IPTEK The Journal of Technology and Science, 36.
, 2025 .
/pISSN:2088-2033, 0853-4.
DOI: 10.
12962/j20882033.
Submitted 01-30-2025.
Reviewed 03-11-2025.
Accepted 03-18-2025
ORIGINAL RESEARCH
Predicting Failure Using Machine Learning and Statistical-Based Method: A Production Machine Case Study Effi Latiffianti* | Stefanus Eko Wiratno | Samuel Aditya Christianta 1 Dept.
of Industrial and Systems Engineering.
Institut Teknologi Sepuluh Nopember.
Surabaya.
Indonesia Correspondence Abstract This research investigates the applicability of failure detection models based on machine learning and statistical approaches to reduce unplanned downtime in a food *Effi Latiffianti.
Dept of Industrial and Systems Engineering.
Institut Teknologi Sepuluh Nopember.
Surabaya.
Indonesia.
Email: effi_l@ie.
production company.
Sensor data is utilized to for identifying early failure symptoms.
To capture temporal and sequential dependencies in time-series data, we employ one of potential network based method so called the Long Short Term Memory (LSTM) Autoencoder.
Furthermore, we contrast the performance of the result with the tradi- Present Address Gedung Teknik Industri.
Kampus ITS Keputih Sukolilo.
Surabaya 60111.
Indonesia tional statistical method, the multivariate Exponentially Weighted Moving Average (EWMA).
While both models successfully detected all failures.
LSTM-AE demonstrated superior performance by reducing false alarms and providing true alarms with a longer time-to-failure.
The findings highlight the potential of leveraging limited data for failure prediction, demonstrating the effectiveness of both models in detecting anomalies while emphasizing their role in enhancing productivity through early failure detection.
KEYWORDS:
Anomaly detection, failure, fault.
Long Short Term Memory Autoencoder.
Multivariate Exponentially Weighted Moving Average.
INTRODUCTION
Maintenance is crucial for optimal performance and lifetime of industrial infrastructure and equipments.
In general, it includes regular inspections, repairs, strategic asset management, and life cycle planning to reduce risks, prevent failures, and improve As industries adopt new technologies, effective maintenance strategies are vital for maintaining operational efficiency and competitiveness.
Unplanned downtime due to sudden failures and emergency maintenance affect around 90% of organizations in European companies .
, with traditional maintenance methods often failing to prevent these disruptions .
Even with planned maintenance, unexpected breakdowns can still occur, leading to inefficiencies.
Latiffianti ET AL.
Detecting a failure before its occurrence can prevent sudden machine breakdowns, which contribute to significant production Early detection allows for proactive maintenance, reducing downtimes and managing costs more effectively.
Data-driven failure detection models have been extensively applied to wind turbines, leveraging sensor data collected through Supervisory Control and Data Acquisition (SCADA) systems to enhance operational reliability .
Ae.
Similarly, recent studies by .
demonstrate the potential of predictive models in forecasting failures within manufacturing machinery.
Building on these advancements, this study aims to implement and evaluate failure detection models within a selected case study to further explore their applicability and effectiveness.
Our case study focuses on a production machine of a leading poultry-based food industry in Indonesia.
The maintenance team has observed several instances of abnormal machine behavior, including increased noise levels, reduced performance, and elevated temperatures, which have resulted in breakdowns.
Some of these breakdowns cause the production line to come to a complete halt, leading to significant financial losses.
In some cases, the unavailability of replacement parts further prolongs the downtime, exacerbating the impact on production.
This highlights the critical importance of implementing effective failure detection mechanisms, such as anomaly detection models, which can monitor equipment conditions, provide early warnings of potential failures, and enable proactive interventions.
By preventing unexpected downtimes and ensuring the availability of necessary parts, such systems can significantly enhance operational efficiency, reduce costs, and maintain the continuity of production.
Anomaly detection identifies deviations from expected patterns in data, signaling potential equipment failures .
However, false alarms, where normal conditions are inaccurately flagged as anomalies, pose a significant challenge .
, .
A systematic approach is needed to filter out false alarms and improve model accuracy.
In light of the above details, this study covers three main objectives.
