Data-Driven Predictive Maintenance for Component Life-Cycle

Extension

Margarida Moreira

1 a

, Eliseu Pereira

1,2 b

and Gil Gonc¸alves

1,2 c

Faculty of Engineering, University of Porto, Portugal

SYSTEC-ARISE, Faculty of Engineering, University of Porto, Portugal

Keywords:

Predictive Maintenance, Industry 4.0, Remaining Useful Life (RUL), Data-Driven Methods, Survival

Analysis.

Abstract:

In the era of Industry 4.0, predictive maintenance is crucial for optimizing operational efﬁciency and reducing

downtime. Traditional maintenance strategies often cost more and are less reliable, making advanced pre-

dictive models appealing. This paper assesses the effectiveness of different survival analysis models, such as

Cox Proportional Hazards, Random Survival Forests (RSF), Gradient Boosting Survival Analysis (GBSA),

and Survival Support Vector Machines (FS-SVM), in predicting equipment failures. The models were tested

on datasets from Gorenje and Microsoft Azure, achieving C-index values on test data such as 0.792 on the

Cox Model, 0.601 using RSF, 0.579 using the GBSA model and 0.514 when using the FS-SVM model. These

results demonstrate the models’ potential for accurate failure prediction, with FS-SVM showing signiﬁcant

improvement in test data compared to its training performance. This study provides a comprehensive evalua-

tion of survival analysis methods in an industrial context and develops a user-friendly dashboard for real-time

maintenance decision-making. Integrating survival analysis into Industry 4.0 frameworks can signiﬁcantly en-

hance predictive maintenance strategies, paving the way for more efﬁcient and reliable industrial operations.

1 INTRODUCTION

In the industrial setting, maintaining system compo-

nents is a critical yet challenging aspect of operations

management. Traditional maintenance approaches,

often reactive or time-based, have limitations in terms

of cost efﬁciency and operational effectiveness. With

the advancement of Industry 4.0 development, the in-

tegration of digital technologies into manufacturing

processes has revolutionized maintenance strategies,

making predictive maintenance not just beneﬁcial but

essential. Predictive maintenance utilizes advanced

data analytics and machine learning to predict com-

ponent failures before they occur, allowing for timely

and cost-effective interventions.

Traditional maintenance methods, such as reactive

maintenance and preventive maintenance, often re-

sult in inefﬁciencies. Reactive maintenance can lead

to unexpected downtime and high repair costs, while

preventive maintenance may cause over-maintenance

https://orcid.org/0009-0000-7290-6467

https://orcid.org/0000-0003-3893-3845

https://orcid.org/0000-0001-7757-7308

and unnecessary expenses. Although there are some

attempts in this area (Rossini et al., 2021), there is a

critical need for predictive maintenance models that

can optimize maintenance schedules, reduce costs,

and improve equipment uptime.

This work focuses on implementing and evalu-

ating various predictive maintenance models using

survival analysis techniques. The methods include

Cox Proportional Hazards, Random Survival Forest

(RSF), Gradient Boosting Survival Analysis (GBSA),

and Survival Support Vector Machines (FS-SVM).

The data used for this analysis comes from two

sources: a dataset from Gorenje company related to

a spot welding robot and a Microsoft Azure predic-

tive maintenance dataset from Kaggle. It will show

the preprocessing of data ingested, the training and

evaluation of different survival analysis models and a

simulation of a maintenance scenario.

The work presented in this paper aims to provide a

comprehensive evaluation of predictive maintenance

models based on survival analysis and their integra-

tion within an Industry 4.0 framework.

This paper is organised as follows: Section 2

presents the literature review. Section 3 presents the

126

Moreira, M., Pereira, E. and Gonçalves, G.

Data-Driven Predictive Maintenance for Component Life-Cycle Extension.

DOI: 10.5220/0013014200003822

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Informatics in Control, Automation and Robotics (ICINCO 2024) - Volume 1, pages 126-136

ISBN: 978-989-758-717-7; ISSN: 2184-2809

system’s architecture implementation. Section 4 dis-

cusses the main results obtained after the implementa-

tion. Finally, Section 5 provides the conclusions and

future work that can still be done.

2 LITERATURE REVIEW

The following literature review summarises Industry

4.0 and the different types of industrial maintenance

that are used, the different predictive maintenance

models, survival analysis and its performance metrics

and in the end, the concept of data preparation and

feature engineering and ends with the presentation of

different data-driven methods. It concludes with a

summary of the primary ﬁndings from the reviewed

studies.

2.1 Types of Industrial Maintenance

Predictive Maintenance (PdM) has been increasing in

popularity in the last few years, mainly in Industry

4.0. It is estimated that ”The impact of maintenance

represents a total of 15 to 60% of the total costs of

operating of all manufacturing” (Zonta et al., 2020).

Nowadays, the three main types of industrial

maintenance are reactive maintenance, preventive

maintenance, and predictive maintenance.

