Real-Time Equipment Health Monitoring Using Unsupervised Learning

Techniques

Nadeem Iftikhar

and Finn Ebertsen Nordbjerg

Centre for Industrial Digital Transformation, University College of Northern Denmark, Aalborg 9200, Denmark

Keywords:

Equipment Health Indication, Real-Time Monitoring, Sensor Data, Unsupervised Learning, Anomaly

Detection.

Abstract:

Reducing unplanned downtime requires monitoring of equipment health. This may not be possible in many

cases as traditional health monitoring systems often rely on the use of historical data and maintenance infor-

mation which is not always available, especially for small and medium-sized enterprises. This paper presents

a practical approach that uses sensor data for real-time equipment health indication. The methodology pro-

posed consists of a set of steps. It starts with feature engineering which may include feature extraction to

transform raw sensor data into a format more suitable for analysis. Anomaly detection follows next, where

various techniques are employed to ﬁnd any deviations in the engineered features indicating potential equip-

ment deterioration or abrupt failures. Then comes the most important stages equipment health indication and

alert generation. These stages provide timely information about the equipment’s condition and any necessary

interventions. These steps make it possible for such an approach to be effective even when there is little or no

historical data available. The applicability of this approach is validated through a lab-based case study.

1 INTRODUCTION

Monitoring machinery and systems’ health is very

important through the equipment health indication

(EHI). This enables early detection of issues based on

real-time data, even when there are no predictive ca-

pabilities to forecast faults in future. The proposed

approach, therefore, improves safety and reliability

in production operations by reducing time for equip-

ment repair and maintenance expenses. Typical meth-

ods employed for EHI usually requires lots of histori-

cal data, maintenance records, particular sensor types

and expert advice for technical purposes. This can

be a challenge to SMEs that might not have enough

resources or expertise. This paper presents a contem-

porary approach of performing EHI using real-time

data, which does eliminate the need for extensive his-

torical data as well as maintenance details. The pro-

posed methodology is divided into two main stages:

unsupervised feature engineering that involves feature

extraction; then proceeded by unsupervised anomaly

detection. In cases where there’s limited historical

data available (with or without failure data), the en-

tire process of feature engineering which includes fea-

https://orcid.org/0000-0003-4872-8546

ture extraction can be used. However, without histor-

ical data, the strategy is limited to feature engineering

only with exclusion of the dimensionality reduction

aspect involved in feature extraction. The main con-

tributions of this paper can be summarized as follows:

• The paper discusses how unsupervised learning

techniques can be used to analyze sensor data in

order to identify complex patterns and anomalies

that cannot be easily recognized without prede-

ﬁned labels, benchmarks, or failure data. This

could enable SMEs to adopt early issue detection

approaches thus bettering the maintenance strate-

gies they use.

• A case study demonstrates the efﬁciency of this

methodology. Speciﬁcally, this case study illus-

trates how adaptable the methodology is when tra-

ditional equipment health monitoring techniques

are lacking as a result of which it has potential for

optimizing operational effectiveness.

The rest of this paper is organized as follows: Sec-

tion 2 presents an overview of the research question.

Section 3 reviews relevant literature. Section 4 ex-

plains the adopted methodology. Section 5 provides

the implementation details. Section 6 evaluates the

methodology through a case study and discusses im-

Iftikhar, N. and Nordbjerg, F.

Real-Time Equipment Health Monitoring Using Unsupervised Learning Techniques.

DOI: 10.5220/0012785500003756

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

In Proceedings of the 13th International Conference on Data Science, Technology and Applications (DATA 2024), pages 401-408

ISBN: 978-989-758-707-8; ISSN: 2184-285X

401

plications for ﬁndings. Lastly, section 7 concludes the

paper and suggests future research directions.

2 RESEARCH PROBLEM

This research is focused on the development of an

EHI system that employs real-time sensor data for

monitoring industrial equipment. To avoid costly

breakdowns and ensure that equipment functions op-

timally, there is a need to come up with effective con-

dition monitoring techniques for detecting anomalies

in the machines’ performance. In most cases, historic

and maintaince data is used by traditional predictive

maintenance systems. Nonetheless, such data may be

limited or absent for many SMEs. Moreover, these

predictive maintenance systems may not capture dy-

namic behavior and complexity of industrial equip-

ment thus leading to false alarms.

