International Journal of Engineering Insights: (2024) Vol. 2, Nro. 1, Regular Paper

https://doi.org/10.61961/injei.v2i1.18

Early Fault Detection in Paper Machine Motors Using

Machine Learning

Cristian P. Chuchico · Oscar Acosta Agudelo

Received: 18 Feb 2024 / Accepted: 02 May 2024 / Published: 16 May 2024

Abstract: This research addresses the application of

a neural network as a tool for early fault detection in

the motors of a paper machine under a simulated en-

various speed setpoints are configured, and the neces-

sary data for training the developed model is collected.

Keywords Machine Learning · Predictive Mainte-

nance · Paper Machine · DTC · Simscape

1 Introduction

In the industrial sector, especially in areas like paper

manufacturing where operational efficiency and pro-

cess continuity are paramount, equipment maintenance

plays a crucial role in the success and profitability of op-

erations. Maintenance management is sometimes con-

the ability to effectively prevent unexpected motor fail-

ures remains a significant challenge [2][3].

Cristian P. Chuchico

Escuela de Ingenier´ıa y Tecnolog´ıa

International University of La Rioja

La Rioja, Spain

Tel.: +593 992842517

E-mail: cristianpaul.chuchico050@comunidadunir.net

Oscar Acosta Agudelo

Escuela de Ingenier´ıa y Tecnolog´ıa

Moreover, artificial intelligence has seen broad evo-

lution across diverse fields, including service sectors,

healthcare [4], robotics [5], and the industrial domain.

specks and various types of diffraction in a laser beam.

The research achieves an accuracy of around 99% with

of conveyor belts is conducted in [9], summarizing the

results and challenges of various methodologies used in

these systems. In addition, [10] shows a multihead neu-

ral network developed under the variability of individ-

ual machine degradations to derive machine-level prog-

nostics. This network learns degradation features and

updates remaining useful lifetime (RUL) distributions

from diverse distribution ensembles. Even though these

articles are closely related to the paper manufacturing

sector, a comprehensive analysis using artificial intelli-

tor operation, thereby preventing mechanical damage

to couplings, crosses, and cardan shafts in the system.

With the incorporation of this tool, the information

from operating variables such as speed reference, speed

feedback, motor current, torque reference, and motor

torque (calculated based on motor voltage and current)

is analyzed. This analysis helps determine whether the

machine’s operation is adequate or if there is a need

to plan activities to address out-of-standard conditions,

thereby avoiding unplanned production stoppages [11][12].

main contribution of this work is the incorporation of

a neural network for anomaly detection in motors in

the paper industry. By validating the contribution of

this neural network, it is possible to integrate the algo-

rithm into a real production environment and connect

it to the SCADA system via an OPC server, simplify-

ing signal interpretation and contributing to industrial

maintenance management.

2 Neural Network Configuration and Training

2.1 Simulation Scenario Setup

that emulate the physical system: power supply, recti-

fier block, inverter, and motor. Fig. 1 shows the cited

setup and its connections.

2.2 Data Generation for Training

As described in Table 1, during the model simulations,

different combinations of speed and torque are applied

by using step signals at specific time intervals at the

back to emulate abnormal system behavior. The results

shown in Figure 3 indicate that the controller exhibits

highly oscillatory behavior.

The simulation behavior aligns with mechanical issues

encountered in a real system, as shown in Figure 4. The

figure displays the torque of four motors: motors a and

b are operating normally, while motors c and d exhibit

signals from systems with mechanical problems.

The data for the analysis are obtained from several sim-

ulations, which include different scenarios such as ac-

celeration, deceleration, and steady state. The training

of the algorithm is based on identifying the behavior of

– Torque control signal (controller output).

– Motor torque.

– Motor current.

2.3 Neural Network Training

The process is divided into several trials to define the

best parameters for the neural network, with the con-

figurations shown in Table 2. After the training pro-

cess, the model’s performance is evaluated using con-

fusion matrices for each trial, allowing us to visual-

Table 1 Example of applied speed and torque combinations.

step Operation State Torque Ref Torque Motor Speed Ref

T1 Null Null Null Null

T2 Acceleration Null Over Reference 0 to 1200 RPM

T3 Stable 600 and 100 Nm According with reference 1200 RPM

T4 Deceleration 100 Nm Under refernce 1200 to 500 RPM

each trial conducted are presented below: In the first

trial, the confusion matrix (Figure 5) shows that the

model achieved a 70.3% accuracy. In the second trial,

whose results are presented in the confusion matrix in

Figure 6, the model’s consistency was confirmed, and a

72.7% accuracy was achieved. Subsequently, in Trial 3,

an increase in the input data was made, and three previ-

ous samples were considered for analysis. That is, to de-

termine the system’s performance at time t

, the values

distinguish between normal and faulty states improved

(Figure 7). In Trial 4, the algorithm’s robustness and its

ability to generalize from the training data were demon-

strated, achieving the best performance (Figure 8). For

Trial 5, an adjustment was made to the data alloca-

tion as follows: 70% for training and 30% for testing, in

order to evaluate whether the algorithm maintains its

performance and to rule out overfitting of the network

(Figure 9).

3 Metrics Calculation

Based on the developed trials, precision, accuracy, and

recall are calculated, considering the values of true pos-

itives (TP), false positives (FP), false negatives (FN),

and true negatives (TN) shown in each of the confusion

matrices from the previous section. Accuracy indicates

the proportion of correct predictions and is defined as

Recall indicates the proportion of actual positives that

Table 2 Configuration of the conducted trials.

Trial Input layer size Hidden layer size Output layer size Training-validation sample split

Trial 1 6 2 2 90 - 10

Trial 2 6 5 2 90 - 10

Trial 3 24 1 2 90 - 10

Trial 4 24 3 2 90 - 10

possible to design a machine learning algorithm that

continuously analyzes motor operation variables with-

out interfering with daily production activities, provid-

ing a basis for future research and practical applications

related to optimizing control systems and reducing op-

License

Agudelo.

ative Commons 4.0 license.

terial in any medium or format — and adapt the docu-

ment — remix, transform, and build upon the material

— for any purpose, even commercially, provided you

comply with the conditions of Attribution. You must

give appropriate credit to the original work, provide a

link to the license, and indicate if changes were made.