Available online at www.sciencedirect.com
ScienceDirect Procedia CIRP 22 (2014) 181 – 185
3rd International Conference on Through-life Engineering Services
Context Driven Remaining Useful Life Estimation Carl-Anders Johansson1*, Victor Simon1, Diego Galar1 1
Luleå University of Technology, Luleå, 971 87, Sweden
* Carl-Anders Johansson, E-mail address:
[email protected]
Abstract In the context of maintenance activities maintainers rely on machine information, their past breakdowns, adequate repair methods and guidelines as well as new research results in the area. They usually get access to information and knowledge by using information systems (nondestructive testing (NDT) or condition monitoring.), local databases, e-resources or traditional print media. Basically it can be assumed that, the amount of available information affects the quality of maintenance decision making and acting positively. Machine health information retrieval is the application of information retrieval concepts and techniques to the operation and maintenance domain. Retrieving Contextual information, describing the operational conditions for the machine, is a subarea of information retrieval that incorporates context features in the search process towards its improvement.. Both areas have been gaining interest from the research community in order to perform more accurate prognostics according to specific scenarios and happening circumstances. Context is a broad term and in this paper the operational conditions and the way the machine has been used is seen as the context and is represented by operational data collected over time. This paper intends to investigate the effects of the interaction of context features on machine tools health information. This interaction between context and health assessment is bidirectional in the sense that health information seeking behavior can also be used to predict context features that can be used, without disturbing the operational environment and creating production disruptions. The extraction of multiple features from multiple sensors, already deployed in this type of machinery, may constitute snapshots of the current health of certain machine components. The mutation status (the way they have changed) of these snapshots, hereafter called Fingerprints, has been proposed as prognostic marker in machine tools problems. Of them, in this work so far only the spindle fingerprint mutation has been validated independently as prognostic for overall survival and survival after relapse, while the prognostic value of rest of components mutation is still under validation. In this scenario, the prognostic value of spindle fingerprint mutations can be investigated in various contexts defined by stratifications of the machine population.
© by Published Elsevier B.V. This is an open access article under the CC BY-NC-ND license ©2014 2014 Published The Authors. by Elsevier B.V. Peer-review under responsibility of the Programme Chair of EPSRC Centre for Innovative Manufacturing in Through-life Engineering (http://creativecommons.org/licenses/by-nc-nd/3.0/). Services. under responsibility of the Programme Chair of the 3rd InternationalThrough-life Engineering Conference Peer-review Keywords: fingerprint;operational data;remaining useful life;RUL;condition based maintenance;CBM;context driven prognostic;CDP
1. Introduction To know and predict the condition of an asset is of most value. During the last 40 years a lot of diagnostic techniques has been developed for this, many of them based on different signal analysis and statistical methods. For stationary operational conditions even the prognostic of the failure works well but the methods used often requires costly installation of transducers and signal analysis equipment handled by special skilled personal.
For some time it can be seen an increasing interest in methods based on available operational data without the need of costly equipment and special skilled personal. There are a number of commercial products in this area, the ARTESIS system is one of them (www.artesis.com). The methods described in this paper are similar to those used in other systems but are more focused also on other context data than those you can get from current, voltage and vibration signals.
2212-8271 © 2014 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the Programme Chair of the 3rd InternationalThrough-life Engineering Conference doi:10.1016/j.procir.2014.07.129
182
Carl-Anders Johansson et al. / Procedia CIRP 22 (2014) 181 – 185 Most interesting is this for those who have a fleet of similar equipment where methods and experience can be reused. This is for instance the situation for producers and owners of windmills, airplanes, ships, cars & trucks, elevators, packing machines and machine tools. This paper is part of the Power consumption driven Reliability, Operation and Maintenance optimisation (Power OM) project (http://www.power-om.eu/) which at this moment is in the middle of the data collection phase of operational data and fingerprint data. The paper is therefore mostly conceptual and will concentrate on possible techniques and methods to know and predict the condition of machine tools and fleet of machine tools, especially problems/faults in the spindle drive train and linear axis. The proposed method for Condition Based Maintenance (CBM), based on fingerprint and