An approach to evaluate tactical decision-making in industrial

maintenance
N. Rodríguez-Padial, M. Marín, R. Domingo*
Abstract

This work comprises a part of a global and modular framework for Maintenance Decision Support Systems, whose general
objective is the proposal of a system that assists an expert in decision making for the design of maintenance program customized
in a productive plant. This system starts from the alignment with the strategic objectives of the company, after with the tactical
and operational maintenance. This paper deals the tactical goal, which defines a dynamic and graphic Balanced Scorecard to
extracts knowledge and prediction of the indicators in the medium term. The proposed custom Balanced Scorecard design
integrates the use of Pivot Tables and Pivot Charts in Microsoft Excel© with Machine Learning in Matlab©. We apply
"TopDown" strategies to address a very abundant data with Pivot Chart filters on a data. Artificial Neural Networks algorithms
are used in order to recognize patterns and to establish a correspondence between input and outputs variables.

Keywords: Pivot Table, Machine Learning, Clustering, Regression, Maintenance Balanced Scorecard"

1. Introduction

This work comprises a second part of a global and modular framework composed of three parts [1], Fig. 1, whose
general objective is the proposal of a system that assists an expert in decision making for the design of customized
maintenance plans to a productive plant, starting from the alignment of the strategic objectives of the company, those
tactical and operational maintenance, according to the personalized maintenance concept established by [2].

Fig. 1. Modular structure of the proposed global system decision support framework MDSS.

This BSC, thus designed allows, as well as the classic intermediate BSC, to measure the current situation and also
to predict its evolution considering historical data [3].

2. Methodology

The proposed custom Balance Scorecard design integrates an exploratory piece of data by using Pivot Charts in
Microsoft Excel©, and another part of discovery and prediction using Machine Learning, ML, in Matlab©, and it is
possible to embed the Matlab code as an add-in in Excel, so it is not necessary to install Matlab in the
user’scomputer. The methodology described in this Section is applied to maintenance data from papermaking
industry; due to the typical continuous process of this industry, the maintenance influence on the efficiency and
sustainability more than other type of plants [4-6] and the indicators can be used to know the overall equipment
effectiveness [7].

2.1 Exploratory phase using Pivot Charts in Excel

A Pivot Table is an interactive way to quickly summarize and analyze large amounts of numerical data. It is
especially designed to: query large amounts of data, obtain subtotals and sums of numerical data, summarize data by
categories and subcategories, and create calculations and customized formulas, expand and collapse data levels to
highlight results and drill down into summary details of areas of interest, move rows to columns or columns in rows
(or "pivot") to see different summaries of the data source, filter, sort, group and conditionally format more useful and
interesting subset of data, allowing you to focus on the information you want, submit concise, attractive, and
annotated online or print reports. Pivot Charts provide graphical representations of the data of their associated Pivot
Tables and thus are interactive, allowing you to sort and filter the underlying data of the pivot chart. Pivot Tables
and Pivot Charts allow you to summarize, analyze, explore and present data, tables and dynamic charts that can
easily see comparisons, patterns and trends, allowing you to make informed decisions about critical company data.
You can also connect to different external data sources, such as SQL Server tables and cubes, Office Data
Connection (ODC) files, XML files, Access databases, and text files to create Pivot Tables or use tables existing
dynamics to create new tables.
The PivotTable (and PivotChart) in Excel is the first part of the custom design of the proposed scorecard:
Microsoft Office software with Excel is often available, which offers the use of PivotTables and Pivot Chart in your
spreadsheets, is a very new improvement that allows to apply filters on a set of data for all its fields, enabling
"TopDown" strategies to approach a very abundant data set [3], as usually found in productive environments.

2.2 Phase of analysis through Machine Learning with Matlab

Machine Learning, ML, is a branch of artificial intelligence whose goal is to programmatically automate the
learning process on a computer, just as they naturally learn from humans and animals through experience [8,9]. The
algorithms of ML directly employ the data without previously establishing an equation as a model, in addition, these
algorithms improve their efficiency with the number of data used as examples during the learning.