The first objective is to apply one of machine learning techniques that has shown some excellent performance in failure detection Ae long short term memory (LSTM), which in our case will be combined with autoencoder and refer to as LSTM-AE.
The second objective is to investigate whether the classical statistical-based approach, i.
e multivariate exponentially weighted moving average (EWMA) can also work comparably well in the case study.
Additionally, we would like to understand how much the two differ by comparing their performance.
The subsequent sections of this document are organized to provide a comprehensive understanding of the study.
Section 2 focuses on the data used in the analysis.
Section 3 provides the methodologies applied, detailing the approaches, techniques, and models employed to address the research objectives.
Section 4 presents the results of the analysis and provides an in-depth discussion of the findings, highlighting their implications and relevance to the study.
Finally.
Section 5 reports the conclusion, the key insights summary, contributions, and potential directions for future research or practical applications.
THE DATA
In this study, the dataset was collected from a production machine in the poultry-based food production line.
The machine, a Townsend NL-17 sausage linker, has been recently equipped with sensors to enhance maintenance and operations.
However, due to resource limitations, the collected data has not been adequately utilized.
The sausage linker is a critical component of the production line, frequently failing and accounting for over 50% of the downtime in sausage production.
The dataset consists of four variables collected from four different types of sensors, with measurements recorded every 5 minutes over one month of machine operation.
The dataset has a dimension of 10,397 rows and 5 columns.
Table 1 outlines the variables included in the dataset.
TABLE 1 Available variables from sensor measurement in the dataset.
Variable measured Current Voltage Power Power factor Description Flow of electrical charge in the linker conductor Electrical potential difference in the linkerAos electrical circuit Rate of energy conversion in the linker component Ratio of real power to total power supplied Unit of measurement Ampere (A) Volt (V) Watt (W) Ratio Latiffianti ET AL.
To validate the early failure detection results, failure log data was also utilized.
Our analysis focuses on breakdowns caused by linker component failures in the sausage linker machine.
Table 2 presents details of these breakdowns, including their duration, timestamp, and description.
TABLE 2 Failure record in the dataset.
Duration .
n minute.
Failure time
2024-03-02 10:20
2024-03-15 03:30
2024-03-22 02:30
2024-04-05 19:20
Failure description Linker chain dislodgement Bearing problem .
ear and tea.
Linker chain tension issue Linker chain problem .
METHOD We apply two methods in the analysis: the LSTM-AE and multivariate EWMA.
The LSTM-AE combines the feature reconstruction capabilities of autoencoders with LSTMAos ability to capture dynamic relationships in time-series data, enabling effective anomaly detection in sequential, unlabeled data.
Because it is an unsupervised model, we do not actually need to split the data into training and testing sets.
However, we will do the splitting to evaluate the performance on the training period to see the consistency of the results when it is applied on the test set.
Multivariate EWMA, on the other hand, employs a multivariate time-series approach with exponentially decreasing weights to the historical data, enhancing its ability for detecting changes in the data especially in time series contexts.
Although multivariate EWMA does not required labeled data, it is categorized as semi-supervised learning.
In order to detect abnormality using multivariate EWMA, a baseline needs to be determined for the normal data.
Because the sensor data is accompanied by the failure logs, the task of identifying normal data is less complicated.
The following subsections explain the LSTM-AE, multivariate EWMA, and the implementation details that we used in this Long Short Term Memory Autoencoder (LSTM-AE) Long Short Term Memory (LSTM) networks is an advanced type of Recurrent Neural Network (RNN) that are particularly effective in capturing long-term dependencies in sequential data such as time series data.
LSTMs use gates and memory cells to selectively store or discard information, enhancing their ability to model sequential data .
, which is similar to the case like predictive maintenance.
The primary advantage of LSTM is its ability to identify long-term dependencies in sequential data.
Figure 1 presents the structure of LSTM.
FIGURE 1 The structure of LSTM.
Latiffianti ET AL.
In Figure 1 , ycuyc is the input vector at timestamp t.
Eayc and EaycOe1 represent the current and previous output, respectively.
yayc and yaycOe1 represent the new and previous cell states or memory at timestamp yc and yc Oe 1, respectively.