This paper therefore introduces an unsupervised

learning approach, which can analyze sensor data

at real-time, extract/create relevant features and spot

anomalies. The suggested approach could be effec-

tive in various situations. It could prove particularly

useful when there is limited data, a lack of histori-

cal data, or even when there are no records of any

failures. It makes this approach particularly useful

for SMEs, who often lack resources and expertise to

implement smart analytics solutions. Hence, the aim

of this approach to achieve optimal equipment per-

formance without depending too much on extensive

historical data or failure incidents.

3 BACKGROUND AND RELATED

WORK

Major progresses have been achieved in real-time

equipment condition monitoring especially with the

integration of unsupervised learning techniques. This

was illustrated in a holistic study adopting such data-

driven techniques as feature extraction, deep learning,

novelty detection and cluster analysis as presented

in the (Eltouny et al., 2023). Furthermore, a thor-

ough study of 71 anomaly detection algorithms for

the time series category on 976 datasets conducted by

(Schmidl et al., 2022) delivered the necessary marks

for technique choice and best practices. Unsuper-

vised methods for detecting concept drift in machine

learning were reviewed together with a taxonomy for

these methods, and their importance in ﬁeld scenarios

where immediate class labels are not available was

shown by (Gemaque et al., 2020).

In the anomaly detection research context, a liter-

ature review is outlined regarding the trend of unsu-

pervised learning from 2000 to 2020, as presented by

(Nassif et al., 2021). The superiority of Local Out-

lier Factor (LOF) and One-Class Support Vector Ma-

chines (OCSVM), reported by (Qasim et al., 2022),

was one of the main focuses of an unsupervised

anomaly detection algorithms comparison for predic-

tive maintenance in SMEs. A real-time anomaly de-

tection scheme for industrial automation through IoT

and unsupervised learning was proposed by (Gul-

tekin and Aktas, 2023). A novel infrastructure mon-

itoring method using a hybrid semi-supervised ap-

proach combining Convolutional Autoencoder (CAE)

and One-Class Classiﬁcation (OCC) was described

by (Saeedi and Giusti, 2022). An improved Autoen-

coder based method for unsupervised anomaly detec-

tion was presented by (Cheng et al., 2021).

Regarding the health monitoring, an industrial

machine health prediction system was proposed,

which was based on unsupervised learning and time

series decomposition to compute the health index us-

ing sensor data as described by (de Lima et al., 2021).

An unsupervised learning method based on Convolu-

tional Autoencoder (CAE) for machine health assess-

ment was presented by (Guo et al., 2022). Further, an

unsupervised structural health monitoring approach

that utilizes Autoencoder and Hidden Markov Model

(HMM) was outlined by (Coraca et al., 2023). The

role of data processing and machine learning model

selection were examined in the context of condition

monitoring in industry was revealed by (Surucu et al.,

2023). Machine learning based approach for real-time

monitoring and fault detection in industrial compo-

nents was introduced by (Yang et al., 2019). An Long

Short-Term Memory (LSTM) model for fault detec-

tion and health management of military vehicles was

proposed by (Shukla et al., 2021). A Convolutional

Neural Network (CNN) based model that can estimate

the remaining useful life of machinery was introduced

by (Wen et al., 2023). Moreover, a predictive main-

tenance framework for Industry 4.0 which employs

machine learning for anomaly detection, demonstrat-

ing high recall levels across various scenarios, was in-

troduced by (Morselli et al., 2021).

To expand on the previous research which points

out the importance of advanced analytics in managing

industrial equipment data, this paper demonstrates a

practical data-driven method for SMEs. The use of

this method that combines real-time sensor data and

unsupervised learning techniques, is well placed in

situations where SMEs may have no historical data,

limited historical data, or only operational data with

no fault data available. This approach connects the

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

402

monitoring of equipment health to the improvement

of operational performance, reducing downtime and

maximizing overall efﬁciency. In addition, it has been

presented in the paper that this proposed method can

be implemented easily as well.