Reactive maintenance is characterized by re-

sponding to equipment failures, in other words, with-

out prior planning. Preventive maintenance refers to

activities performed on a ﬁxed schedule, normally in

regular intervals. This technique reduces the likeli-

hood of equipment failure. However, it may lead to

over-maintenance, where components are often ﬁxed

more than necessary. This will cause an increase in

costs, which will cause the companies to lose more

than necessary.

Predictive maintenance relies on the real-time

monitoring of equipment conditions to predict when

maintenance should be performed. It uses a combi-

nation of data analysis, machine learning, and sen-

sor technologies to indicate potential failures before

they happen. The main goal is to optimize mainte-

nance frequency, reduce costs, and increase equip-

ment uptime. It predicts and prevents component fail-

ures. It has been widely adopted due to its superior

results compared to the previously mentioned tech-

niques (Moat and Coleman, 2021).

In the last years, a new kind of maintenance has

been developed. This type, called proactive main-

tenance, is considered the latest type in this evolu-

tion process. It consists of identifying the source

of the fault and remediating conditions that can fail

(Ramezani et al., 2023).

2.2 Predictive Maintenance Models

There are three main models, as presented below:

• Condition-Based Maintenance (CBM)

• Prognostics and Health Management (PHM)

• Remaining Useful Life (RUL)

The main focus of this article will be on the Re-

maining Useful Life (RUL) strategy. RUL refers to

the estimated lifespan of a device before it requires

maintenance or replacement and also the remaining

time until the equipment’s health conditions reach the

failure threshold (Gupta et al., 2024). The accuracy of

the RUL estimation depends on the quality and quan-

tity of available health monitoring data and the meth-

ods used for such analysis (Ferreira and Gonc¸alves,

2022).

Figure 1: Models for RUL prediction, based on (Achouch

et al., 2022).

RUL uses different models for its prediction-

making, as shown in ﬁgure 1.

2.3 Survival Analysis

Survival analysis encompasses various methods for

analyzing data related to time-to-event outcomes. It

is also comprehended as a reliability theory and an

event history analysis. The term ‘survival’ is equiva-

lent to probabilities in survival analysis, implying that

survival can be ‘measured’. The survival function can

be expressed as (Moat and Coleman, 2021):

S(t) = P(T > t) = 1 − F(t) (1)

Where S(t) is the survival function representing

the probability that a system or component survives

beyond a certain time t. The T is the number of cy-

cles until the component is worn out. The whole func-

tion describes the probability that the component is

not worn out by t cycles.

Data-Driven Predictive Maintenance for Component Life-Cycle Extension

127

The survival function present on 1 can be esti-

mated using the Kaplan-Meier Estimator (KM). It is a

non-parametric estimator used to measure the fraction

of living subjects after a certain amount of time with

respect to one or more speciﬁc events (Yang et al.,

2022). It is deﬁned as:

S(t) =

∏

≤t



1 −



(2)

In the equation 2 (Frumosu et al., 2020), the t

the time when at least one event (worn-out) occurred,

the d

is the total number of events (worn-out) at t

number of shots and n

is the number of components

at risk at t number of shots.

The Kaplan-Meier survival curve starts from 1

when all components work correctly and decreases to

0 as events (worn-outs) occur.

To evaluate and compare the performance of sur-

vival analysis models, the scikit-survival module from

Python can be used, which offers comprehensive tools

and metrics tailored for survival data. The bigger dif-

ference between survival analysis and traditional ma-

chine learning is the fact that parts of the training

data can only be partially observed. (P

olsterl, 2024).

Consequently, the performance metrics used are the

Concordance index (C-index), Brier Score, integrated

Brier score and Time-Dependent ROC.

2.4 Performance Metrics

Performance metrics are used to evaluate and estimate

the survival models. Some of the most common in-

clude the Concordance Index (C-index), Brier Score

and the integrated Brier score (IBS) (Zeng et al.,

2023).

The C-index, also called Harrell’s Index (Frumosu

et al., 2020), is one of the most common evaluation

metrics used for measuring the quality and efﬁciency

of survival models. It ranges from 0 to 1, where re-

sults closer to 1 indicate a better performance of the

model (De Santis et al., 2022).

The Brier score measures the difference between

the predicted probability and the true outcome, with

higher scores (Shen et al., 2024). The Brier score

ranges from 0 to 1, and contrarily to C-index, 0 in-

dicates a perfect prediction while 1 indicates poorer

prediction accuracy and calibration.

The IBS extends the Brier score to the context of

time-to-event data, such as survival analysis. It re-

ﬂects calibration over all time points, with a smaller

value indicating greater accuracy. Like the Brier

score, a smaller value indicates a greater accuracy

(Tran et al., 2023).

2.5 Data Preparation and Feature

Engineering

The dataset is usually chosen based on the data col-

lected by the sensors, more precisely, temperature,

voltage, pressure, votation, and vibration, among oth-

ers.

Normally, two different datasets are used for test-

ing, one real and another one already used in other

works, to compare the results.

