40 50

200

400

600

800

1000

1200

Bins

Frequency

Figure 5: Histogram of the rate of change of slew. This is

used to identify the levels for the symbolic aggregate ap-

proximation.

During the exploratory phase of operation, a k-

means clustering is applied of the histogram for the

rate of slewing, see Figure 5. For simplicity of

demonstration three clusters are used and assigned the

symbols [1, 2, 3] with the associated human readable

texts [slewing-left, no-slewing, slewing-right].

The generated token table for the slew data now

enables symbolic queries. The result of a sym-

bolic query for an interrupted right slew movement

is shown in Figure 6.

4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 5

x 10

190

200

210

220

230

240

250

260

270

280

Slew

Query = Right-stop-right

Figure 6: A symbolic query for an interrupted right slewing

of the machine, abbreviated as right-stop-right. The occur-

rences of the query in the data are marked in red.

4.2 MCLA and Knowledge Discovery

Here we demonstrate the implementation of SCLAs

for the slew, long-position and tonnage of this ma-

chine, each with a different LDO, see Figure 7 for a

visual representation. Then the token tables for the

three channels are merged automatically to identify

more complex machine operations.

The data set has n = 86225 samples for each sen-

sor, i.e. a full day of operation samples at t

= 1s. The

combined token table has m = 578 entries, only these

entries need to be searched during a symbolic query

- this corresponds to a very high degree of compres-

sion. Additionally the system generates a statistical

analysis for the identiﬁed tokens, see Table 2. It can

be seen that certain combinations of events never oc-

cur, this is knowledge discovery.

To get a measure for performance data was taken

from a second machine with a sampling time t

20ms. The data acquisition was started at t

= 22 −

Jan − 201410 : 39 : 09 and stopped t

= 24− Jan −

201422 : 18 : 17, that is for a time period of approxi-

mately 60 hours. The generation of the token table for

a single channel required t

= 0.84sec this includes

the solution of the inverse problem. A symbolic query

on this data set required between t

= 20. . . 30ms de-

pending on the query. This computation was per-

formed on a 2.4GHz dual core processor.

Figure 7: Three SCLA channels and the automatically gen-

erated vocabulary for the combination of these channels.

This is only for a small zoomed portion of the data. The data

set has n = 86225 samples for each sensor, i.e. a full day

of operation samples at t

= 1s. The combined token table

has m = 578 entries, only these entries need to be searched

during a symbolic query.

5 CONCLUSIONS

It is concluded that the concepts of the philosophy of

phenomenology are of great support when designing

sensor-data analytics, data mining and knowledgedis-

covery systems. If complex events are to be automat-

ically discovered and identiﬁed it will be necessary to

consider the use of language to implement metaphoric

abbreviations, i.e. terms which abbreviatethe descrip-

tion of complex sensorial events. Furthermore, it has

SENSORNETS2015-4thInternationalConferenceonSensorNetworks

180

Table 2: Statistical summary of the result of merging the

token tables for slew, long-position and tonnage. Note some

operations are automatically identiﬁed as not occurring.

been argued that the solution of inverse problems is

indispensable if the results are to have physical mean-

ing. The demonstration of the application of the sys-

tem to a large mechatronic system has demonstrated

the functionality of the concept.

The issue of scale space has still not been ad-

dressed. That is, there are features which are relevant

at different time scales: some are relevant in millisec-

ond ranges and some in hours. We are presently in-

vestigating a structured decimation of the data.

REFERENCES

Aggarwal, C. C. (2013). Managing and Mining Sensor

Data. Springer Publishing Company, Incorporated.

Baheti, R. and Gill, H. (2011). Cyber-physical systems. The

Impact of Control Technology, pages 161–166.

Davis, J. H. and Thompson, E. (2013). From the Five Aggre-

gates to Phenomenal Consciousness, in A Companion

to Buddhist Philosophy (ed S. M. Emmanuel). John

Wiley and Sons, Ltd, Chichester, UK.

Geisberger, E. and Broy, M. (2012). agendaCPS: Integri-

erte Forschungsagenda Cyber-Physical Systems, vol-

ume 1. Springer.

Gugg, C., Harker, M., O’Leary, P., and Rath, G. (2014).

An algebraic framework for the real-time solution

of inverse problems on embedded systems. CoRR,

abs/1406.0380.

Han, J. (2005). Data Mining: Concepts and Techniques.

Morgan Kaufmann Publishers Inc., San Francisco,

CA, USA.

Hmc, O. (2000). The American Heritage Dictionary of the

English Language. Houghton Mifﬂin.

Husserl, E. and Hardy, L. (1999). The Idea of Phe-

nomenology. Husserliana: Edmund Husserl – Col-

lected Works. Springer Netherlands.

IOSB, F. (2013). Industry 4.0 information technology is

the key element in the factory of the future. Press In-

formation.

Johnson, S. C. (1975). Yacc: Yet another compiler-

compiler. Technical report, Computing Science Tech-

nical Report No. 32, Bell Laboratories, Murray hill,

New Jersey.

Lee, E. A. (2008). Cyber physical systems: Design chal-

lenges. In Object Oriented Real-Time Distributed

Computing (ISORC), 2008 11th IEEE International

Symposium on, pages 363–369. IEEE.

Lin, J., Keogh, E., Lonardi, S., and Chiu, B. (2003). A sym-

bolic representation of time series, with implications

for streaming algorithms. In Proceedings of the 8th

ACM SIGMOD Workshop on Research Issues in Data

Mining and Knowledge Discovery, DMKD ’03, pages

2–11, New York, NY, USA. ACM.

Lusthaus, D. (2002). Buddhist Phenomenology: A Philo-

sophical Investigation of Yog¯ac¯ara Buddhism and the

Cheng Wei-shih Lun. Curzon critical studies in Bud-

dhism. Routledge Curzon.

Merleau-Ponty, M. (2002). Phenomenology of Perception.

Routledge classics. Routledge.

NIST (2012). Cyber-physical systems: Situation analysis

of current trends, technologies, and challenges. Tech-

nical report, National Institute of Standards and Tech-

nology (NIST).

NIST (2013). Strategic r&d opportunities for 21st century

cyber-physical systems. Technical report, National In-

stitute of Standards and Technology (NIST).

O’Leary, P. and Harker, M. (2010). Discrete polynomial

moments and savitzky-golay smoothing. In Waset

Special Journal, volume 72, pages 439–443.

O’Leary, P. and Harker, M. (2012). A framework for the

evaluation of inclinometer data in the measurement

of structures. IEEE T. Instrumentation and Measure-

ment, 61(5):1237–1251.

Park, K.-J., Zheng, R., and Liu, X. (2012). Cyber-physical

systems: Milestones and research challenges. Com-

puter Communications, 36(1):1–7.

Spath, D., Ganschar, O., Gerlach, S., H¨ammerle, M.,

Krause, T., and Schlund, S. (2013a). Produktion-

sarbeit der Zukunft-Industrie 4.0. Fraunhofer IAO

Stuttgart.

Spath, D., Gerlach, S., H¨ammerle, M., Schlund, S., and

Str¨olin, T. (2013b). Cyber-physical system for self-

organised and ﬂexible labour utilisation. Personnel,

50:22.

Tabuada, P. (2006). Cyber-physical systems: Position paper.

In NSF Workshop on Cyber-Physical Systems.

Sensor-dataAnalyticsinCyberPhysicalSystems-FromHusserltoDataMining

181