Comparative Prediction of Physical Fatigue Patterns in Bandung, Indonesia Workers using CNN and ANN
M Fikri Raihan Ardiansyah*, Rifki Wijaya, Gia Septiani Wulandari Departement of Informatics, Faculty of Informatics, Telkom Univeristy, Bandung, Indonesia Email: 1,*[email protected], 2[email protected],
Correspondence Author Email: [email protected]
Abstract−This research explores the impact of physical fatigue on task performance and evaluates the effectiveness of Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) in predicting fatigue levels. Physical fatigue, as a critical factor influencing performance and safety, serves as a signal for the body's need for rest. Utilizing a smartwatch with heart rate sensors, this study applies ANN for subjective fatigue assessments and CNN for time series analysis. With a structured approach encompassing data collection, preprocessing, and model training, a confusion matrix evaluates the model's performance. Results indicate an accuracy of 92.4% for the ANN model with an RMSE of 0.275, while the CNN model achieves an accuracy of 85.46% with an RMSE of 0.381. These findings affirm the effectiveness of both models in predicting fatigue, providing valuable insights for future research and emphasizing the importance of comprehensive data analysis for a nuanced understanding of individual performance (Number of data: 149,796 from 6 subjects).
Keywords:Physical Fatigue; Artificial Neural Network (ANN);Convolutional Neural Network (CNN) ; Time Series; Heart Rate
1. INTRODUCTION
Engaging in tasks that require physical effort will cause someone to experience physical fatigue and will decrease efficiency in performing tasks if done continuously[1],[2]. Every individual will have different physical endurance, so it is important to understand and be aware of one's body condition as it significantly impacts performance, as fatigue is one of the main causes of accidents[3], [4],[5][6]. The reduction in physical or mental performance due to physical factors will significantly affect overall physical abilities[7], [8], [9]. Fatigue is one of the body's mechanisms signaling that the body needs rest.
Heart rate tends to increase when an individual is experiencing physical fatigue. A smartwatch can be used as an assessment method for fatigue because it has sensors capable of estimating heart rate intervals and can be considered a valid accelerometer[3]. The data collection process using a smartwatch connected to a smartphone has the potential to be influenced by various factors, such as body movement and sensor placement on the device, which can impact the collected data. Systematic prediction involves formulating the potential future events based on historical data and current information.
The urgency lies in the user's ability to detect fatigue before it occurs or receive fatigue notifications, thereby preventing overwork or over-exercise. Fatigue notifications could also serve as an early warning if integrated into vehicles, ensuring that a vehicle does not accept or allow someone with a certain level of fatigue to operate it. Artificial Neural Networks (ANN) are utilized for predictive purposes due to their ability to process large volumes of data and provide predictions that are sometimes accurate. This method is highly effective in making predictions as it is based on the principles of biological neural networks[10]. And for Convolutional Neural Network (CNN), it is a specialized variant of a multilayer perceptron. In several cases, CNN has shown good performance in image categorization, object recognition, and can be used for prediction tasks. It has been employed in predictive tasks as well. [10],[11], [12].
This research aims to conduct a comprehensive comparative analysis of the effectiveness of Artificial Neural Network and Convolutional Neural Network in the context of how accurately the models predict fatigue.
It will utilize a confusion matrix to evaluate the performance of each model.
In the research conducted by Z. Ahmad, M. Jamaludin, and K. Soeed in 2018 titled "Prediction of exhaustion threshold based on ECG features using the artificial neural network model"[13], the developed mathematical model using Artificial Neural Network (ANN) shows promise in converting subjective assessments into objective ones. Sensitivity analysis of variables indicates that ANN achieves maximum accuracy related to its output. The confusion matrix demonstrates a very good prediction accuracy level, reaching 89.3%. However, the study has limitations, including a small number of participants and the absence of variables such as sleep patterns and eating habits that could influence physical fatigue.
In the research conducted by A. Zulkifli, J. Mohd Najeb, and J. Ummu Kulthum in 2020 titled "Physical Fatigue Prediction Based on Heart Rate Variability (HRV) Features in Time and Frequency Domains Using Artificial Neural Networks Model During Exercise"[14], fatigue prediction was conducted using the Artificial Neural Network (ANN) method, achieving an accuracy rate of 77.02%. However, the shortcomings of the journal lie in the brief explanation of the signal processing techniques employed to filter and analyze ECG signals, lacking comprehensive details. This limited information may impede other researchers from replicating or advancing the
research methods. In this study, we adopt the Adaptive Artificial Neural Network (ANN) method, proficient in predicting physical fatigue based on collected time series data.