4 METHODOLOGY

The methodology for real-time EHI encompasses sev-

eral steps, each of which is elaborated upon subse-

quently. The process (refer to Fig. 1) commences

with raw sensor data, which is subjected to prepro-

cessing. This is followed by an unsupervised fea-

ture engineering step that generates features from the

cleansed data. As part of this step, there exists an op-

tional process for reducing dimensionality to further

reﬁne these features. The features are then utilized in

unsupervised anomaly detection and to compute the

EHI score. Ultimately, based on the EHI score, alerts

are generated if the scores surpass critical thresholds.

Sensor Data

Data

Preprocessing

Unsupervised Feature

Engineering

Unsupervised Anomaly

Detection & EHI Score

Alert

clean

(

)

(

)

engineer

(

)

Alert

(

alert

(

EHI

(

)

, Θ

warning

,Θ

critical

)

Figure 1: Distinct components of the equipment health in-

dicator (EHI) system. The input for this proposed solution

comprises a dataset containing raw sensor data, typically

in the form of time series, while the outputs include EHI

scores and a variety of alerts.

4.1 Data Preprocessing

The ﬁrst step is raw sensor data preparation which is

represented by D. This data passes through a clean-

ing function that can be written mathematically as fol-

lows:

′

= f

clean

(D) (1)

This equation represents that the cleaned data, D

′

results from applying the cleaning function denoted

as f

clean

to raw sensor data, D. The cleaning func-

tion, f

clean

, usually involves several steps. There exist

noise reduction approaches used in eliminating distor-

tions caused by additive noise in true signal. Sensor

data may have different measurement scales. Normal-

ization can thus be performed on all data to ensure

uniformity but at the same time maintaining original

differences in value range. Missing values may occur

in the sensor data due to malfunctioning of sensors

or errors while transmitting data. Various methods

are used for missing value handling like imputation

where missing values get replaced with substitutes

values. Furthermore, outliers can distort real under-

lying patterns in the information contained within a

dataset being analyzed for any kind of research pur-

pose. Thus, determination and treatment of such out-

liers are accomplished through various outlier detec-

tion methods for the relevant analysis purposes.

4.2 Unsupervised Feature Engineering

This process transforms the cleaned data, denoted as

′

into F, a set of features that capture the dynamic

nature of the equipment. This transformation can be

stated mathematically as:

F = f

engineer

′

) (2)

Here, f

engineer

denotes the function of feature en-

gineering that is applied to D

′

. Feature engineer-

ing is a very important part of this process since it

extracts useful information from sensor data. The

goal is to change preprocessed data such that it be-

comes easier for analysis, thus exposing hidden pat-

terns. One common method used in feature engineer-

ing involves windowing whereby continuous streams

of data are divided into discrete windows for which

features can be extracted after they have been com-

puted and stored as window features. This helps to

capture temporal relationships between various nu-

merical values recorded over time on one hand and

others taken at different intervals on the other hand.

The function for feature engineering can incorporate

multiple methods. Examples include statistical mea-

sures like mean, median among others which sum-

marize the data distribution; frequency domain fea-

Real-Time Equipment Health Monitoring Using Unsupervised Learning Techniques

403

tures can also be created showing periodic compo-

nents within the dataset. The option to use feature en-

gineering or extraction techniques depend on whether

there is enough data. If there is limited historical data,

both methods can be used. Feature engineering will

convert the preprocessed data into a more useful form

while feature extraction will help in identifying which

features are most important for anomaly detection.

Unsupervised learning techniques could help a lot in

this regard, particularly for feature extraction. Such

techniques may include clustering, dimensionality re-

duction strategies such as Principal Component Anal-

ysis (PCA) or autoencoders based approaches among

others. In presence of no historical data at all, feature

engineering might still be used to convert any new in-

coming data in a format, suitable for further analysis.

4.3 Unsupervised Anomaly Detection

and Equipment Health Indication

Score Calculation

The process of unsupervised anomaly detection and

the calculation of an EHI score is the main element of

equipment condition monitoring. This process locates

such data points that do not follow the usual pattern

of behavior by the system, without the need for prior

labels or knowledge. This often occurs in cases where

a sudden failure or degradation of equipment may not

have occurred or been documented previously. The

anomaly score of every data point is computed, in

which case the corresponding input feature x

in the

feature set F is computed as the degree of difference.

Each higher score suggests the point is considered as

more unusual and the probability is less likely to be

equal to the normal.