The dataset that would be chosen to be compared

with the real one should contain a large number of

points in order to have more accurate results.

The ﬁrst step normally consists of the removal

of low-coverage parameters, replacing missing values

with the median in order to maintain data integrity,

removing non-numerical parameters, and eliminating

non-numeric features that are not useful for analysis.

Following that, the aggregation process begins by

summarizing and consolidating the data. Following

that, the feature selection is realized, which starts by

ﬁrst summarising and consolidating the data and then

proceeding with the selection of relevant features and

removal of redundant ones.

The use of multiple datasets allows for the valida-

tion of the models that are being tested and, conse-

quently, the generalization of the results. It also al-

lows us to compare the efﬁciency of predictive main-

tenance tools against established benchmarks. Fi-

nally, the comparative dataset can help interpret re-

sults from the real dataset and vice-versa.

When these processes have been concluded, some

data transformation is performed, where the normal-

ization of the features occurs.

To end this procedure, the data is normally split

into training and testing to posteriorly evaluate the

model’s performance.

Feature engineering is a technique that creates

new features based on the existing ones from the

dataset (Zhou et al., 2022). Some of the most com-

mon techniques are feature creation and feature ex-

traction. Feature creation involves generating new

ones by creating new variables used in the machine

learning models (Madasamy et al., 2023). Feature ex-

traction is based on extracting the most relevant fea-

tures from a raw data sample used as input to the

models. The most common domains to be used to

extract these variables are time, frequency and time-

frequency (Gupta et al., 2024).

2.6 Data-Driven Methods

To predict the behaviour of the machines and deter-

mine when an intervention should be conducted to re-

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

128

store some components, it is necessary to evaluate and

analyze the available data from the sensors.

There are three main types of data-driven meth-

ods, as shown below (Arena et al., 2021):

1. Statistical approaches;

2. Stochastic approaches;

3. Machine Learning Techniques.

The ﬁrst two methodologies are best suited for

examining complex systems whose evolutionary pro-

cesses are not straightforward to anticipate.

The third one is the approach that will be consid-

ered in this work.

This can be summarised in the following ﬁgure 2.

Figure 2: Types of data-driven methods and the fourth main

machine learning methodologies.

Regarding machine learning algorithms, they are

divided into three main categories, as shown in Figure

Figure 3: Machine learning algorithms divided into the

three main categories.

The ﬁrst technique involves using simpler algo-

rithms that are more easily interpreted.

The second one uses multilayered artiﬁcial neu-

ral networks to create a more suitable mapping func-

tion between given inputs and outputs. It normally

requires a large amount of data to achieve better ac-

curacy.

Ensemble learning methods involve combining

two or more ML algorithms to achieve better perfor-

mance than when using the individual algorithms sep-

arately. Instead of depending on a single model, the

predictions from the individual algorithms are com-

bined using a speciﬁc rule to produce a more accu-

rate single prediction. Ensemble methods are typi-

cally categorized into parallel (bagging) and sequen-

tial ensembles (boosting) (Mienye and Sun, 2022).

2.7 Gap Analysis

The reviewed studies present different methods and

models to predict the remaining useful life of different

components. Some of the most common approaches

are traditional machine learning techniques, and with

the development of technology, there are more studies

where deep learning methods are being developed.

Nevertheless, signiﬁcant gaps prevail when using

a survival approach. When searching for articles that

used this package in Industry 4.0, the search returned

null results on Scopus. However, this technique is

widely used in the ﬁeld of medicine. By leveraging

the strengths of survival analysis, industries could im-

prove their predictive maintenance strategies, leading

to more efﬁcient operations and reduced downtime.

3 IMPLEMENTATION

In this chapter, the implementation of the Data-Driven

Predictive Maintenance system is designed to extend

the life cycle of critical components. The implemen-

tation is structured into distinct modules, each re-

sponsible for a speciﬁc aspect of the data processing

and maintenance prediction workﬂow. This modular

approach ensures scalability, ﬂexibility, and ease of

maintenance.

3.1 Systems’ Architecture

The architecture of the system can be categorized into

the following modules, as shown in the Figure 4 be-

low:

Each module plays a critical role in the overall

functionality of the predictive maintenance system.

Figure 1 illustrates the interaction between these mod-

ules and their respective components.

Data-Driven Predictive Maintenance for Component Life-Cycle Extension

129

Figure 4: Systems’ Architecture Overview.

The data acquisition module is responsible for col-

lecting and ingesting raw data from various sensors

and data sources. This data forms the backbone of the

predictive maintenance system, providing the neces-

sary inputs for subsequent analysis and modelling.

Once the data is ingested, it undergoes prepro-

cessing to clean, normalize, and transform it into a

suitable format for analysis on the data preprocess-

ing module. This module includes feature engineer-

ing processes that extract meaningful features from

raw data, enhancing the accuracy and efﬁciency of the

predictive models.