The y and in the circle are vector multiplication and vector addition, respectively.
tan1 and tan2 are tanh neural networks and tangent function.
The equations for the gates are provided in Equation 1 Ae 3.
ycnyc = yua.
cycn [EaycOe1 , ycuyc ] ycaycn ), .
yceyc = yua.
cyce [EaycOe1 , ycuyc ] ycayce ), .
ycuyc = yua.
c0 [EaycOe1 , ycuyc ] yca0 ), .
where yceyc ,ycnyc , and ycuyc represent forget gate, input gate, and output gate, respectively.
The sigmoid function is represented by yua, ycycu is relevant weight in respective gate ycu associated with each block, and ycaycu is bias neurons at gate ycu.
The equations for the cell state, candidate cell state and the final output are presented in Equation 4Ae6 as follows:
( [
yaEyc = tanh ycyca EaycOe1 , ycuyc ycayca yayc = yceyc O yaycOe1 ycnyc O yaEyc , .
( ) Eayc = ycuyc O tanh yayc , .
where yaEyc refers to a tanh output and represents candidate for cell state at timestamp t, and * represents the element wise multiplication of the vectors.
LSTMs have demonstrated superior performance over traditional machine learning models in predicting the remaining useful life of equipment .
, .
, as well as fault detection in different application such as aircraft engines .
, .
, and wind turbine systems .
Ae.
LSTM was also reported to perform better that other models like Decision Tree.
Random Forest.
Convolutional Neural Networks, and XGBoost in terms of accuracy and other metrics .
, .
Autoencoders are a specialized type of feedforward neural network designed to reproduce input data as accurately as possible in the output.
They achieve this by transforming the input into a lower-dimensional representation, often called the latent space or code, and then reconstructing the original input from this compact representation .
The latent code serves as a condensed summary, capturing the most essential features of the input.
An autoencoder is composed of three primary components: the encoder, the code, and the decoder as illustrated in Figure 2 .
The encoder compresses the input data to generate the latent code, while the decoder reconstructs the input using only this compressed representation.
This structure enables autoencoders to learn efficient data representations and uncover underlying patterns.
In manufacturing, combining Autoencoders with Long Short-Term Memory (LSTM) deep learning models is recommended for anomaly detection .
LSTM, with its Encoder-Decoder architecture, can process input sequences of varying lengths and generate corresponding output sequences.
LSTM-Autoencoder (LSTM-AE) is particularly effective in capturing the dynamics and complex relationships within time-series data, making it well-suited for detecting anomalies in manufacturing processes .
Studies highlight LSTM-AEAos effectiveness in different applications.
demonstrates its success in detecting anomalies in industrial gas turbines and CPU utilization.
Similarly, .
showcases its potential for fault detection in electric drives.
Another study integrates LSTM and Autoencoder networks into a specialized encoder-decoder LSTM architecture, enhancing monitoring capabilities in complex dynamic processes .
In our case study, we set the sequence length of data used in the model to 5, which is equal to 25 minutes based on the nature of the data The selection was based on the result from autocorrelation analysis.
Furthermore, we used Mean Squared Error (MSE) as the loss function in training and testing of the model.
MSE is particularly suited for this study because it squares the error term, making it more sensitive to large deviations or outliers.
This means that large errors .
disproportionately impact the final MSE value, making significant deviations easier to identify .
Anomalous data points are identified as those Latiffianti ET AL.
FIGURE 2 Autoencoder illustration (Reprinted from Wolpe and Waal, 2.
with higher reconstruction errors compared to the majority of the data.
Finally, a threshold was selected by targeting the need for proper lead time in detecting failures in the training data and acceptable number of false alarms justified by the maintenance engineers in the observed production line.
Multivariate EWMA The exponentially weighted moving average (EWMA) control chart is a well known statistical process control (SPC) method renowned for its effectiveness in identifying small shifts in process parameters.
In contrast to traditional Shewhart control charts, which analyze individual data points.
EWMA prioritizes trend detection by applying exponentially decreasing weights to older This characteristic enhances its sensitivity to subtle deviations, often serving as early indicators of potential failures.