An array of machine learning models can be used

for this task. Predominantly, each model has its par-

ticular strengths and some scenarios where they can

be utilized effectively. There are cases where histor-

ical data is absent, unsupervised anomaly detection

could be carried out with Isolation Forest, being just

one of the many potential algorithms. This algorithm

aims to detect the unusual nature of abnormal data

rather than common patterns. With this feature in

place, the model is well-equipped to face data with

many complex dimensions and also provide a great

amount of efﬁciency. In the case when there is lim-

ited historical data which includes both the normal

and the fault data, the Autoencoders, which is a type

of artiﬁcial neural network, can be used for anomaly

detection. During the training using a normal and

fault data set, the autoencoder with be able to adopt

some of its weights and ﬁlters to learn and reconstruct

the frequencies parts of it and mostly, it will recon-

struct the normal operating conditions with greater

accuracy as compared to the least frequent patterns

that may between faults or anomalies. After training,

reconstruction error is computed that is just the dif-

ference between the input and the output generated

by the algorithm. It is called anomaly score. The

point data that have greater reconstruction errors has

high probability to be anomalous. In cases in which

only a limited amount of historical data is available

and fault data is missing, One-Class Support Vector

Machine (OCSVM) can be utilized. This model will

learn a boundary around the normal data, and any

data point beyond this boundary will be detected as

an anomaly by the model. This makes it a powerful

tool for anomaly detection especially when there is a

small dataset with only good or normal data.

The EHI score at a given time point t is then ob-

tained through the use of the actual scores. This task

can be implemented via two detection methods: an

aggregate or an individual detection method. The

aggregate detection method ﬁnds already aggregated

anomaly scores for each data point, represented as:

EHI(t) = f

detect.aggregate

, F) (3)

Conversely, the individual detection method ﬁnds

individual anomaly scores for each data point and

then aggregates them, represented as:

EHI(t) = f

aggregate

( f

detect.individual

, F)) (4)

This EHI score is considered to be a health con-

dition of the equipment, and the lower the score, the

greater the chance of a problem.

4.4 Real-Time Alert Generation

The ﬁnal stage of the process is the generation of real-

time alerts mechanism. These alarms are triggered

based on the EHI scores reaching the corresponding

threshold levels. The predetermined threshold values

are set in advance, denoted by θ

warning

and θ

critical

order to classify the equipment state as normal state,

warning state and critical state. There is an issuance

of alerts that are based on those standard thresholds.

The alert generation function is represented mathe-

matically as:

Alert(t) =











Normal if EHI(t) ≤ θ

warning

Warning if θ

warning

< EHI(t) ≤ θ

critical

Critical if EHI(t) > θ

critical

(5)

In this equation, ‘Normal’ signiﬁes a normal op-

erational state, ‘Warning’ indicates a warning level

alert, and ‘Critical’ denotes a critical level alert. The

function Alert(t) generates an alert at time t based on

the EHI score and the predeﬁned thresholds θ

warning

and θ

critical

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

404

5 IMPLEMENTATION

The proposed methodology has been effectively im-

plemented, comprising several stages: data col-

lecting, preprocessing, storing, extracting features,

anomaly detection, alerting system setup and real-

time monitoring dashboard creation. Real-time data

has been continuously collected from the sensors and

transmitted. A light-weight, scalable and efﬁcient

communication protocol (Message Queuing Teleme-

try Transport (MQTT)) is used for data transmission

from the sensors to a central server. After, the sen-

sor data is pre-processed, which is done using the

Python scripts, it is then stored in a highly efﬁciently

and ﬂexible relational database, PostgreSQL. A data

processing web steaming mechanism has been built

using Node-RED, a effective ﬂow-based development

platform to ease the integration and ﬂow in the data

processing pipeline. Such a script start off by us-

ing Python’s native libraries like NumPy and Pandas

and then exploring for example, Isolation Forest algo-

rithm in Scikit-learn’s collections. Node-RED is ad-

ditionally used as a tool to create the alert system. The

last thing that has been employed is Grafana, which

basically creates dashboards that are interactive in real

time that show sensor data, deviation results and EHI

scores and provide an immediate picture of the condi-

tion of the system.

6 EXPERIMENTAL RESULTS

This section describes the outcomes derived from the

deployment of a real-time monitoring system, speciﬁ-

cally designed to evaluate the performance and health

status of robotic arms (refer to Fig. 2). The analyt-

ics procedure encompasses data pre-processing, engi-

neered features to improve data representation, unsu-

pervised anomaly detection methodologies and alert

generation in real time based on EHI scores.