In the predictive model module, machine learning

algorithms are employed to predict the degradation

of components. These predictions help prevent unex-

pected failures and extend the component life cycle.

The maintenance module interprets the predic-

tions and provides actionable insights for mainte-

nance planning. It generates dashboard reports that

aid in decision-making and execution of maintenance

actions. This module ensures that the right mainte-

nance activities are performed at the right time based

on data-driven insights.

For the validation of the solution, two datasets

were used: one provided by Gorenje company that

is related to component welding and another that

was taken from Kaggle, a Microsoft Azure predic-

tive maintenance dataset created for predictive main-

tenance model building, being a synthetic dataset.

For the development of the architecture, the open-

source scikit-learn package used version 1.3.2, the

scikit-survival package used version 0.22.2, the scipy

package version used version 1.10.1, and the Python

version 3.9.0.

3.2 Data Acquisition and Preprocessing

The data related to the company Gorenje were related

to a spot welding robot. It was presented in diverse

CSV ﬁles, with the minimum, mean and maximum

current from 2022-09-30 until 2023-03-08. It pre-

sented a total of 629444 samples. In another ﬁle, there

was information related to the values of a rotary ta-

ble, with a total of 96869 samples. There was also a

ﬁle that presented the time of failure when the failure

occurred, the error code number, and a total of 248

samples.

Based on these ﬁles, the ﬁrst step was to merge the

sensor data with the counter data to align timestamps

correctly, using the pd.concat function, incorporating

sensor readings and counter value.

After merging the data, the difference between

successive counter values was calculated in order to

determine the machine cycles. For each cycle, statis-

tics such as the mean and standard deviation of cur-

rent sensor readings were computed. Each cycle was

then labelled as either working or resulting in a fail-

ure based on historical failure data. This labelling was

used to train predictive models.

The Microsoft Azure predictive maintenance

dataset was designed for building predictive main-

tenance models and includes various operational

metrics and sensor readings, all collected hourly.

The dataset comprises several CSV ﬁles such as

PdM telemetry.csv that contain hourly averages of

voltage, rotation, pressure, and vibration from

100 machines in 2015, totalling 876,101 sam-

ples. PdM errors.csv includes 3,920 samples of

machine errors, detailing the datetime, machineID,

and speciﬁc error. These errors occur without

causing machine shutdowns. PdM maint.csv con-

tains 3,287 samples with the datetime, machineID,

and the component that was replaced. It records

both proactive maintenance during scheduled visits

and reactive maintenance after component failures.

PdM failures.csv is a subset of the maintenance data,

showing component replacements due to failures and

PdM machines.csv provides each machine’s model

type and age. Only telemetry data from 2015 was

used in the preprocessing phase.

3.3 Predictive Models

For both Gorenje and Microsoft Azure datasets, the

preprocessed data obtained in the previous module

was used to train various survival analysis models.

First, the preprocessed data was loaded, and then rel-

evant features, including cycle time and cumulative

metrics, were selected. The dataset was divided into

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

130

a training set and a test set in order to evaluate the

performance of each model and proceed to the vali-

dation of the system. The x train contained the fea-

tures of each dataset, while the y train included the

status of the components (if they have failed) and the

time to failure for each component. After dividing the

datasets, the ﬁrst step was the training phase. During

training, the model used x train (cycles) and y train

as inputs. After the model was trained, the outputs

produced were the different models and the associ-

ated c index train for each model. After completing

the training phase, the testing phase was conducted to

obtain the model predictions. It was used the x test

and y test as inputs, and the trained models produced

in the testing phase. The outputs obtained included

the predictions associated with each model (such as

the survival functions and the time to failure for the

Fast Survival Support Vector Machine), as well as the

c index test to evaluate the performance. The differ-

ent survival models that were trained and evaluated

included the Cox Proportional Hazards, Random Sur-

vival Forest, Gradient Boosting Survival Analysis and

Fast Survival Support Vector Machine.

For this, was created the training phase and, after

that, the testing phase. In the ﬁrst one, the system had

as inputs the x train. After this, the models were eval-

uated using the concordance index, and survival func-

tions were plotted to visualize the estimated survival

probabilities over time. This procedure was applied

to each component individually to create this module.

After this step, the models were stored on the disk

using the dump function to be used in the maintenance

module.

3.3.1 Cox Proportional-Hazards Model

The Cox proportional hazard model (PH) is a semi-

parametric model. It is used to investigate the effect

of several variables on the time a speciﬁed event takes

to happen, and it has the following hazard function

(Moat and Coleman, 2021):

h(t, X) = h

(t)exp(βX) (3)

In the equation 3, h

(t) is the baseline hazard

function, β is the coefﬁcient, and X is the covariate.

The Cox PH model assumes that survival times t are

independent, the hazard is proportional, i.e., the haz-

ard ratio is proportional, the hazard function is the

linear function of the numerical covariates, and the

values of X’s do not change over time.