In the research conducted by Shahnawaz Anwer, MPT in 2021, the focus is on predicting periodic multivariate time series using the Multiple CNNs model, This approach involves the utilization of Convolutional Neural Networks (CNNs) to enhance prediction accuracy[15]. Experiments on two real-world datasets demonstrate the superior performance of Multiple CNNs compared to baseline methods, reinforcing its capability to improve the accuracy of periodic information prediction. Related to the effective extraction of periodic information, this research has the potential to serve as a source of inspiration and comparison for the development of models.
In the research conducted by S. Mehtab and J. Sen in 2020, titled "Stock Price Prediction Using Convolutional Neural Networks on a Multivariate Time series"[12]. Combines machine learning and deep learning approaches to predict stock price movements using training data from January 2015 to December 2018 (Case I) and tests the model on test data from December 2018 to December 2019 (Case II). Classification methods are employed to predict index movement patterns, while regression models are constructed to forecast the actual
"Close" values. Model accuracy is enhanced through the use of a deep learning-based regression model using Convolutional Neural Network (CNN) with a walk-forward validation. The research findings reveal an overall Root Mean Squared Error (RMSE) value of 0.895 for models with a training data size of 5 (representing one week prior). Additionally, interesting patterns in RMSE are identified, notably a significant increase on Tuesdays compared to other weekdays. The study concludes that deep learning-based models exhibit higher feature extraction and learning capabilities than traditional machine learning models, and multivariate analysis improves prediction accuracy.
2. RESEARCH METHODOLOGY
2.1 Research Stages
This research adopts a structured approach that commences with the process of data collection and organization into a dataset. Pre-processing and normalization procedures are applied to ensure optimal data quality.
Subsequently, the dataset is split to train Convolutional Neural Network (CNN) and Artificial Neural Network (ANN) designed to capture relevant patterns. The performance evaluation of CNN and ANN is conducted using separate data, and the evaluation results are documented in Figure 1 for collection data and preprocessing data, and Figure 2 Testing ANN and CNN, represented in the form of a flowchart
Figure 1. Flowchart of Testing of ANN and CNN Models
Figure 2. Flowchart of Data Collection, Preprocessing 2.2 Data Colection
In the data acquisition phase, participants will wear a smartwatch on their wrists and engage in their daily activities for two days. They are expected to continuously wear the smartwatch, with instructions to remove it only during showering, charging the smartwatch, or as needed for a few minutes. After the two-day period, the data will be exported from the smartwatch to an Excel file. This process will be repeated for the next participants until the predetermined target is achieved.
2.3 Prepocessing
In the pre-processing stage of data, the focus is on transforming and cleaning the data to make it ready for further analysis. This process aims to eliminate noise, handle missing values, and enhance the analysis and performance of models. It also helps reduce the risk of bias due to incomplete or inconsistent data, as machine learning heavily relies on accurate input[16],[17]. Therefore, data pre-processing is a crucial step to be undertaken. In this current research, it includes data interpolation, normalization, and converting data to appropriate formats.
2.4 Convolutional Neural Network (CNN)
This model is a specialized variant of a multilayer perceptron, but it differs from other deep learning architectures as the underlying neural network cannot comprehend complex characteristics. In various applications, CNN algorithms have demonstrated excellent performance in image categorization, object recognition, and medical image analysis. Moreover, this model can be used for prediction tasks and has been employed in predictive tasks as well[10], [11] . The algorithm to be used is the Convolutional Neural Network Convolution-1D (Conv1D) with three main layers. This model uses 64 filters and a kernel size of 3, followed by the ReLU activation function. The optimizer used is Adam, and the loss function is implemented with Mean Squared Error. Additionally, this model also utilizes EarlyStopping to prevent overfitting.
2.4.1 Convolutional Layer
Convolutional Neural Network (CNN), a technique that aids in understanding how data is processed to extract patterns, represents a powerful approach in the realm of deep learning. In this exploration, our attention is directed towards the Convolutional Layer (1), a pivotal component within CNN architectures that plays a pivotal role in capturing spatial hierarchies and discerning intricate features.