As .
notes, the cumulative approach of EWMA offers a distinct advantage in identifying gradual changes in processes, making it particularly valuable for applications that demand timely failure detection.
EWMA tracks the mean value of a process by giving recent observations more importance through exponentially decreasing weights.
This method provides a smoothed estimate of the process mean, making it sensitive to subtle changes in data that the traditional control chart method might miss .
In multivariate anomaly detection.
EWMA has been adapted from its original use with single variable to handle complex datasets with multiple interconnected variables.
This adaptation brings several advantages.
For instance, multivariate EWMA can effectively capture relationships between variables by incorporating covariance matrices.
This capability allows it to detect anomalies that affect multiple variables simultaneously, which is crucial and has become the most rapidly developing area in industries like manufacturing and complex systems monitoring .
The application of EWMA in detecting anomalous data points has been widely reported to be effective.
highlights that EWMA excels in the early detection of anomalies across interconnected variables, providing a proactive approach to system monitoring, particularly in scenarios where anomalies can impact multiple parameters simultaneously.
When combined with other methods.
EWMA performs well in simulated data for failure detection .
, .
Successful implementations of EWMA, as well as its modified version, the Unscented Kalman Filter (UKF)-based EWMA, in fault detection for wastewater treatment plants further demonstrate its applicability across various domains .
, .
The study in .
also compares EWMA with m-CUSUM, revealing that EWMA outperforms its counterpart in fault detection for wastewater treatment plants.
In photovoltaic systems, the EWMA control chart effectively monitors abnormal states and accurately identifies the specific string where a fault occurs .
Its detection power can be further enhanced when integrated with other methods .
Similarly, in wind turbine systems.
EWMA has successfully detected outliers up to one week before bearing failure .
, allowing for early fault alarms and proactive maintenance.
Based on its capability and the nature of the case study, we consider multivariate EWMA model as a potential alternative of solution for early failure detection.
Unlike machine learning approaches, there is no iterative training on the multivariate EWMA Latiffianti ET AL.
This study uses the basic form of the multivariate EWMA chart to monitor shifts in the mean vector of a multivariate The multivariate EWMA score per data point is the output of the multivariate EWMA approach and can be obtained using Equation 7 and 8.
yeuyc = yuI.
Oe yuI)yeuycOe1 ycnOe1 Oc yuI.
Oe yuI)yc yeoycnOeyc .
Oe yuI)ycn yeu0 , .
yc=1 ycNycn2 = .
euyc Oe yyA0 ]yc .
yoOe1 euyc Oe yyA0 ] yc where yyA0 is the mean vector of yeoycn and yoycycn are the the covariance matrix of yeuycn .
When ycn is large, i.
ycn Ie O, yoycycn can also be yuI estimated as: 2OeyuI yo0 , where yo0 is the covariance matrix of yeoycn .
In our case, yeoycn was defined as the vector of variables in data points that are considered normal, i.
at least 24 hours away from the failure events.
Furthermore, the implementation of EWMA requires two parameters to be determined: the smoothing constant yuI, and the threshold or the control limit that acts as a boundary between the anomalous and normal.
proposed the use of a smoothing constant .
uI) that ranges between 0.
05 and 0.
In our study, we used yuI = 0.
20 as it has been proven robust in anomaly detection .
and it also produced best result in our case study.
The selection of the control limit was based on the training data by targeting the true detection of the actual failure.
Any multivariate EWMA value higher than the threshold indicates anomalous data behavior and to be flagged as an alarm.
Implementation Details We begin our analysis with data preprocessing, where we normalized the variables using a robust scaling approach.
This technique adjusts features by subtracting the median and scaling them based on the interquartile range (IQR).
Unlike methods that rely on the mean and standard deviation, robust scaling is less sensitive to outliers and extreme values .
, making it particularly suitable for anomaly detection context.
The scaled values were using the formula in Equation 9.
ycuAycn = ycuycn Oe ycE2 .
ycE3 .
Oe ycE1 .
where ycuAycn is the transformed value of the ycnycEa variable, ycuycn is the original value of the ycnycEa variable, and ycE1 .
, ycE2 .
, ycE3 .
represent the first, second, and third quartile of the data, respectively.