The Algorithm 1 was implemented as follows:

3.3.2 Random Survival Forests

The Random Survival Forest is a collection of tree-

based models that ensures that each tree is built on

Data: Training data x

train

, y

train

, Test data x

test

Result: Trained Cox model, Feature

importance, Survival functions

Initialize: Cox Proportional Hazards model

cox model;

cox model ← CoxPHSurvivalAnalysis();

Call: TrainModel(cox model);

Print: “Feature importance:”;

Print: pd.Series(cox

model.coe f ,

index=x f eatures);

Algorithm 1: Cox Proportional Hazards Model Training.

a different bootstrap sample of the original training

data, consequently removing correlations between the

trees. At each node, only a randomly selected subset

of features and thresholds are used to evaluate the split

criterion. The ﬁnal predictions are made by combin-

ing the predictions of each individual tree in the en-

semble. The following algorithm 2 demonstrates the

implementation of this model.

Data: Training data x

train

, y

train

, Test data x

test

Result: Trained RSF model, Concordance

index

Initialize: Random Survival Forest model rs f ;

rs f ←

RandomSurvivalForest(n estimators=100,

min samples split=10, min samples leaf=15,

max features=”sqrt”, n jobs=-1,

random state=20);

Fit the model: rsf.ﬁt(x

train

, y

train

);

Evaluate the model using concordance

index;

cindex train ← rsf.score(x

train

, y

train

);

cindex test ← rsf.score(x

test

, y

test

);

Print: “ RSF Model ’;

Print: “cindex train: ”, round(cindex train,

3);

Print: “cindex test: ”, round(cindex test, 3);

Algorithm 2: Random Survival Forest Model Training.

3.3.3 Gradient Boosting Survival Analysis

Gradient Boosting leverages the principle of strength

in numbers by combining the predictions of multiple

base learners to create a robust overall model.

A gradient boosted model shares similarities with

a Random Survival Forest in that both rely on mul-

tiple base learners to generate an overall prediction.

However, they differ in their combination methods. A

Random Survival Forest ﬁts several Survival Trees in-

dependently and averages their predictions, whereas

a gradient boosted model is built sequentially in a

greedy stagewise manner. The Algorithm 3 demon-

strates the implementation of this model.

Data-Driven Predictive Maintenance for Component Life-Cycle Extension

131

Data: Training data x

train

, y

train

, Test data

test

, y

test

Result: Trained Gradient Boosting model,

Concordance index

Initialize: Gradient Boosting Survival

Analysis model gbsa;

gbsa ← GradientBoostingSurvivalAnaly-

sis(n

estimators=100, learning rate=0.01);

Fit the model: gbsa.ﬁt(x

train

, y

train

);

Evaluate the model using concordance

index;

cindex train ← gbsa.score(x

train

, y

train

);

cindex test ← gbsa.score(x

test

, y

test

);

Print: “Gradient Boosting Survival Analysis

Model”;

Print: “cindex train: ”, round(cindex train,

3);

Print: “cindex test: ”, round(cindex test, 3);

Algorithm 3: Gradient Boosting Survival Analysis Model

Training.

3.3.4 Survival Support Vector Machines

Survival Support Vector Machines (Survival SVM)

are an extension of traditional Support Vector Ma-

chines adapted for survival analysis. The Survival

SVM method integrates the principles of SVM into

survival analysis, aiming to ﬁnd a hyperplane that

maximizes the margin between different survival

times. Unlike traditional SVM, which is used for

classiﬁcation or regression, Survival SVM deals with

censored data, where the event of interest has not oc-

curred for all subjects during the observation period.

The following Algorithm 4 illustrates the implemen-

tation of this model.

3.4 Maintenance Module

This module was only applied to the second dataset

since this last one presented more features and data

samples.

Firstly, a simulation was created where the dif-

ferent models produced in each component would be

called using the load() function. In this simulation, a

machine class was created to simulate a machine with

multiple components that may fail and need replace-

ment. Inside this class, diverse methods were created,

such as:

• log event: that logs signiﬁcant events, such as

component failures and repairs, for debugging and

record-keeping, including timestamps and costs

associated with the maintenance of the compo-

nents;

Data: Training data x

train

, y

train

, Test data

test

, y

test

Result: Trained FS-SVM model,

Concordance index

Procedure: TrainSVM();

Initialize: Fast Survival SVM model

svm model;

svm

model ←

FastSurvivalSVM(rank ratio=0.0,

max iter=1000, tol=1e-5, random state=42);

Print: “FS-SVM Model”;

Fit the model: svm model.ﬁt(x

train

, y

train

);

Evaluate the model using concordance

index;

cindex train ←

concordance index censored(y

train

[

′

status

′

train

[

′

tt f comp1

′

-svm model.predict(x

train

));

Print: “cindex train: ”,

round(cindex train[0], 3);

cindex test ←

concordance index censored(y

test

[

′

status

′

test

[

′

tt f comp1

′

-svm model.predict(x

test

));

Print: “cindex test: ”, round(cindex test[0],

3);

Return: svm model, cindex train,

cindex test;

Algorithm 4: Fast Survival SVM Model Training.