operational data, gives information about both operational conditions and power consumption without increasing the complexity and can be seen as a Green Condition Based Maintenance platform (Green CBM) [1] for both CMB and energy optimization. In this paper we will concentrate on CBM and RUL estimation. The platform, based on the Knowledge and Advanced Services for E-Maintenance (KASEM) platform from PREDICT, uses techniques for Context Driven Prognostic (CDP) for Remaining Useful Life (RUL) estimation based on available fingerprint and operational data, considered as context data [2]. The aim is to use the energy consumption monitoring and profiling and the operational conditions to improve the overall business effectiveness, under a triple perspective: x Maintenance optimization: optimizing maintenance strategies based on the prediction of potential failures and guiding the planning of maintenance operations to schedule maintenance operations in convenient periods and avoid unexpected equipment failures x Energy optimization: managing the energy as a production resource and reduce its consumption and cost x Machine reliability: providing the machine tool builder with real data about the behavior of their product and its critical components By integrating all the information from machines, fleet of machines and even between different companies the CBM platform can act as a hub of technology federating methods and analysis tools in order to provide the different user profiles (machine tool users, maintenance service providers, machine tool manufacturers) with a unified framework delivering business processes targeting: x Maintenance through fleet-wide predictive maintenance strategy deployment including: ¾ Support diagnostic process with fleet-wide comparison ¾ Associated monitored data with component operation condition in a larger scale ¾ Associated monitored data with component health for a regular update of the prediction models used locally ¾ Data-context of alert and predictive diagnostic ¾ Past diagnostic and past solution for similar abnormal situation
Optimise maintenance strategy under cost-effective parameters considering all the cost factors for planned maintenance operations, predictive inspections, machine breakdowns time, cost to repair, etc. x Operation through fleet-wide performance assessment by means of services allowing: ¾ Fleet-scale Key Performance Indicators (KPI) aggregation ¾ Machine fleet relative performance (e.g. point out weakest machine with respect to the fleet) ¾ Fleet energy efficiency assessment, based on consumption patterns. For this the context will be relevant, as the operational conditions have a decisive influence ¾ Fleet energy optimization, sharing cause-effect relationships for abnormal consumption x Product reliability through a closed loop with machine engineering and design by providing continuous feedback on: ¾ Machine operation health condition ¾ Estimation of Reliability, Availability, Maintainability and Safety (RAMS) parameters at the machine level and at the main components level: Spindle, linear guides, etc. ¾ Failure root causes discovery ¾ The use of data mining tools to support the analysis process of consolidation, aggregation and synthesis ¾
Figure 1. Machine tool (source Siemens)
2. Data collecting and handling 2.1. Input data for CDP The proposed method uses two types of input data, fingerprint data and operational data. 2.1.1. Fingerprint data Fingerprint data are collected in a standardized way every day/week/month under a test procedure. This standardized procedure implies a routine, a special CNC-program, that the machine runs every time the fingerprint is collected. There is also a possibility to use standard sequences in ordinary production programs like tool changing for part of the
183
Carl-Anders Johansson et al. / Procedia CIRP 22 (2014) 181 – 185 fingerprint collection. The data are collected with a sampling frequency between 100 Hz and 50 kHz, depending of type of data. Typical data collected and synchronized in time are: x Vibration data x Motor power for spindle and linear axis (current signal and motor current signature analysis) x RPM and speed for spindle and linear axis and axis position The data can be analyzed in both time domain and frequency domain [3] and a number of features (table 1) can be calculated for each signal.
Normally the spindle rotates clock wise. In some operation like threading and some milling it operates counter clock wise. The counter clock wise operation is normally with less power/torque. Therefore, comparing the difference in feature values in different rotational directions gives valuable information.
Table 1. Time domain features
Feature
Definition
Peak Value
ͳ ܲ ݒൌ ሾሺݔ ሻ െ ሺݔ ሻሿ ʹ
Root Mean Square
ͳ ܴ ܵܯൌ ඩ ሺݔ ሻଶ ݊ ୀଵ
Figure 3. Spindle head in test bench
Standard Deviation
ͳ ܵ ݀ݐൌ ඩ ሺݔ െ ݔҧ ሻଶ ݊
In table 2 a spindle head before and after repair is tested in both rotational directions, and where it can be seen that the degradation was bigger in the ‘normal operation’ direction (clock wise direction).
Table 2. Motor Current Analysis of the Spindle Motor showing the difference in behaviour for a spindle head rotating in different directions. CCW is the normal rotating direction
ୀଵ
Kurtosis Value
భ
1.
Crest Factor
2.
Clearance Factor
3.
Impulse Factor
4.
Shape Factor
5.
Normal Negative Likelihood value
6.