Figure 2: Robotic arms to run lab experimental assembly

line.

6.1 Data Understanding

Data Understanding is the initial step of this research.

The data used was sensor data from two different op-

erating states of robotic arms: “gradual failure” and

“sudden failure”. The fact that the proposed moni-

toring systems has not encountered this type of data

before is rather noteworthy. It is this states’ data that

is believed to be streamed in a near real-time, which

is one of the key assumptions that depict an opera-

tional environment in which the monitoring system is

deployed without a historical data or failure records

to support it. This case is one of the challenges

that SMEs usually confront. The research method-

ology employed here is developed to evaluate the ef-

ﬁciency of the proposed system against the scenarios

where the conventional predictive maintenance mod-

els based on the historical data may not be applicable.

This method, thus, reinforces the effectiveness and re-

liability of unsupervised learning approaches. These

techniques can analyze time-series or incoming sen-

sor data to ﬁnd unusual data patterns and deviations

which indicate equipment health issues at the moment

without the need for pre-processed training data.

Figure 3: Subdivided plot displays each of the six sensors

in the gradual failure scenario separately, with each subplot

dedicated to one sensor.

The datasets are given here with an intention that

they would simulate the type of data that the system

might face in the real-world settings. The gradual

failure dataset (refer to Fig. 3). These variations, par-

ticularly in the vibration, temperature and power con-

sumption sensors after the 700th data point, can serve

as indicators of approaching equipment failure and

must not be neglected. The demonstration of these

gradual alterations intensiﬁes the signiﬁcance of prior

monitoring to avoid a continuous deterioration. On

the other hand, the sudden failure dataset (refer to

Fig. 4) is a situation where the system must rapidly re-

spond to changes that are a sign of instant equipment

failure. The dataset is comprised of the suddenly in-

Real-Time Equipment Health Monitoring Using Unsupervised Learning Techniques

405

Figure 4: This visualization presents each of the six sensors

for the sudden failure scenario, with each subplot dedicated

to one sensor.

troduced fault, for example, sudden drops in power

consumption and torque in a short span of time and

elevated values for vibration and temperature read-

ings together with the optical quality readings, espe-

cially evident between the 500th and 700th data point.

These datasets, after generating the failures synthet-

ically, will act as a base for validation of the given

technique.

6.2 Data Preparation

The ﬁrst step of the analysis is the data preparation

phase, carried out on the incoming data. It is a pivotal

part of the real-time operation because it enhances

the data quality due to the fact that data comes in a

steady, real-time stream. It includes the operations

like noise reduction, ﬁlling of missing values, removal

of outliers and so on. After the calibration process,

the sensor data was adjusted to a uniform scale using

normalization procedure. Normalization empowers a

comprehensive approach to the processing of sensor

records through data homogenization. In this manner,

data is prepared for the following analytical tasks.

6.3 Feature Engineering

This phase has been perhaps rather crucial in turning

cleaned sensor data into a format that has been more

representative of all the dynamic conditions the mon-

itored equipment has been going through. This paper

use aggregation of the sensor data to determine the

mean value for every operational condition, like grad-

ual failure or sudden failure. On the other hand, other

statistical measures may be employed depending on

the analysis needs. The rolling windowing technique

was used as a tool for effective feature engineering,

and the data was segmented into blocks of 50 obser-

vations each. Inside each window, the average of the

sensor readings were extracted. This step is the key

for deﬁning and modeling all temporal patterns sig-

nifying the abnormal operation or normal functioning

of the equipment. Disclaimer- in this case, feature

extraction was not applied for the lack of historical

data. This unsupervised feature engineering process

formed a solid basis for subsequent anomaly detection

and equipment health assessment, thus demonstrating

its potential use in real-time monitoring.