After the transformation, we employ the autocorrelation function (ACF) to analyze the relationships between past and current data points for each feature, determining if the data sequence should be treated in a time-series context.
The ACF results showed significant short-term dependencies, especially up to five lags, indicating that recent past values influence future values, which informed our decision to set the LSTM sequence length to five.
While the voltage did not show time-dependent behavior, it was still included as a predictor due to its relevance in indicating potential failures.
The data exhibited short-term temporal dependencies, underscoring the need for a time-series approach in our predictive modeling to capture these patterns effectively.
Next, we divided the data into two subsets: 70% for the training set and 30% for the testing set.
While the two methods primarily rely on unsupervised and semi-supervised learning techniques, we performed the split in the same way as in supervised learning.
This approach allows us to simulate real-world conditions, where future observations are unavailable during model training.
Consequently, we tuned the model parameters based solely on the training set.
We applied two models.
LSTM-AE and multivariate EWMA, to the subsets.
The LSTM-AE model was trained using the training set and evaluated on the test set.
The multivariate EWMA approach, however, was implemented in two distinct phases.
In Phase I, normal data was selected to establish yeayen0 , yo0 , and the control limit .
tailored for anomaly detection.
In Phase II, the entire dataset was analyzed using all those parameters defined in Phase I to perform the fault detection.
Finally, we evaluated the performance of both approaches using two metrics: the total number of false alarms and the failure detection lead time.
The failure detection lead time refers to the time interval between the first alarm or detection and the actual machine failure.
For this case study, we established a maximum lead time of 24 hours, based on input from maintenance experts Latiffianti ET AL.
familiar with the production machines and processes.
This threshold was chosen to reflect the machineAos historical behavior and the characteristics of the components monitored by the sensors.
Furthermore, since production operates on three shifts per day, a lead time exceeding 8 hours was considered preferable in This would provide the maintenance team with sufficient time to plan and execute maintenance activities during shift Consequently, we defined a false alarm as any alarm triggered more than 24 hours before a failure or less than 8 hours before a failure.
The numerical analysis was conducted using Python, with PyTorch specifically used for the LSTM-AE model.
Since the dataset is not large, the implementation runs efficiently on any computer with at least 8 GB of memory.
RESULT AND DISCUSSION
Once the models are implemented, we calculate a score that is used for failure detection.
To trigger an alarm, a threshold is If the score exceeds this threshold, an alarm is raised.
The alarm will remain active until the score drops back below the threshold.
The tricky part is where to set the threshold, considering the balance between the true detection and false alarms.
Naturally, lowering the threshold increases the likelihood of false alarms.
However, it also reduces the chances of missing a true failure, resulting in a higher true positive rate.
Conversely, raising the threshold decreases the number of false alarms but increases the risk of failing to detect actual failures.
Practitioners often look for the balance between the two, aiming at the highest detection rate with considerably acceptable number of false alarms.
In practice, missed detections are the most critical issue to avoid, as their impact is significantly more severe.
Missed detection results in an unscheduled production line stoppage, leading to lost production and reduced productivity for all other resources that depend on the production process.
Furthermore, unexpected failures often create additional challenges, such as the unavailability of necessary inventory to restore operations, further compounding delays and inefficiencies.
For this reason, sometimes the performance of a model can be measured from the number of miss detections produced from the model.
In many machine learning applications, model performance is commonly evaluated using metrics such as F1 score, accuracy, precision, and recall.
While these metrics are effective in many contexts, they are not always ideal for failure detection.
practice, production managers often prioritize the practical benefits of a model, such as the cost savings or financial impact it brings to real-world operations.
Although this type of metric is less commonly used, it has been introduced in certain studies.
For example, .
calculated the total savings generated by a method based on real-world maintenance costs, including replacement costs, inspection costs, and repair costs.
In our case, we lack the necessary data on these costs to perform a similar analysis.
However, we can still evaluate the models using alternative metrics, such as average lead time, the number of false alarms, and the number of missed detections, which can provide comparable insights into the practical effectiveness of the models.
Table 3 and Table 4 present a small range of threshold, denoted as H, that were implemented on the two methods.
From these two tables, we can see how different values of H affect the results of the detection.
