1

Predictive maintenance framework:

1.1. Abstract
This paper presents a predictive maintenance framework for CNC machines
focusing on a multi-stage prediction of machines’ health status. For the
implementation of such a multi-stage prediction, the proposed approach includes two
prediction layers: the cloud prediction layer and the local prediction layer. Each layer
provides a prediction of machine health status in different timescale. The local
prediction layer, based on data analysis techniques, is responsible to predict the health
status of the machine for a short time period. Thus, this prediction can be used as an
alarm aiming to prevent un-expected breakdowns. The cloud prediction layer, based
on digital physical-based models, is responsible to provide a more general overview
of machine health status using Prognostics and Health Management (PHM)
techniques, useful for long timespan strategies definition. This paper presents the
proposed approach and its benefits are described and discussed. The proposed
approach will be implemented in the PROGRAMS project.

Chapter written by P. Aivaliotis1, K. Georgoulias1,*, R. Ricatto2, M. Surico2

1 Laboratory for manufacturing Systems and Automation (LMS), University of Patras, 26504,
Greece, Tel: +30-2610-910160
2 Fidia S.p.A., C.so Lombardia, 11 - 10099 San Mauro Torinese TO, Italy, Tel: +39-011-

2227111

* Corresponding author email: [email protected]

1.2. Introduction
Maintenance and its related costs continue, over the years, to draw the attention of
production management since the unplanned failures decrease the reliability of the
system, and also the return on investments [MOU 15]. Taking under consideration
that the maintenance accounts for as much as 60 to 70% of the production lifecycle
total costs, it is a core activity of most industrial sectors [DHI 06]. More specifically,
industrial studies have revealed that the cost of replacing worn-out components may
be as high as 70% of the total maintenance cost [VEN 10].
Nowadays, the high complexity is an important characteristic of the CNC
machines which are used for industrial applications. Several hundreds of components
are required to allow the CNC machine functionality. For example, the
implementation of a maintenance strategy requires continuous monitoring of each
piece of equipment to be aware of its real deterioration status. Due to the high
complexity of the CNC machines, a multi-stage predictive maintenance framework is
needed to enable the prediction of machine health status over different timescales,
aiming to ensure the prediction over both long and short time periods of operation.
One the one hand, a long-time period prediction of the CNC machines’ health
status is useful for the precise selection and scheduling of the best moment for
maintenance, right before a component breaks down. However, prediction like this
requires a long time period simulation and, accordingly, huge computational time. On
the other hand, a short time period prediction is required to prevent un-expected
breakdowns in the near future. Such predictions are based on the use of historical data
and comparison techniques, which require small computational time.

1.3. Method
This paper presents an approach which implements this integration of long and
short term prediction in a multi-stage predictive maintenance framework with two
different layers: the cloud prediction layer and the local prediction layer. The cloud
prediction layer is responsible to provide a more general overview of machine health
status using Prognostics and Health Management (PHM) techniques over a long time
period. The local prediction layer is responsible to predict the health status of the
machine using data driven models over a short time period.
The main pillar of a PHM system is the prediction of the Remaining Useful Life
(RUL) that depicts the time after which a system or a machine’s component no longer
performs its intended function [AIV 2017]. According to previous research, a
maintenance action should be executed taking under consideration both health
assessment information (such as RUL) as well as additional information from multi
criteria mechanisms [PAP 07]. The RUL is the most important parameter to be taken
under consideration for the creation and execution of a maintenance plan [OKO 04].
Thus, a number of predictive maintenance platforms are based on RUL prediction to
fulfil the needs of the data analysis and knowledge management [MYC 10]. These
platforms are based on three main stages: the first stage is responsible for data
3 Predictive maintenance framework: Implementation of local and cloud processing for multi-
stage prediction of CNC machines’ health

extraction and processing; the second one focuses on the maintenance knowledge
modelling and calculation of RUL; and the third stage provides advisory capabilities
for maintenance planning [EFT 12].
In the proposed research, machine component physical-based models will be used
for the calculation of the RUL for each machines’ components. The proposed method
will be used in the PROGRAMS project. More specific, some data will be gathered
by the machines’ controllers and external sensors, which will be structured and
uploaded to a cloud database. Some of this data will be used for the simulation of the
digital models, while some other, will be used to update the simulation models, aiming
to ensure that the simulated functionalities of the machines will be the same as the
real one. Therefore, a digital twin of the real production equipment will be created.
Finally, the upcoming process plan of the machine will make up the input for the
simulation model. The output of the simulation, in combination with the reliability
parameters of the machines and the real time monitored data, will be used for the final
RUL calculation. The above procedure is depicted in Figure 1.