𝓏(1)=Conv1D(X; W(1)X + b(1))
a(1)=ReLU(𝒵(1)) (1)
a(2)=ReLU(W(2)a(1)+ b(2)) ŷ=W(3)a(2)+ b(3)
Information for convolutional layer used, X for input sequence, W(1) Convolutional kernel, b(1) bias, W(2)and b(2) weights and bias in the first dense layer, W(3)and b(3) weights and bias of the output layer.
2.5 Artificial Neural Network
Artificial Neural Network (ANN) is an information processing model that emulates the activities and structure of biological neural networks found in the human brain. As one of the most predominant artificial intelligence algorithms for prediction[10],[18]. In this research, the loss function employed is Mean Squared Error (MSE), and the backpropagation method is utilized to compute the gradient of the loss function with respect to the weights and
biases. The model training is conducted with a feedforward, followed by invoking the 'fit' function on the training data using the backpropagation method and employing the Adam Optimizer. Additionally, EarlyStopping is implemented to prevent overfitting.
2.5.1 Feedforward
Feedforward, as a fundamental technique in Artificial Neural Networks (ANN), serves as the backbone for understanding the complex process of transmitting input data through various layers towards the final output. This mechanism, as illustrated in Equation (2), reveals the sequential activations from the initial layer a(1) to the final prediction ŷ
a(1)=ReLU(W(1)X + b(1))
a(2)=ReLU(W(2)a(1)+ b(2)) (2)
a(3)=ReLU(W(3)a(2)+ b(3)) ŷ=a(4)a(3)+ b(4)
Information for ANN feedforward used, X is an input vector with a length of 3, W(1) is the weight matrix for the i-th layer,and b(i) is the bias vector for the i-th, the activation function used is ReLU for all layers, except for the output layer which uses linear activation. And ŷ is the predicted output.
2.6 RMSE
RMSE, as an evaluation metric, measures how close the model predictions are to the actual values. The calculation is done by taking the square root of the average of the squared differences between predictions and actual values.
The model's accuracy is reflected in a low RMSE value[19].
RMSE = √∑ (yi
N pred
i=1 − yipred) 2
N (3)
2.7 Confusion Matrix
Confusion matrix is a table used to evaluate the performance of a prediction or classification model by comparing the model's predicted outcomes with the actual values from the observed data. This evaluation method utilizes a confusion matrix. The following is the table 1 for the confusion matrix.
Tabel 1. Confusion Matrix
Actual Value
Positive Negative Predictive Value Positive True Positive (TP) False Positive (FP)
Negative False Positive (FN) True Negative (TN)
In the confusion matrix table, there are four crucial evaluation metrics: true positive (TP), true negative (TN), false positive (FP), and false negative (FN).
Accuracy = (TP + TN + FP + FN) (TP + TN) (4)
3. RESULT AND DISCUSSION
In this section, I conducted several preprocessing steps, including interpolating data at 5-second intervals, labeling based on fatigue signs during the interview session, converting dateTime to datetime, replacing fatigue labels with numbers, converting each column to integers, normalizing the data and split data. A more detailed explanation can be found in sections 3.1 to 3.5
3.1 Interpolation
Interpolation is a mathematical method used to fill or estimate values between two known data points. In the applied linear interpolation method, values that are not recorded are filled with the nearest neighboring values[20].
Interpolation helps complete missing or unrecorded values, enabling more efficient data processing.
Before interpolation : Time : [0, 5, 10, 15, 20]
BPM: [77, 79, NaN, NaN, 89]
After interpolation:
Time : [0, 5, 10, 15, 20]
BPM: [77, 79, 82, 85, 89]
3.2 Labeling Data
The purpose of this step is to label the data based on the interview results with participants, using the label "ya" to indicate the presence of fatigue and "tidak" to signify the absence of fatigue. Subsequently, these labels will be transformed into binary representation, where "ya" is represented as 1 and "tidak" as 0.
Table 2. Labelling data fatigue “ya”
Id dateTime bpm step fatigue ID001 23/09/2023 15:27 86 6 ya ID001 23/09/2023 15:27 88 6 ya ID001 23/09/2023 15:27 89 6 ya
The table 2 illustrates the results of the labeling process on data obtained from interview sessions with participants experiencing physical fatigue.
Table 3. Labelling data fatigue “tidak”
Id dateTime bpm step fatigue ID001 23/09/2023 00:00 96 0 tidak ID001 23/09/2023 00:00 95 0 tidak ID001 23/09/2023 00:00 93 0 tidak
The table 3 illustrates the output of the data labeling process obtained from interview sessions with participants who did not experience physical fatigue.