• predict time to failure: where is speciﬁed the

model to predict the remaining useful life of a

component based on its features;

• run machine: It simulates the machine operation

and decrements the remaining life of components,

checks for failures, and starts repairs when neces-

sary. It also checks component lifetimes at the end

of each simulated day;

• log remaining times: It logs the remaining times

for all components at regular intervals for moni-

toring purposes;

• update remaining times: Regularly updates the

remaining times of components every 8 hours;

• start repair: Manages the repair process of a failed

component, including scheduling repairs within

working hours and handling repair costs;

• complete repair: Completes the repair of a com-

ponent and updates its predicted remaining life.

In this simulation, a cost of 40 C per repair of the

component was also added, and 40C was referred to

as the displacement of the repairmen. It was also

added some restrictions, such as the repayment only

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

132

does the maintenance of the components from Mon-

day to Friday, from 08:00 to 17:00. Also, the ﬁrst

component takes four hours to be repaired, and the

subsequent on that day takes only one hour.

After the simulation was performed, a dashboard

was created with the support of Streamlit. This Dash-

board ﬁrst opens a webpage where the people can se-

lect a date and time between 01/01/2015 06:00 and

01/01/2016 06:00. The person can also choose the

time of simulation he wants to produce. This time

is referred to as the weeks. After the person presses

the button of the simulation, a report with the total

cost associated with the repairs and displacement is

demonstrated, as well as the start and end date of the

simulation. It then shows a table with the repair log of

each component, the time it started being repaired and

the model that was used to predict the failure. It ends

with showing graphs, one with the remaining time of

each component during the weeks simulated and an-

other with the failures of each component.

4 EXPERIMENTS AND RESULTS

The results can be divided between the performance

of each model for each dataset and the dashboard cre-

ated with the help of the Streamlit application.

4.1 Predictive Models Performance

The ﬁrst results were obtained after the training of the

different models. The results were evaluated using the

C-Index for training and for test samples.

The initial ﬁndings were performed on Gorenje’s

dataset and can be observed in the table 1.

Table 1: Evaluation of the different models trained using

C-index performance measure for the Gorenje’s dataset.

Model Cox

Model

RSF GBSA FS-

SVM

C-index

train

0.763 0.897 0.789 0.722

C-index

test

0.974 0.791 0.789 0.865

The Cox Model was observed to perform prop-

erly on the training data (0.763), but it performed ex-

ceptionally well on the test data (0.974). This unusu-

ally high test c-index compared to the training c-index

might indicate potential overﬁtting. Generally, the

test performance is expected to be slightly lower or

comparable to the training performance.

RSF shows high performance on the training data

(0.897), suggesting it has learned it well. However,

its performance drops on the test data (0.791), main-

taining a good level of predictive accuracy on the test

data.

The GBSA model shows consistent performance

on both training and test datasets with identical c-

index values (0.789). This consistency suggests that

the model generalizes well and has neither overﬁtted

nor underﬁtted the data.

FS-SVM has the lowest performance on the train-

ing data (0.722) but performs signiﬁcantly better on

the test data (0.865). This improvement could imply

that the model is better suited to the distribution of the

test data.

When testing for Microsoft Azure’s dataset, the

performance was evaluated for each component of

machine 1. The results are represented in table 2.

Based on the results, it can be conﬁrmed that the

CoxPH Model presents a pattern of overﬁtting across

all components, with test c-index values signiﬁcantly

higher than training c-index values. On the other

hand, the RSF model represents a good generalization

on components one and four, but components two and

three present a big drop between the train and test c-

index, which may suggest an overﬁtting. The GBSA

model presents a consistent performance for compo-

nents one, two and four and component three shows a

moderate drop in test performance, suggesting some

overﬁtting. In the end, FS-SVM does present a con-

sistent performance for components three and four, al-

though the results are not that high.

4.1.1 Simulation and Dashboard Report

When simulating, based on the results of the c-index

values, the models for each component were chosen.

The models chosen were:

• model choice comp1 = ’rsf’

• model choice comp2 = ’gbsa’

• model choice comp3 = ’svm’

• model choice comp4 = ’svm’

These models could be changed in the simulation

script.

With the created interface, the person can choose

the date and the weeks they want to simulate, accord-

ing to the rules mentioned in section 3.4. For the fol-

lowing results, the data chosen was 2015/01/05 06:00,

and the time for simulation was 5 weeks. The ﬁrst

ones to acknowledge are the simulation report and the

repair log, present in the ﬁgure 5.

Data-Driven Predictive Maintenance for Component Life-Cycle Extension

133

Table 2: Evaluation of for the different models trained using C-index performance measure, for the Microsoft Azure’s dataset.