Fig. 2. General Vision of Machine Learning Techniques. Adapted from [10]

Fig. 3. (a) Machine Learning Techniques; (b) Clustering. Adapted from [10]

ML finds natural patterns in the data and helps make better decisions and establish predictions, Machine Learning
is divided into two techniques: supervised learning, training a model on a known input and output data to predict
future outputs, and unsupervised learning, finding patterns hidden and intrinsic structures in the input data, Fig. 2.
For each technique, different algorithms can be used, Fig. 3a, where choosing the ideal is done by trial and error.
Supervised learning uses classification and regression techniques to develop predictive models. The difference
between them is that the classification predicts responses in discrete or categorical variables, whereas regression
predicts responses in continuous variable. Unsupervised learning uses the clustering technique commonly used in
exploratory data analysis to find hidden patterns as clusters in data. Fig. 3 b.
We will use algorithms modeled with Artificial Neural Networks (ANNs), for their versatility, both for
nonsupervised techniques (clustering) in order to group the input data in order to recognize patterns and define the
natural groups present in the data, as well as for techniques Supervised (regression), whose purpose is to establish a
correspondence or mapping between input values or predictors, and output variables or objectives to predict. For
this, the model is trained (or adjusted) by a knowledge base formed by historical examples of known inputs and
outputs.
In the preparatory phase of data, previous to the Machine Learning, in addition the statistical technique of
Principal Component Analysis (PCA) has been used to reduce the data dimension and to find the main axes that best
represent said variation of data. These axes are orthogonal to each other and are calculated by a linear base change
application by choosing a new coordinate system for the original data set in which the largest variance of the dataset
is captured on the first axis, called the first principal component, the second largest variance is the second axis, and
so on. It reduces to a problem of eigenvalues and eigenvectors on the covariance matrix of the data, obtaining a
reduction of the dimensionality of the data on those axes that contribute more to its variance, in general so many
principal axes are taken whose sum represents at least 80% of the variation of the original data.

3. Results

It is possible to design a customized BSC and adapted to the real and specific needs of the industrial plant. In the
first phase, exploratory, through the use of PivotCharts in Excel, you obtain a BSC that makes it possible in an easy,
intuitive, fast and interactive way, to visualize the general panorama of the plant and to expand in detail those
productive areas and sections of the plant to the level desired, in order to evaluate those areas of the plant that show
the user's greatest interest. The input data are collected in an Excel sheet with the work orders, and its associated
information (table of 46 variables and 1291 instances) of the entire industrial plant, for the whole year, Fig. 4a,
shows a Pivot Chart corresponding to general overview of the plant for all workshop interventions for a year,
presenting in the upper left-hand the filters of: urgency, date of request of work orders, month and year, in front of
sections of the plant numbered as: 11, 12, 20 and 22.

Fig. 4. (a) General outlook for industrial plant (number of interventions); (b) Outlook with filter data.

In the same way the interventions corresponding to the different intervening workshops coded as OR, TE, TM,
TT, fields that are also filterable are represented by color. To show the ease and interaction of the PivotTable, by
pressing the filter and selecting Month 2, Fig. 4b shows the adjustment of the PivotTable for the month of February.
This way of presenting the data by sections of the plant for the variables number of interventions and costs, allows a
quick overview as an exploratory analysis of the productive area analyzed, so that quickly they are evident those
values that disagree comparatively on the total, in addition to making it possible to reveal those sections where this
occurs.
This same type of graphics is made, recursively, for each section of the plant, allowing another filter for those
parts or locations more detailed, using the strategy "top-down" for an adaptive study, detailed and visual from the
general to the most relevant locations. Fig. 5a shows a PivotChart for the filtered section 11 for the month of
February, consistent with the graph of Fig. 4b, showing that the portions of the section 11 with the greatest number
of interventions are the installations 12 and 55, using an install filter are both emphasized in Fig. 5b.

Fig. 5. (a) PivotChart Section 11; (b) Section 11 with installations filter data.

In the second phase, of Machine Learning (ML), developed with Matlab and integrated in the same Excel
spreadsheet, also allows to know the inherent structure of the data, through analysis of principal components analysis
PCA, to discover the minimum dimensions that explain the variation of the data and the clustering technique to
discover patterns hidden in data, as the natural grouping of the same ones.

Fig. 6. (a) Principal components variance; (b) Projected data on principal components

Applying the PCA analysis on the set of work orders of productive area 2, AP2 (composed of sections 11 and 12)
as a result of the strategic module 1, three principal dimensions or axes are found are sufficient to explain the whole
99% variation of the original data contained in the work orders of the plant; the standard projections of the original
data and variables on the main axes can be seen in Fig. 6. Previously the data were normalized with mean 0, standard
deviation 1.
In this case, Fig. 7a, it is observed that for average Euclidean distances of 3000 to 6000 it presents two natural
groups, from 1500 to 3000, three groups, from 1000 to 1500, four groups, already below 1000 the number of groups
grow considerably, because of this compression, a value of 750 is taken, by pruning the tree into 6 natural groups,
the different groups of color data can be visualized in the Fig. 7b. The clustering technique is again used, using a
second algorithm of a self-organized map artificial neural network (SOM) on the sub-set of total cost data and repair
time, as chosen variables reflecting costs and times of the intervention maintenance of the plant. A self-organized
map (SOM) or Kohonen consists of a competitive layer that can classify a set of vector data with any number of
dimensions in as many classes as neurons have the layer [11]. Neurons are arranged in a two-dimensional topology
of the data set. The trained network with 2 variables and 1291 input data is shown in Fig. 7c.