3.3 Data Format Conversion
This step is taken to ensure that the data applied to the model meets the correct format and can be processed efficiently.
1. Convert dateTime to datetime :
data['dateTime'] = pd.to_datetime(data['dateTime']) 2. Replace coloum “kelelahan” :
data['kelelahan'] = data['kelelahan'].replace({'ya': 1, 'tidak': 0}) 3. Variabel to integer :
data['bpm'] = data['bpm'].astype(int) data['step'] = data['step'].astype(int)
data['kelelahan'] = data['kelelahan'].astype(int) data['dateTime'] = data['dateTime'].astype(int) 3.4 Normalize Data
The purpose of normalizing a dataset is to ensure that the values are within a similar range[21]. In this study, we employ the 'MinMaxScaler' method from sklearn, which scales the data to bring its values into the range [0, 1].
3.5 Split Data Train & Test
In developing the model, I divided the dataset into two parts: the training set (70%) and the test set (30%). The training set is used to train the model, while the test set is used to evaluate how well the model performs on new data. This split provides an indication of the model's performance beyond the training data.
3.6 Result of ANN
In the context of this study, the researcher will project predictions related to physical fatigue. This process involves the use of a dataset that has undergone preprocessing. The dataset consists of two classes, namely "tidak" (class 0) and "ya" (class 1). The main focus of this research is on training an Artificial Neural Network model using the training dataset.
3.6.1 Result of Model ANN
Figure 3. Accuracy Model Artificial Neural Network
The graph in Figure 3 displays accuracy of the Artificial Neural Network (ANN) model during training, divided into training accuracy and validation accuracy. The training accuracy tends to increase until the 30th epoch but plateaus afterward. To address overfitting, early stopping is applied at the 40th epoch, where the validation accuracy shows no improvement. Although the model with early stopping has lower training accuracy, its validation accuracy is higher compared to the model without early stopping. Consequently, the model stops at the 40th epoch, achieving a balanced accuracy between training and validation.
Figure 4. Loss for Artificial Neural Network Training
The graph in Figure 4 displays loss values of an Artificial Neural Network (ANN) model during training, separated into training loss and validation loss. The training loss generally decreases until the 30th epoch and then becomes stagnant. To prevent overfitting, the early stopping technique is implemented at the 40th epoch, activated when the validation loss does not decrease for the last 10 epochs. The model with early stopping exhibits higher training loss but lower validation loss compared to the model without early stopping. Consequently, the training of the model is halted at the 40th epoch, resulting in a model with a balance between higher training loss and lower validation loss.
3.6.2 Result of Test
Testing results in this research utilize artificial neural network methods with a focus on classification evaluation using a validation confusion matrix. The results include accuracy and Root Mean Squared Error (RMSE) values.
The model evaluation demonstrates impressive results, achieving an accuracy rate of 92.4%. The RMSE of 0.275 indicates that the model provides reasonably accurate predictions with minimal errors in estimation. The confusion matrix testing results are divided into four parts: True Positive (TP), indicating actual data is true, and the model predicts it correctly. True Negative (TN), indicating actual data is false, and the model predicts it as false. False Positive (FP), indicating actual data is false, but the model predicts it as true. False Negative (FN), indicating actual data is true, but the model predicts it as false.
Figure 5. Confusion Matrix result from Artificial Neural Network (ANN) 1405/1405 [==============================] - 2s 1ms/step
confusion matrix:
True Positive (TP): 16240 False Positive (FP): 1626
False Negative (FN): 1788 True Negative (TN): 25285 3.7 Result of CNN
In the context of this study, the researcher will project predictions related to physical fatigue by utilizing a pre- processed dataset. The labeled dataset consists of two classes, namely "tidak" (class 0) and "ya" (class 1). The primary focus is on training using a Convolutional Neural Network (CNN) model.
3.7.1 Result of Model CNN
Figure 6. Accuracy Model Convolutional Neural Network
The graph in Figure 6 displays accuracy of the Convolutional Neural Network (CNN) during training with two lines, namely training accuracy and validation accuracy. The training accuracy continues to increase, indicating the model's learning from the training data. The validation accuracy reaches its peak at epoch 40 and then decreases. Early stopping is implemented at epoch 50 to prevent overfitting, stopping the training if the validation accuracy remains stagnant for 10 consecutive epochs. With this strategy, the model can be achieved with optimal accuracy without the risk of overfitting.