CoxPH RSF GBSA FS-SVM

C-index train C-index test C-index train C-index test C-index train C-index test C-index train C-index test

Comp1 0.531 0.926 0.709 0.689 0.627 0.645 0.510 0.553

Comp2 0.518 0.858 0.702 0.387 0.627 0.509 0.516 0.544

Comp3 0.480 0.649 0.751 0.458 0.722 0.543 0.511 0.496

Comp4 0.518 0.792 0.687 0.601 0.602 0.579 0.507 0.514

Figure 5: Simulation Report and Repair Log for the simu-

lated dates.

After the report, the following charts were pre-

sented. In ﬁgure 6, it can be seen on the line chart

the remaining time of each component. In ﬁgure 7, it

is presented with the help of a bar chart, the times that

a component failed during the simulated time.

Figure 6: Remaining time for each component.

Overall, the analysis of the predictive models’ per-

formance across different datasets highlights key in-

sights. The CoxPH model, despite its strong perfor-

mance on test data, exhibited signs of overﬁtting, ne-

Figure 7: Failures of each component during the simulated

time.

cessitating further validation and regularization tech-

niques. The RSF model demonstrated good general-

ization in some components but struggled with over-

ﬁtting in others, indicating the need for potential ad-

justments or alternative strategies for certain data sub-

sets. GBSA consistently delivered reliable results

across most components, conﬁrming its robustness

and generalizability. Finally, FS-SVM showed varied

performance, excelling in some test scenarios while

underperforming in others, suggesting it might bene-

ﬁt from further reﬁnement or hybrid approaches.

In the simulation phase, the models selected for

each component, based on their c-index values, pro-

vided a solid foundation for generating accurate pre-

dictions and facilitating the creation of a user-friendly

interface for real-time decision-making. This compre-

hensive approach not only validated the chosen mod-

els but also highlighted areas for further enhancement.

5 CONCLUSIONS AND FUTURE

WORK

The application of survival analysis methods for pre-

dictive maintenance within the framework of Indus-

try 4.0 has demonstrated signiﬁcant potential and ef-

fectiveness. This document evaluated several pre-

dictive models, including the Cox Proportional Haz-

ards model, Random Survival Forests (RSF), Gradi-

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

134

ent Boosting Survival Analysis (GBSA), and Survival

Support Vector Machines (FS-SVM), on datasets

from both Gorenje and Microsoft Azure.

The results reveal that survival methods can be

highly effective for predicting maintenance needs in

an industrial setting. As it was seen, the Cox model

showed strong performance on test data, although it

exhibited overﬁtting, which necessitates further inves-

tigation and validation. The RSF model displayed

good generalization in some components but also

signs of overﬁtting in others, indicating the need for

model adjustments. The GBSA model consistently

performed well across different datasets, suggesting

it is a robust choice for generalizable predictive main-

tenance tasks. Meanwhile, FS-SVM showed poten-

tial, especially in speciﬁc test scenarios, though it may

beneﬁt from further reﬁnement.

A key advantage of using these survival analy-

sis models is the ability to generate Remaining Use-

ful Life (RUL) estimates, which are strongly impor-

tant for planning maintenance activities. The simula-

tion and dashboard interface developed using Stream-

lit provided a practical tool for visualizing these pre-

dictions, allowing maintenance schedules to be opti-

mized to minimize costs and downtime. The inter-

face facilitated real-time decision-making by enabling

users to simulate various scenarios and view detailed

reports on component health and predicted failures.

Despite the promising results, there are several ar-

eas for future work to enhance the application of pre-

dictive maintenance, which can be highlighted:

• Real-Time Data Integration: Implementing mod-

els that can handle real-time data streams will im-

prove the timeliness and accuracy of maintenance

predictions. This integration will also allow for

dynamic updating of RUL estimates and more re-

sponsive maintenance planning.

• Enhanced User Interface: Developing a more

complex and detailed user interface could provide

additional insights and functionality, such as more

granular control over simulation parameters and

detailed visualizations of component health trends

over time.

• Advanced Model Reﬁnement: Further reﬁnement

of models, particularly those showing signs of

overﬁtting or underperformance, could involve

hybrid approaches that combine the strengths of

multiple models or the integration of additional

data sources to improve prediction accuracy. The

use of other performance measures, such as the

Brier Score mentioned in section 2.4, can help to

compare the results obtained in the future.

• Scalability Testing: Testing the models and simu-

lation framework in larger, more diverse industrial

settings will ensure the scalability and robustness

of the solutions proposed.

In conclusion, this study emphasises the potential

of survival analysis methods in improving predictive

maintenance within Industry 4.0 environments. By

using advanced predictive models and user-friendly

interfaces, industries can move towards more efﬁ-

cient, cost-effective, and reliable maintenance strate-

gies. Continued research and development in this

ﬁeld will further solidify these methods’ role in the

next generation of intelligent and efﬁcient mainte-

nance.

ACKNOWLEDGEMENTS

This work was partially supported by the

HORIZONCL4-2021-TWIN-TRANSITION-01

openZDM project, under Grant Agreement No.