Fig. 7. (a) Dendrogram; (b) data; (c) Artificial Neural Network SOM trained; (d) SOM topology sample hits.

The network is configured by 2 dimensions, 2x3, discovering a pattern of 6 natural groups in the data, distributed
with a clear linear relationship between them. Fig. 7d shows the 6 natural groups and the linear relationship between
the two variables (cost, time)."
Finally, the regression technique allows to predict the future values of the mean times between failures, MTBF
and repair, MTTR, simultaneously. We use a trained neural network with input data (predictors) that combines: a
variable time (month), ten production variables, and a maintenance variable (number of interventions), the two target
variables, output, are average times between failures MTBF and till repair MTTR, measured for the 12 months of the
year.
 Fig. 8. (a) Artificial neural network regression trained; (b) Performance; (c) Regression fit.

Fig. 8a depicts a feed-forward network with 10-layer hidden neurons and 2 layers of linear output neurons, can
arbitrarily fine-tune multidimensional mapping problems, given consistent data and sufficient neurons in its hidden
layer. The network will be trained with 70% of the data using the backpropagation algorithm of Bayesian
regularization, 15% of the data will be used for validation and the remaining 15% for testing. A mean square error
performance function, MSE, will be used, Fig. 8b. The result of the trained network is given by good according to
the regression adjustment coefficients, Fig. 8c. The result of the adjustment is observed by comparing the output
variables MTBF and MTTR, in blue, measured with those foreseen by the MTBF and MTTR model, in red, Fig. 9.
In short, this network can predict the behavior of the plant, for MTBF and MTTR time values, feeding the model
with predictive values of production and maintenance. As more input / output data are being input, the network will
be re-trained and more reliable, as it will be adjusted with more real examples, thus gaining more experience and
knowledge.

Fig. 9. Measured output variables (blue) vs predicted (red). (a) MTBF; (b) MTTR.

4. Conclusions

A customized scorecard model has been developed for tactical definition, allowing to control the maintenance of
an industrial plant, in the medium term, in such a way that it is possible to measure indicators of those productive
areas, resulting from the exit of the strategic module of the stage previous. The BSC environment is developed with
Pivot Charts in Excel, a graphical tool for an analyst user to perform quick, intuitive and interactive analysis of data
in a graphical rather than a numerical way. In addition, it is given to the BSC of Machine Learning in order to
discover structures and patterns of behaviour, relatively hidden, in work orders and production data. By grouping or
clustering technique, natural groups are determined on the variables cost and maintenance work, in addition to
making predictions about the mean MTBF failure times and MTTR repair. The scorecard model on a paper mill has
been validated.

References

[1] N. Rodríguez-Padial, M. Marín, R. Domingo. Procedia Eng, 132 (2015) 903-910.

[2] G. Waeyenbergh, L. Pintelon. Int. J. Prod. Econ. 77 (2002) 299-313.
[3] S.W. Palocsay, I.S. Markham, S.E. Markham. J. Bus. Res. 63 (2010) 191-206.
[4] L.M. Calvo, R. Domingo. J. Clean Prod. 96 (2015) 253-262. [5] L.M. Calvo, R. Domingo. Procedia Eng. 63 (2013) 678-686.
[6] L.M. Calvo, R. Domingo, Sustainability. 2017: 9, 514.
[7] R. Domingo, S. Aguado, Sustainability. 2015: 7, 9031-9047.
[8] I. Bose, R.K. Mahapatra, Inform Manage. 39 (2001) 211-225.
[9] G. Köksal, I. Batmaz, M.C. Testik, Expert Syst Appl. 38 (2011) 13448–13467.
[10] Mathworks. Introducing Machine Learning. https://www.mathworks.com [consulted in 2017].
[11] J. Hernández, M.J. Ramírez, C. Ferri. Introducción a la minería de datos. Spain. Pearson Prentice Hall, 2005.