Figure 7. Loss for Convolutional Neural Network Training
The graph in Figure 7 displays the loss values for both the model and training during the training of a Convolutional Neural Network (CNN), with termination occurring at epoch 50 due to the implementation of early stopping. Early stopping is employed to prevent overfitting, a condition where the model becomes overly tailored to the training data, impeding its generalization to new data. In the graph, the validation loss begins to rise at epoch 50, prompting the cessation of training. Despite the relatively low model loss value at this epoch (0.12), early stopping facilitates the model in achieving good performance on the training data without the risk of overfitting.
3.7.2 Result of Test
The assessment results of this study employ the Convolutional Neural Network (CNN) method, with an emphasis on classification evaluation through a validation confusion matrix. The obtained results encompass accuracy and Root Mean Squared Error (RMSE) values. The evaluation of the model showcases remarkable performance,
achieving an accuracy rate of 85.46%. The RMSE of 0.381 indicates that the model provides reasonably accurate predictions with minimal estimation errors.
Figure 8. Confusion Matrix result from Convolutional Neural Network (CNN) 1405/1405 [==============================] - 2s 1ms/step
Confusion Matrix:
True Positive (TP): 23676 False Positive (FP): 3397 False Negative (FN): 3137 True Negative (TN): 14729
4. CONCLUSION
The results of the Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) models were presented, showcasing their accuracy and loss during training. The ANN achieved an impressive accuracy of 92.4%, with early stopping applied at the 40th epoch to prevent overfitting. The CNN exhibited a peak validation accuracy at epoch 40, and early stopping at epoch 50 ensured optimal accuracy without overfitting risks. For the test results, the ANN demonstrated an accuracy rate of 92.4% with an RMSE of 0.275, while the CNN achieved an 85.46% accuracy with an RMSE of 0.381. Comparing the two models, the ANN outperformed the CNN in terms of accuracy. However, the CNN still provided reasonably accurate predictions with minimal errors. In summary, the study concludes that the ANN model, with its higher accuracy, performs better than the CNN model in the context of the fatigue prediction task. Nevertheless, the CNN model also exhibits commendable performance, and further research could explore optimizations for both models. The limitations of this study provide avenues for future improvements, emphasizing the need for continuous refinement in fatigue prediction methodologies and These findings can serve as a foundation for future research development, with prospects involving the addition of variables such as sleep and eating patterns to better understand the factors influencing physical fatigue.
ACKNOWLEDGMENT
This research was conducted in the Department of Informatics, Faculty of Informatics, Telkom University Bandung, Indonesia.
REFERENCES
[1] S. Rahimian Aghdam, S. S. Alizadeh, Y. Rasoulzadeh, and A. Safaiyan, “Fatigue Assessment Scales: A comprehensive literature review,” Arch. Hyg. Sci., vol. 8, no. 3, pp. 145–153, 2019, doi: 10.29252/archhygsci.8.3.145.
[2] P. Yin, L. Yang, C. Wang, and S. Qu, “Effects of wearable power assist device on low back fatigue during repetitive lifting tasks,” Clin. Biomech., vol. 70, pp. 59–65, 2019, doi: 10.1016/j.clinbiomech.2019.07.023.
[3] S. Park, S. Seong, Y. Ahn, and H. Kim, “Real-Time Fatigue Evaluation Using Ecological Momentary Assessment and Smartwatch Data: An Observational Field Study on Construction Workers,” J. Manag. Eng., vol. 39, no. 3, pp. 1–14,
2023, doi: 10.1061/jmenea.meeng-4953.
[4] U. Techera, M. Hallowell, and R. Littlejohn, “Worker Fatigue in Electrical-Transmission and Distribution-Line Construction,” J. Constr. Eng. Manag., vol. 145, no. 1, pp. 1–9, 2019, doi: 10.1061/(asce)co.1943-7862.0001580.
[5] F. K.-W. Wong, Y.-H. Chiang, F. A. Abidoye, and S. Liang, “Interrelation between Human Factor–Related Accidents and Work Patterns in Construction Industry,” J. Constr. Eng. Manag., vol. 145, no. 5, pp. 1–8, 2019, doi:
10.1061/(asce)co.1943-7862.0001642.
[6] M. Z. Liu, X. Xu, J. Hu, and Q. N. Jiang, “Real time detection of driver fatigue based on CNN-LSTM,” IET Image Process., vol. 16, no. 2, pp. 576–595, 2022, doi: 10.1049/ipr2.12373.