6.4 Anomaly Detection

The paper implemented Isolation Forest, a kind of un-

supervised anomaly detection algorithm. This model

is very applicable for the cases in which the histori-

cal data is either limited or nonexistent. It operates

by detecting anomalies in high-dimensional datasets,

which usually occur in sensor data for the equipment

health monitoring. These anomalies, efﬁciently pin-

pointed by the model (refer to Fig. 5), could be an

early warning or alarm suggesting faulty equipment

functions and promptly the corrective actions. The

models’ output associated each observation with a

normal (1) or anomalous (-1) status created the foun-

dation for the development of EHI scores. EHI score

represents a generated number from the results of

anomaly detection and provides a numerical measure

of equipment status. The visualization of EHI scores

over time for both gradual failure and sudden fail-

ure scenarios provided insightful views on equipment

health (refer to Fig. 6). For the case of gradual de-

cline, the EHI scores showed considerable succession

periods of stability were followed by visible health

drops. In the sudden failure case, the scores of EHI

that were stable earlier showed a noticeable drop that

is usually related to the multiple abnormalities occurs

within a short span of time. In general, EHI scores

could be used in early failures identiﬁcation, therefore

would permit fast response to the potential equipment

problems and minimizing of the downtime.

6.5 Real-Time Alert

This phase permits proactive maintenance interven-

tions and helps to convert the insights obtained from

EHI scores to actions. The paper featured visualiza-

tion techniques that enabled straightforward presen-

tation of equipment’s performance over a given time

period, as shown as condition status: good condition,

gradual failure condition and sudden failure condi-

tion, each indicating an EHI score with corresponding

real-time alerts. Visual indicators, represented by yel-

low and red vertical lines, represent the warning and

critical alert zones, respectively (refer to Fig. 6). The

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

406

Figure 5: These visualizations highlight where the isolation forest algorithm detected anomalies within each dataset based on

the combined mean of all sensor features, distinguished by color: gradual failure in blue and sudden failure in red.

Figure 6: The visualizations display the EHI scores and the corresponding real-time alerts together for each condition, sepa-

rately.

zones deﬁned by predeﬁned threshold levels signal

that immediate action is required. In the case where

everything works as planned, the device is demon-

strated by stable EHI scores, for example, from 200th

to 500th data point. Furthermore, the gradual fail-

ure condition (refer to Fig. 6 - upper) displays ﬂuctu-

ations in the EHI scores, indicating presence of drops

where the scores have slipped into the danger or crit-

ical zones between 700th and 1000th marks. These

alerts are the transitional ones which tell about the

current state so that maintenance activities could be

taken immediately to maintain smooth functioning.

With respect to the sudden failure condition (refer to

Fig. 6 - lower) where EHI scores suddenly drop from

the 500th data point to lower levels, generating criti-

cal alert and suggesting catastrophic failure.

6.6 Discussion

The methodology suggested in this paper viewed the

dynamics of operation and the conditions of equip-

ment for good health also under different operational

circumstances. The transformation of sensor data

into a structured format by applying the feature en-

gineering process was the next step in the process,

and it captured the dynamic states of the equipment.

The anomaly detection algorithms highlighted devia-

tions from standard working patterns, thereby signal-

ing possible equipment health problems. EHI score

assessed the equipment health as a qualitative mea-

sure over time. The combination of EHI and real

time alerts can give a holistic view of the equipment’s

health status. This strategy could provide a rea-

sonable mechanism for maintaining workplace efﬁ-

ciency. Similarly, this real-time alerting system could

Real-Time Equipment Health Monitoring Using Unsupervised Learning Techniques

407

be used as the core of the maintenance team’s work as

the team members can receive notiﬁcations about the

possible issues in a timely manner, so it can guarantee

greater efﬁciency and reduce production delays.

7 CONCLUSIONS AND FUTURE

WORKS

The paper proposes a practical application of real-

time equipment health monitoring. The approach

eliminates the need for extensive historical data and

maintenance records, which provides a considerably

advantageous avenue for the SMEs. The core con-

tributions of the research involve the implementation

of an unsupervised learning system and a real-time

notiﬁcation system. This framework incorporates un-

supervised learning algorithms, which helps analyze

sensor data and highlight any abnormalities, which is

useful for implementing efﬁcient machine monitoring

system. The real time alert system ensures the equip-

ment reliability and durability and so it leads to the

further improvement of the operation efﬁciency. It is

worth mentioning that this approach is not only bene-

ﬁcial for SMEs but also simple to implement, making

it a practical solution for real-time equipment health

monitoring.

The future task will be devoted to increasing the

system’s capability to handle various IoT devices.