Suppose our goal is to detect all failures within 8 to 24 hours of their occurrence.
Based on Table 3, we have two options to consider for detection: setting the threshold at ya = 4.
5 or ya = 4.
Considering the nature of the case, where the production line operates in three working shifts, we can conclude that the average lead time between the two thresholds is not significantly different.
In other words, detections will typically occur within one or two shifts of the actual failure.
This lead time should provide sufficient opportunity to arrange maintenance actions during shift changes, when the machine is normally shut down.
Therefore, ya = 4.
5 is the more favorable option to be used with LSTM-AE scores, as it leads to fewer false alarms.
In a similar fashion, using information presented in Table 4, ya = 7 or ya = 8 are two very good options for the multivariate EWMA scores.
With the thresholds set for both methods, we can now compare their performance in terms of the number of false alarms and the average lead time.
There are two approaches to this comparison:
Set ya to values that result in a comparable average lead time for both methods, then evaluate the number of false alarms.
Set ya to values that produce a similar number of false alarms for both methods, then assess the average lead time.
Latiffianti ET AL.
TABLE 3 LSTM-AE performance based on selected threshold.
Threshold Average lead time .
# of false alarms # of true alarm # of total alarm # of miss detection ya = 5.
ya = 4.
ya = 4.
TABLE 4 Multivariate EWMA performance based on selected threshold.
Threshold Average lead time .
# of false alarms # of true alarm # of total alarm # of miss detection ya = 12 ya = 11.
ya = 10 ya =9 ya =8 ya =7 ya =6 Both approaches are expected to lead to the same conclusion.
On the first approach, let ya = 4.
5 for the LSTM-AE and ya = 8 for multivariate EWMA.
Based on Table 3 and Table 4, this will result in the average lead time of 14.
54 hours and 13.
67 hours for LSTM-AE and multivariate EWMA, respectively.
With these thresholds, the total number of false alarms are 10 and 28 for LSTM-AE and multivariate EWMA, respectively, which means LSTM-AE outperformed multivariate EWMA.
On the second approach, let ya = 4.
0 for the LSTM-AE and ya = 11.
5 for multivariate EWMA.
This will give us 18 and 17 false alarms for LSTM-AE and multivariate EWMA, respectively.
In terms of average lead time.
LSTM-AE was able to detect all 4 failures with average lead time of 16.
92 hours, while multivariate EWMA could not detect the first and second failures, resulting in only 7.
90 hours lead time in average.
Again.
LSTM-AE outperformed multivariate EWMA.
Table 5 presents detection time and time to failure produced from LSTM-AE with ya = 4.
5 and MEWMA with ya = 8.
The detection plots for both methods are presented in Figure 3 .
TABLE 5 Failure detection times resulted from LSTM-AE and multivariate EWMA
Failure time
LSTM-AE
EWMA
Earlierst alarm Lead time .
Earlierst alarm Lead time .
2024-03-02 10:20 2024-03-01 15:40 2024-03-01 16:10 2024-03-15 03:30 2024-03-14 08:00 2024-03-14 11:25 2024-03-22 02:30 2024-03-21 05:05 2024-03-21 06:25 2024-04-05 19:20
2024-04-04 16:45
2024-04-04 15:00
Both the LSTM-AE, a machine learning-based method, and EWMA, a statistical approach, effectively detect anomalies to provide early failure warnings.
However, the LSTM-AE outperforms the multivariate EWMA model by detecting failures earlier, about an hour or 5.
05% earlier.
Furthermore, when the multivariate EWMA threshold is adjusted to match the failure detection lead time of the LSTM-AE, the multivariate EWMA model generates more false alarms, i.
two times higher .
r 211%).
This highlights the LSTM-AEAos superiority over the multivariate EWMA model, both in terms of earlier failure detection and reduced false alarms.
Latiffianti ET AL.
FIGURE 3 Detection plots of multivariate EWMA .
and LSTM-AE .
The solid horizontal line represents the threshold ya, the vertical solid lines mark the time of the actual failuresAo occurrence, and the vertical dashed lines are the time of the earliest alarms.
In practice, an early warning system enables management to anticipate failures more effectively.