ݒܭൌ
ర σ సభሺ௫ ି௫ҧ ሻ
ோெௌ ర
݂ݎܥൌ ܲݒൗܴܵܯ ݂݈ܥൌ
௩ భ
మ ሺ σ సభ ඥȁ௫ ȁሻ
݂݉ܫൌ ݄݂ܵ ൌ
௩
భ σ ȁ௫ ȁ సభ
ோெௌ
Feature
భ σ ȁ௫ ȁ సభ
ܰܰ ܮൌ െ݈݊ܮǢ ܮൌ ςே ୀଵ ݂ሺݔ ǡ Ɋǡ ߪሻ
In the frequency domain the vibration level on known frequencies like gear mesh frequencies, bearing frequencies, rotational speed etc. and their harmonics will be followed. For faults/problems in the gear train, like bearing and gear problem, the most sensitive features has been chosen through test in a test bench (see figure 2) where different type of fault can be simulated and through test on faulty components sent in by customer for repair (see figure 3).
Figure 2. Test bench for gear train
Crest Factor
Rotational Frequency sideband
Value
Rotational direction
Spindle Status
1,96
Counter clockwise
Repaired
2,24
Counter clockwise
Faulty
1,96
Clockwise
Repaired
1,96
Clockwise
Faulty
3,56
Counter clockwise
Repaired
12,7
Counter clockwise
Faulty
4,73
Clockwise
Repaired
4,64
Clockwise
Faulty
For the linear axis there can be problem both with the drive train (motor/gearbox/ball screw/nut/rack/pinion) and the linear bearings (see figure 4).
Figure 4. Linear guides ball screw/nut
184
Carl-Anders Johansson et al. / Procedia CIRP 22 (2014) 181 – 185 For the drive train the machine tool axis dynamics is an important factor which can be analyzed by looking at the position/speed/acceleration and jerk values and by comparing differences in commanded and actual position [4]. With high resolution power measurements or vibration measurements on the linear axis it is possible to indicate problems in ball screw/nut, rack/pinion, hydraulic counter balance system, linear bearings [5]. 2.1.2. Operational data Operational data are collected with a sampling frequency between 1-100 Hz. The data are collected via interfacing with the Computer Numerical Control (CNC) controller of the Machine Tool (see figure 5). In the case of the Power-OM project, the research toolbox GEM OA (Open Architecture) hardware from Artis will be used for the data collection [6].
Figure 5. Local data collection unit
Typical data collected are: Spindle power and rpm Motor power and position for linear axis Difference between commanded and actual position Temperatures Program number Tool number Alarms (sampled from the CNC or taken from the log file) The operational data describes the way the machine has been used between the fingerprints, in this paper considered as the context. x x x x x x x
From the operational data a number of KPI can be calculated, some examples are: x Number of start/stop/acceleration/retardation x Total travelled length for linear axis and its distribution over the axis or ‘travelled load’ calculated from power need during acceleration of the axis x Mean power/torque for spindle and axis and its distribution over the axis x Difference in behavior in difference rotational directions for both spindle and linear axis x Running time in different rpm, direction and power intervals x Number of alarms per type/group 2.2. Remaining Useful Life estimation For each machine/component in the fleet typical faults/problems are identified and for each of them the most
sensitive fingerprint features and KPI’s are chosen. This means that for each machine/component there is a number of faults, and for each of them, a number of features and KPI’s to trace in a multi-dimensional space to estimate the RUL. The threshold for the estimation of the RUL is set based on result from test with known faults in test bench, test of faulty components (so far only for repaired and faulty spindle head) and experience from this or similar machines in the fleet. As a start value in the beginning the estimation can be based on the history of the machine tool like: x Age of the component/machine tool x Designed life time for the component/machine tool x Type of production/use (8h/24h/7d, heavy, medium, low) x Maintenance history x Experience from similar machines in the fleet But after a while the estimation is more and more based on results from fingerprint and operational data. The change of the different features value between two fingerprints indicates the degradation of the component and depends on the initial previous condition and the way the machine has been used, described by the operational data. This means that the future condition, the feature value Fn, is a function of previous condition value Fn-1 and the coming operational data. ܨ ൌ ݂ሺܨିଵ ǡ ܱܽݐܽ݀ ݈ܽ݊݅ݐܽݎ݁ሻ
(1)
Modern data mining algorithms can be used to extract the knowledge from the fingerprint and operational data and find the correlation between the change in condition and the way the machine tool is used. The result from this is used to estimate the RUL for this machine/component and its uncertainty. An overview of different techniques for RUL estimation can be seen in [7]. The prediction of remaining useful life in engineering systems is affected by several sources of uncertainty, and it is important to correctly interpret this uncertainty in order to facilitate meaningful decision-making. Thus, it is clear that the uncertainty of RUL depends on: uncertainty in collected data and feature calculations, uncertainty in prediction algorithms, uncertainty in future operating conditions and uncertainty in threshold setting [8]. Sometimes, the probability distribution of RUL may be extremely skewed and this can vary with the distance to ‘end of life’. By classification of Operational data into groups, for instance based on power and speed, prediction can be made for each group and if also the program number is classified into the same groups the user can have an estimation of the RUL depending on what products/program number they plan to run (different scenarios). In other words, the historical data gives information about how the degradation of the condition of the machine or component (change in fingerprint values) depends on how the machine has been used (represented by the operational data, the context). This also means that knowing the planned future use of the machine (operational data) makes it possible to predict the RUL for different scenarios. See figure 6.