101058673.

REFERENCES

Achouch, M., Dimitrova, M., Ziane, K., Sattarpanah Kar-

ganroudi, S., Dhouib, R., Ibrahim, H., and Adda, M.

(2022). On Predictive Maintenance in Industry 4.0:

Overview, Models, and Challenges. Applied Sciences,

12(16):8081.

Arena, F., Collotta, M., Luca, L., Ruggieri, M., and Ter-

mine, F. G. (2021). Predictive Maintenance in the Au-

tomotive Sector: A Literature Review. Mathematical

and Computational Applications, 27(1):2.

De Santis, R. B., Gontijo, T. S., and Costa, M. A. (2022).

A Data-Driven Framework for Small Hydroelectric

Plant Prognosis Using Tsfresh and Machine Learning

Survival Models. Sensors, 23(1):12.

Ferreira, C. and Gonc¸alves, G. (2022). Remaining Use-

ful Life prediction and challenges: A literature review

on the use of Machine Learning Methods. Journal of

Manufacturing Systems, 63:550–562.

Frumosu, F. D., Rønsch, G. Ø., and Kulahci, M. (2020).

Mould wear-out prediction in the plastic injec-

tion moulding industry: a case study. Interna-

tional Journal of Computer Integrated Manufactur-

ing, 33(12):1245–1258.

Gupta, S., Kumar, A., and Maiti, J. (2024). A critical review

on system architecture, techniques, trends and chal-

lenges in intelligent predictive maintenance. Safety

Science, 177:106590.

Madasamy, S., Shankar, B. P., Yadav, R. K., and P, J. K.

(2023). A Machine Learning Approach in Predic-

tive Maintenance in the IoT Enabled Industry 4.0. In

2023 4th International Conference on Smart Electron-

ics and Communication (ICOSEC), pages 418–423,

Trichy, India. IEEE.

Data-Driven Predictive Maintenance for Component Life-Cycle Extension

135

Mienye, I. D. and Sun, Y. (2022). A Survey of Ensemble

Learning: Concepts, Algorithms, Applications, and

Prospects. IEEE Access, 10:99129–99149.

Moat, G. and Coleman, S. (2021). Survival Analysis and

Predictive Maintenance Models for non-sensored As-

sets in Facilities Management. In 2021 IEEE Inter-

national Conference on Big Data (Big Data), pages

4026–4034, Orlando, FL, USA. IEEE.

olsterl, S. (2024). scikit-survival: A Python package for

time-to-event analysis. Accessed: 2024-07-18.

Ramezani, S. B., Cummins, L., Killen, B., Carley, R., Amir-

latiﬁ, A., Rahimi, S., Seale, M., and Bian, L. (2023).

Scalability, Explainability and Performance of Data-

Driven Algorithms in Predicting the Remaining Use-

ful Life: A Comprehensive Review. IEEE Access,

11:41741–41769.

Rossini, R., Prato, G., Conzon, D., Pastrone, C., Pereira,

E., Reis, J., Goncalves, G., Henriques, D., Santi-

ago, A. R., and Ferreira, A. (2021). AI environment

for predictive maintenance in a manufacturing sce-

nario. In 2021 26th IEEE International Conference

on Emerging Technologies and Factory Automation

(ETFA ), pages 1–8, Vasteras, Sweden. IEEE.

Shen, J., Yang, D., Zhou, Y., Pei, J., Wu, Z., Wang, X.,

Zhao, K., and Ding, Y. (2024). Development of

machine learning models for patients in the high in-

trahepatic cholangiocarcinoma incidence age group.

BMC Geriatrics, 24(1):553.

Tran, T. T., Lee, J., Gunathilake, M., Kim, J., Kim, S.-

Y., Cho, H., and Kim, J. (2023). A comparison of

machine learning models and Cox proportional haz-

ards models regarding their ability to predict the risk

of gastrointestinal cancer based on metabolic syn-

drome and its components. Frontiers in Oncology,

13:1049787.

Yang, Z., Kanniainen, J., Krogerus, T., and Emmert-Streib,

F. (2022). Prognostic modeling of predictive mainte-

nance with survival analysis for mobile work equip-

ment. Scientiﬁc Reports, 12(1):8529.

Zeng, C., Huang, J., Wang, H., Xie, J., and Zhang, Y.

(2023). Deep Bayesian survival analysis of rail use-

ful lifetime. Engineering Structures, 295:116822.

Zhou, B., Pychynski, T., Reischl, M., Kharlamov, E., and

Mikut, R. (2022). Machine learning with domain

knowledge for predictive quality monitoring in resis-

tance spot welding. Journal of Intelligent Manufac-

turing, 33(4):1139–1163.

Zonta, T., da Costa, C., da Rosa Righi, R., de Lima, M.,

da Trindade, E., and Li, G. (2020). Predictive mainte-

nance in the Industry 4.0: A systematic literature re-

view. Computers and Industrial Engineering, 150.

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

136