For instance, if the warning provides a lead time of more than eight hours, the maintenance team can schedule an inspection during the next shift change when the machine is normally stopped, minimizing production loss.
Additionally, if maintenance personnel are unavailable when the alarm is triggered, the advance notice allows sufficient time to bring them on-site for the scheduled inspection.
Based on the inspection, appropriate maintenance actions can be determined.
If immediate repairs are not possible, preventive measures can be taken to avoid failure until the necessary maintenance is This proactive approach helps prevent prolonged downtime, production losses, and productivity declines.
To understand what happened during the period after the earliest alarms, we perform additional analysis on the measured The begin by dividing the data into five subsets: the normal data points, and four subsets of data points where the score from one or both methods are greater than the control limit.
We plotted the density of each variable from those subsets in Figure 4 .
We differentiated the plots from each subset using different colors.
Based on Figure 4 , we can clearly see that voltage does not provide pronounced clue about the failure except for Failure #1 and Failure #3 where the sensor readings yield wider density, with spikes on the right side, indication some higher values compared to the rest of the data.
The current and the power on the other hand, present similar patterns across different failures, showing sort of bimodal distribution with some density raising on the opposite direction of the normal ones.
Finally, on the power factor, small portions of failure signatures show larger values, but at the same time, the density from all failures also shifted to the left, indicating that all four failures also present nearly zero power factor as the one of the symptoms.
In summary, the exploration of the sensorsAo reading values reveals that all four variables we used in the model contributes to the detection resulted by the two models.
However, we cannot isolate nor differentiate the exact cause of each of the failures solely based on this analysis.
The early failure warning capability demonstrated in this study is based on a limited dataset.
Nevertheless, the results serve as a validation of the anomaly detection model, showing that it can successfully identify anomalies as early indicators of failure and provide timely warnings.
While these results may vary when tested on larger datasets, the findings presented here offer an initial understanding of the modelAos ability to provide early failure warnings.
CONCLUDING REMARKS
This research highlights the benefits of using machine learning and statistical models for early anomaly detection in predictive By analyzing time-series data, the LSTM-AE and multivariate EWMA models effectively identified early signs Latiffianti ET AL.
FIGURE 4 Variable exploration based on combined subset from both LSTM-AE and multivariate EWMA.
of equipment failure.
These models detected failure patterns that were not apparent in the raw data, with the LSTM-AErelying on reconstruction errors and the multivariate EWMA model using statistical measures.
While the results may vary with larger datasets, they provide an initial understanding of the modelsAo ability to deliver early failure warnings.
A crucial aspect of effective anomaly detection is setting appropriate detection thresholds.
Lower thresholds improve precision but result in shorter lead times, whereas higher thresholds allow for longer lead times but generate more false alarms.
Finding the right balance between early detection and reliability is the key.
Among the two models, the LSTM-AE showed better performance, particularly in detecting subtle anomalies and providing earlier warnings, while also generating fewer false alarms.
However, the multivariate EWMA model remains valuable for its ability to effectively handle short-term patterns.
In summary, the LSTM-AE demonstrates strong potential for enhancing early failure detection by providing accurate and timely Future research should aim to test these models with larger datasets and refine threshold optimization to further improve their reliability and effectiveness.
Another promising direction is the integration of machine learning-based anomaly detection with real-time industrial monitoring systems, such as Industrial IoT platforms, to enable automated decision-making and predictive maintenance.
While this study is centered on a food production company, further research could investigate the applicability of these models in other industries, such as automotive manufacturing, pharmaceuticals, and energy production, to assess their robustness in different operational environments.
Additionally, conducting a cost-benefit analysis of implementing machine learning-based predictive maintenance would provide decision-makers with valuable insights into cost savings, return on investment, and overall productivity improvements.
Latiffianti ET AL.
CREDIT
Conceptualization.
Methodology.
Writing - literature review, formatting, reviewing, editing.
Formal analysis and investigation, finalizing draft: Effi Latiffianti.
Conceptualization.
Writing - reviewing.
Data collection, and Research supervision: Stefanus Eko Wiratno.
Data collection.
Formal analysis and investigation.
Writing - first draft: Samuel Aditya Christianta.
References