Figure 1. RUL calculation main concept [4]

For the local prediction layer and its prediction techniques, the proposed approach
is based on algorithms that use the data-driven approach to prognosis learn models
directly from the data, rather than using a hand-built model based on human expertise
[SCH 05]. Considering that data batches collected from a repetitive operation are
generally similar unless in the presence of an abnormality, a condition change can be
inferred by comparing the monitored data against an available nominal batch [BIT
14]. Other researches attempt to address this problem with intelligence-oriented
techniques, such as dynamic wavelet neural networks (DWNNs). DWNNs
incorporate temporal information and storage capacity into their functionality so that
they can predict the future, carrying out fault prognostic tasks [WAN 01].
In this research, an already existing data-driven prediction methodology based on
Artificial Neural Networks will be used for the local prediction. This proposed data
driven method is not a part of the PROGRAMS project. Available historical data will
4 Predictive maintenance framework: Implementation of local and cloud processing for multi-
stage prediction of CNC machines’ health

be used for the training of the neural network and the real time gathered data will be
compared with the historical one aiming to identify potential upcoming failures.
The strong point of the presented approach is the integration of the physical-based
and data-driven methodologies for the prediction of the machines’ health status in a
multi-dimensional time frame. The benefits of this approach when applied to CNC
driven industrial equipment is presented in detail in the following chapter.

1.4. Approach
Two main factors have limited the diffusion of predictive maintenance so far: on
the one hand, the huge amount of data that must be preventively collected in order to
train and validate the health status prediction algorithms; on the other hand, the
somewhat limited choice left to the monitored production equipment owner (replace
the worn component or live with it). This scenario however changes drastically when
machine tools equipped with CNCs are considered. In fact, latest generation of smart
CNCs allow to record internal signals with sample rate near or equal to the CNC cycle
time (usually 1-2 ms), to internally analyse real time data to detect if an equipment
component is working and under which stress conditions, and to exploit knowledge
of equipment’s health status to adjust the processing performance.
The data produced can successively be moved to a remote cloud storage (for
example a factory wide database) for applications that go beyond the straight process
management. The added value of implementing these features directly inside the
CNC, as opposed to using external Programmable Logic Controllers (PLCs) and
Programmable Automation Controllers (PACs), is that the bulk of machine tool data
and the capability for process parameters management are already available and ready
to be exploited. In addition, only a limited set of CNCs allow to relinquish the control
of process parameters to third party applications, both for quality management and for
security/safety reasons. It then becomes clear that a new maintenance paradigm can
be originated, by combining these functionalities with an advanced tool capable of
using smart (data driven or model based) algorithms, to understand the equipment’s
current status and to predict the time when it will become irreversibly worn out. With
this objective in mind, a smart CNC producer (like FIDIA) can decide to interface its
controllers with a new maintenance platform, in order to offer a whole new set of
functionalities to its customers.
The health status of production equipment can be computed by data driven
algorithms. The main advantage of these tools is that they do not rely on complicated
system modelling and can be thus applied directly to the historical CNC records (for
training) or even to real time data (for fast evaluation of the current equipment’s
status). Once the health status of main equipment’s components is available, the CNC
must then correlate it with the impact it could have on the process. In fact, a worn
component is usually not an important enough reason to stop the production and
request a maintenance activity. In most cases, when the equipment’s status deviates
from the nominal one, the operator evaluates the deviation and, based on it, decides
5 Predictive maintenance framework: Implementation of local and cloud processing for multi-
stage prediction of CNC machines’ health