Different unsupervised learning algorithms shall be

tested to ﬁnd out those best performing ones for

anomaly detection. Furthermore, a variety of feature

engineering techniques will also be studied in order

to further improve the performance. The adaptability

and scalability of the system through the use of real

data, comprising real failures instances, will be tested

in real production environments.

REFERENCES

Cheng, Z., Wang, S., Zhang, P., Wang, S., Liu, X., and Zhu,

E. (2021). Improved autoencoder for unsupervised

anomaly detection. International Journal of Intelli-

gent Systems, 36(12):7103–7125.

Coraca, E. M., Ferreira, J. V., and Nobrega, E. G. O. (2023).

An unsupervised structural health monitoring frame-

work based on variational autoencoders and hidden

markov models. Reliability Engineering and System

Safety, 231:109025.

de Lima, M. J., Crovato, C. D. P., Mejia, R. I. G.,

da R. Righi, R., de O. Ramos, G., da Costa, C. A.,

and Pesenti, G. (2021). Healthmon: An approach for

monitoring machines degradation using time-series

decomposition, clustering, and metaheuristics. Com-

puters and Industrial Engineering, 162:107709.

Eltouny, K., Gomaa, M., and Liang, X. (2023). Unsu-

pervised learning methods for data-driven vibration-

based structural health monitoring: A review. Sensors,

23(6):3290.

Gemaque, R. N., Costa, A. F. J., Giusti, R., and Santos,

E. M. D. (2020). An overview of unsupervised drift

detection methods. Wiley Interdisciplinary Reviews:

Data Mining and Knowledge Discovery, 10(6):e1381.

Gultekin, E. and Aktas, M. S. (2023). Real-time anomaly

detection business process for industrial equipment

using internet of things and unsupervised machine

learning algorithms. In International Conference on

Computational Science and Its Applications, pages

16–31. Springer Nature Switzerland.

Guo, L., Yu, Y., Duan, A., Gao, H., and Zhang, J. (2022).

An unsupervised feature learning based health indica-

tor construction method for performance assessment

of machines. Mechanical Systems and Signal Process-

ing, 167:108573.

Morselli, F., Bedogni, L., Mirani, U., Fantoni, M., and

Galasso, S. (2021). Anomaly detection and classiﬁca-

tion in predictive maintenance tasks with zero initial

training. IoT, 2(4):590–609.

Nassif, A. B., Talib, M. A., Nasir, Q., and Dakalbab, F. M.

(2021). Machine learning for anomaly detection: A

systematic review. IEEE Access, 9:78658–78700.

Qasim, M., Khan, M., Mehmood, W., Sobieczky, F., Pich-

ler, M., and Moser, B. (2022). A comparative analy-

sis of anomaly detection methods for predictive main-

tenance in sme. In International Conference on

Database and Expert Systems Applications, pages 22–

31. Springer International Publishing.

Saeedi, J. and Giusti, A. (2022). Anomaly detection for

industrial inspection using convolutional autoencoder

and deep feature-based one-class classiﬁcation. In In-

ternational Conference on Computer Vision, Theory

and Applications, pages 85–96.

Schmidl, S., Wenig, P., and Papenbrock, T. (2022).

Anomaly detection in time series: a comprehensive

evaluation. Proceedings of the VLDB Endowment,

15(9):1779–1797.

Shukla, I., Silas, A., Dozier, H., Hansen, B. E., and Bond,

W. G. (2021). Data driven hybrid approach for health

monitoring and fault detection in military ground ve-

hicles. In International Conference on Data Science,

Technology and Applications, pages 300–307.

Surucu, O., Gadsden, S. A., and Yawney, J. (2023). Con-

dition monitoring using machine learning: A review

of theory, applications, and recent advances. Expert

Systems with Applications, 221:119738.

Wen, L., Su, S., Wang, B., Ge, J., Gao, L., and Lin, K.

(2023). A new multi-sensor fusion with hybrid con-

volutional neural network with wiener model for re-

maining useful life estimation. Engineering Applica-

tions of Artiﬁcial Intelligence, 126:106934.

Yang, C., Liu, J., Zeng, Y., and Xie, G. (2019). Real-

time condition monitoring and fault detection of com-

ponents based on machine-learning reconstruction

model. Renewable Energy, 133:433–441.

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

408