Carl-Anders Johansson et al. / Procedia CIRP 22 (2014) 181 – 185 fingerprints. By use of data mining knowledge extraction techniques the correlation between operational data and change in fingerprint can be found. This information is used in the prediction models for RUL. Acknowledgements The research has received funding from the European Community´s Framework Programme FP7-2012-NMP-ICTFoF. Work programme: FoF.NMP.2012-2 “Methodologies and tools for the sustainable, predictive maintenance of production equipment” under grant agreement no. “314548 - PowerOM”. The Consortium consists of TEKNIKER, ARTIS, FAGOR AOTEK, PREDICT, MONITION, GORATU and Luleå University of Technology (LTU). Figure 6. Uncertainty in RUL estimation
For this estimation several self-learning algorithms can be used and the prediction hopefully will get better and better with more data, see figure 7.
References [1]
[2]
[3]
[4]
Figure 7. RUL algorithms
3. Conclusion [5]
This paper is a conceptual paper based on work in the Power OM project. In the final solution the proposed system will be integrated with the KASEM e-maintenance platform and therefore have the possibility to work on fleet, machine and component level. A method for CDP for RUL estimation based on available operational data and fingerprint is proposed. The method is based on available data with a minimum of new sensors. Selflearning algorithms can estimate both the RUL and the uncertainty for the estimation. Prediction can be done based on either historical production mixes or on planned production for the nearest future. The fingerprint represents the condition/status of the machine and the change in condition represents the degradation of the machine/component. The operational data represents the way the machine has been used between the
[6]
[7]
[8]
Johansson, C.A., Galar, D., Villarejo, R., Monnin, M. (2013). Green Condition based Maintenance - an integrated system approach for health assessment and energy optimization of manufacturing machines. The Tenth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies, 18-20 June 2013, Krakow, Poland. Prado, A., Alzaga, A., Konde, E., Medina-Oliva, G., Monnin, M., Johansson, C.A., Galar, D., Euhus, D., Burrows, M. & Yurre, C. (2014). Health and Performances Machine Tool Monitoring Architecture. The 3rd International Workshop and Congress on eMaintenance, 17-18 June 2014, Luleå, Sweden. Galar, D., Kumar, U., Lee, J. & Zhao, W. (2012). Remaining Useful Life Estimation using Time Trajectory Tracking and Support Vector Machines. 25th International Congress on Condition Monitoring and Diagnostic Engineering, Conference Series 364 (2012) 012063, 18-20 June 2012, Huddersfield, United Kingdom. doi: 10.1088/1742-6596/364/1/012063 Santiago-Pérez, J.J., Osornio-Rios, R.A., & Romero-Troncoso, R.J., Herrera-Ruiz, G. & Delgado-Rosas, M. (2008). DSP algorithm for the extraction of dynamics parameters in CNC machine tool servomechanisms from an optical incremental encoder. International Journal of Machine Tools & Manufacture 48 (2008) (1318– 1334). Huang, B., Gao, H., Mingheng, X., Wu, X., Zhao, M. & Guo, L. (2010). Life Prediction of CNC Linear Rolling Guide Based on DFNN Performance Degradation Model. Seventh International Conference on Fuzzy Systems and Knowledge Discovery (FSKD 2010), 10-12 August 2010, Yantai, China. Alzaga, A., Konde, E., Bravo, I., Arana R., Prado A., Yurre C., Monnin M., Medina-Oliva G. (2014). New technologies to optimize reliability, operation and maintenance in the use of Machine-Tools. Euro-Maintenance conference, 5-8 May 2014, Helsinki, Finland. Butler, S., (2012). Prognostic Algorithms for Condition Monitoring and Remaining Useful Life Estimation. Doctoral dissertation. NUI Maynooth, Ireland. Sankararaman, S. & Goebel, K. (2013). Why is the Remaining Useful Life Prediction Uncertain?. Annual Conference of the Prognostics and Health Management Society 2013, 14-17 October 2013, New Orleans, Texas, USA.
185