whether the machine is still fit for its purpose. In most cases (like for mass produced
parts), the operator limits itself to warn the operators that will use the machine in the
next shift to monitor it closely. Only in presence of drastic deviations, the operator
decides to request a maintenance activity but even in those cases the production is not
stopped. Instead the processing parameters are modified (usually by lowering speeds
and loop gains) to continue to safely produce (albeit at a slower rate) while waiting
for a replacement/repair intervention. By implementing such a functionality directly
inside the machine’s CNC, it would be possible to directly exploit smart CNC
customization features. This functionality (such as FIDIA’s Look ahead) allows
modifying large sets of parameters with a single command. If specific sets of
parameters are created for specific machine tool’s health status and process
conditions, it then becomes a trivial task to correlate the output of the status evaluation
platform with an automatic process parameter adaption feature. However, the limits
of data driven analysis should be taken into account while using such a method for
health status evaluation. The most important one is that these algorithms are able only
to detect when the machine performs out of nominal boundaries, but cannot provide
any information on which is the cause of such behaviour. On the contrary, a model
based approach would allow to actually correlate the deviation from nominal
conditions to a specific restricted set of responsible components. This approach has
several benefits, since it would allow: (1) to adapt only parameters directly related to
the component’s behaviour, thus reducing the impact on the whole process; (2) to
store information about both the signals and the components that have failed, for
components’ models improvement; (3) to restrict the set of components that could be
malfunctioning, making easier and shorter the duration of the maintenance
intervention; (4) a wider time span between the prediction of the incoming failure and
insurgence, granting a better maintenance scheduling. It is worth noting that a model
based health detection status would greatly benefit from a cloud deployment. On the
one side, this would allow using data coming from whole families of identical but
widely spread similar machine tools (thus supporting also the machine tool design
phase). On the other side, this would limit the computational burden at the factory
level, relieving resources that could be better exploited for different objectives. In
order to allow a cloud based approach however, the CNC should allow functionalities
that go beyond the simple process management. These functionalities include the
availability of web based services for sending information to the cloud (relevant
sensor data for health status computation), retrieving data from the cloud (the
computed components’ status) and the writing of information on a database (for
successive analysis and model training). In the framework of the PROGRAMS
project, FIDIA will integrate its CNCs with the project’s platform and test the new
functionalities on the pilot line of one of the end users.
6 Predictive maintenance framework: Implementation of local and cloud processing for multi-
stage prediction of CNC machines’ health

1.6 Acknowledgments
Part of the work reported in this paper makes reference to the EC research project
“PROGRAMS – PROGnostics based Reliability Analysis for Maintenance
Scheduling” (www.programs-project.eu), which has received funding from the
European Union’s Horizon 2020 research and innovation programme under the Grant
Agreement No. 767287.

1.7 References
[MOU 15] Mourtzis D., Vlachou E., Milas N., Xanthopoulos N., "A cloud-based approach for
maintenance of machine tools and equipment based on shop-floor monitoring", Procedia
CIRP, 48th CIRP Conference on Manufacturing Systems, Vol. 41, 2015, p. 655-660.
[DHI 06] Dhilon BS., “Maintenability, Maintenance, and Reliability for Engineers.”, Taylor
and Francis, 2006.
[VEN 10] Venkataraman V., “Maintenance engineering and management.” PHI Learning
Private Limited 2010.
[AIV 17] Aivaliotis P., Georgoulias K., Chryssolouris G., "A RUL calculation approach based
on physical-based simulation models for predictive maintenance", International Conference
on Engineering Technology and Innovation (ICE IEEE), 2017.
[PAP 07] Papachatzakis P., Papakostas N., Chryssolouris G., “Condition based operational risk
assessment an innovative approach to improve fleet and aircraft operability: Maintenance
planning”, 1st European Air and Space Conference, Germany, 2007, p. 121-126.
[OKO 14] Okoh C., Roy R., Mehnen J., Redding L., “Overview of Remaining Useful Life
Prediction Techniques in Through-Life Engineering Services, Product Services Systems
and Value Creation”, Proceedings of the 6th CIRP Conference on Industrial Product-
Service Systems, 2014, p. 158 – 163.
[MYC 10] MyCar Deliverable D2.3.1 – D3.3.1 – D4.3.1 – D5.3.1 – Refinement and Industrial
Implementation.
[EFT 12] Efthymiou K., Papakostas N., Mourtzis D., Chryssolouris G., “On a Predictive
Maintenance Platform for Production Systems.”, Procedia CIRP, vol. 3, 2012, p. 221-226.
[SCH 05] Schwabacher M., “A Survey of Data-Driven Prognostics”, Infotech@Aerospace,
Infotech@Aerospace Conferences, 2005.
[BIT 14] Bittencourt A. C., Saarinen K., Sander-Tavallaey S., Gunnarsson S., Norrlöf M., “A
data-driven approach to diagnostics of repetitive processes in the distribution domain –
Applications to gearbox diagnostics in industrial robots and rotating machines”,
Mechatronics, vol. 24: 8, 2014, p. 1032-104.
[WAN 01] Wang P., Vachtsevanos G., “Fault prognostics using dynamic wavelet neural
networks”, Artificial Intelligence for Engineering Design and Manufacturing, vol. 15, 2001,
p. 349-365.