[7] M. Moshawrab, M. Adda, A. Bouzouane, H. Ibrahim, and A. Raad, “Smart Wearables for the Detection of Cardiovascular Diseases: A Systematic Literature Review,” Sensors, vol. 23, no. 2, pp. 1–25, 2023, doi: 10.3390/s23020828.
[8] IMO, “MSC.1/Circ.1598 Guidelines on fatigue,” Imo-Msc, vol. 44, no. June 2001, 2019, [Online]. Available:
https://www.register-iri.com/wp-content/uploads/MSC.1-Circ.1598.pdf.
[9] T. Theodoridis and J. Kraemer, “Evaluation of physiological metrics as a real-time measurement of physical fatigue in construction workers: State-of-the-Art Reviews,” vol. 147, no. 5, 2021, doi: http://dx.doi.org/10.1061/(ASCE)CO.1943- 7862.0002038.
[10] O. Bamisile et al., “Comprehensive assessment, review, and comparison of AI models for solar irradiance prediction based on different time/estimation intervals,” Sci. Rep., vol. 12, no. 1, pp. 1–26, 2022, doi: 10.1038/s41598-022-13652- w.
[11] A. Rai, A. Shrivastava, and K. C. Jana, “A CNN-BiLSTM based deep learning model for mid-term solar radiation prediction,” Int. Trans. Electr. Energy Syst., vol. 31, no. 9, pp. 1–13, 2021, doi: 10.1002/2050-7038.12664.
[12] S. Mehtab and J. Sen, “Stock Price Prediction Using Convolutional Neural Networks on a Multivariate Timeseries,”
2020, doi: 10.36227/techrxiv.15088734.v1.
[13] Z. Ahmad, M. N. Jamaludin, and K. Soeed, “Prediction of exhaustion threshold based on ECG features using the artificial neural network model,” 2018 IEEE EMBS Conf. Biomed. Eng. Sci. IECBES 2018 - Proc., pp. 523–528, 2019, doi:
10.1109/IECBES.2018.8626605.
[14] N. Juliana, I. F. Abu, N. A. Roslan, N. I. Mohd Fahmi Teng, A. R. Hayati, and S. Azmani, Muscle Strength in Male Youth that Play Archery During Leisure Time Activity. 2020.
[15] K. Wang et al., “Multiple convolutional neural networks for multivariate time series prediction,” Neurocomputing, vol.
360, pp. 107–119, 2019, doi: 10.1016/j.neucom.2019.05.023.
[16] G. Bilquise, S. Abdallah, and T. Kobbaey, Predicting Student Retention Among a Homogeneous Population Using Data Mining, vol. 77. 2021.
[17] Y. Zhang, M. Safdar, J. Xie, J. Li, M. Sage, and Y. F. Zhao, A systematic review on data of additive manufacturing for machine learning applications: the data quality, type, preprocessing, and management, vol. 34, no. 8. Springer US, 2023.
[18] P. C. Rodrigues, O. O. Awe, J. S. Pimentel, and R. Mahmoudvand, “Modelling the Behaviour of Currency Exchange Rates with Singular Spectrum Analysis and Artificial Neural Networks,” Stats, vol. 3, no. 2, pp. 137–157, 2020, doi:
10.3390/stats3020012.
[19] M. Ayitey Junior, P. Appiahene, and O. Appiah, “Forex market forecasting with two-layer stacked Long Short-Term Memory neural network (LSTM) and correlation analysis,” J. Electr. Syst. Inf. Technol., vol. 9, no. 1, pp. 1–24, 2022, doi: 10.1186/s43067-022-00054-1.
[20] M. Lepot, J. B. Aubin, and F. H. L. R. Clemens, “Interpolation in time series: An introductive overview of existing methods, their performance criteria and uncertainty assessment,” Water (Switzerland), vol. 9, no. 10, 2017, doi:
10.3390/w9100796.
[21] Ahmad Harmain, P. Paiman, H. Kurniawan, K. Kusrini, and Dina Maulina, “Normalisasi Data Untuk Efisiensi K-Means Pada Pengelompokan Wilayah Berpotensi Kebakaran Hutan Dan Lahan Berdasarkan Sebaran Titik Panas,” Tek. Teknol.
Inf. dan Multimed., vol. 2, no. 2, pp. 83–89, 2022, doi: 10.46764/teknimedia.v2i2.49.
