The application of machine learning algorithms in the estimation of production lead times A case stu

Abstract

Keywords: Machine learning; production time; random forest regressor; regression; steel construction.

Introduction

In the contemporary business environment, the efficiency of production processes is a crucial factor in achieving success in a competitive market. Accurate forecasting of production lead times offers numerous advantages, including the efficient use of resources, cost reduction, and increased

customer satisfaction. It also prevents delays in order deliveries. Therefore, estimating production lead times is of great importance for businesses [1]. All planning in the process, including cost, delivery, and other factors, is based on the realization time of the work steps [2]. In a manufacturing sector comprising a vast array of products, time studies are conducted on existing products to ascertain the time re-

*Corresponding author. *E-mail address: seyma.duymaz@rumeli.edu.tr Published by Yıldız Technical University Press, İstanbul, Türkiye This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).

quired for production. These times are incorporated into the calculation of the delivery date or the production planning program. One of the methodologies employed to estimate production times is machine learning algorithms. Machine learning algorithms are employed in a multitude of fields. Their applications extend to diverse domains, including forestry studies [3], the retail sector [4], land cover mapping [5], the clustering of court decisions [6], and the detection and diagnosis of chronic diseases [7]. The deployment of machine learning algorithms facilitates the identification of production failures, the estimation of remaining useful life, and the estimation of production times. This paper presents a review of selected studies in the literature. In a study by Lim et al. [8], support vector machines were employed to forecast production time based on work order and production data. The accuracy of the predictions was quantified at 84.62%. In a study by Schneckenreither et al. [9], an artificial neural network was employed to construct a flow time prediction model that could dynamically adjust lead times. In their study, Yüce et al. [10] employed artificial neural network, support vector regression, and gradient boosting algorithms from the field of machine learning to estimate production times in specific production areas of a production facility. In their study, Haeussler and colleagues [11] examined workload control models with fixed and dynamic lead times. In a related study, El Mekkaoui et al. [12] employed artificial neural networks and a random forest algorithm to predict the arrival time of ships. In a previous study, Agwu et al. [13] employed an artificial neural network to predict oil flow rates. In a study on production time estimation in a construction factory, Alsakka et al. [14] employed a methodology based on machine learning. Gyulai et al. [15] utilized machine learning methodologies to estimate production time, thereby facilitating informed decision-making in a manufacturing context. Chen et al. [16] employed machine learning in manufacturing studies. Dehghani et al. [17] employed a random forest algorithm to predict shear wave velocity. In this study, for different product types of a manufacturing company producing in the field of steel construction; 1 The objective is to determine the time required to produce 99 distinct products on a single machine, each with a defined production time, 2 The objective is to calculate production times for products with varying length, volume, and weight values that have not yet been produced. This approach enables the provision of more precise delivery time estimates to customers without the necessity of manufacturing the products. This study makes a significant contribution to the existing literature by providing insight into the time required to produce a diverse range of products by a company engaged in steel construction. The study also identifies the time required to deliver these products to the customer. Furthermore, a dynamic production planning schedule can be prepared by enhancing the efficiency of machine utilization. The utilization of machine learning algorithms will facilitate the estimation of production times. This study will enable the analysis of higher-dimensional data from production and

the automatic calculation of the duration of instant orders. In contrast to previous studies, a data set comprising products that have never been manufactured was subjected to analysis. The model offers the company a number of benefits. Firstly, it will enable the company to ascertain the delivery times for instant orders. Secondly, it will allow the company to assign parts to machines at the optimal level. This will facilitate the creation of production planning schedules. The calibration time of the machines will be reduced. In the Materials and Methods section of the study, the algorithms and performance metrics utilized are delineated. In the Results and Discussion section, the data is described and the application is made to a data set derived from a steel construction company. In the final section, the results of the study are interpreted and practical benefits are enumerated.

Materials And Methods

Machine learning algorithms are employed in a multitude of sectors. A number of algorithms exist, including those based on artificial neural networks, support vector machines, random forests, decision trees, and gradient boosting. In this study, the random forest, decision tree, and gradient boosting algorithms are employed. As the problem addressed in this study is a forecasting problem, regression techniques are utilized. Random Forest Algorithm The random forest algorithm is based on tree-based models, which are a type of supervised machine learning algorithm. Tree-based models involve recursively partitioning the dataset into two groups, depending on a stopping condition. Each node depends on the previous node. This algorithm is useful for both classification and regression problems. In classification problems, the splitting criterion is determined by entropy calculation. In regression problems, the most commonly used splitting criterion is the mean squared error at each internal node, as described in [18]. The Random Forest Regressor is an ensemble model comprising a multitude of decision trees. Its function is to reduce variance by averaging the results. When forecasting for a given data set, it utilises the prediction of each decision tree and averages these predictions to create the final prediction. The tree structure grows incrementally and is averaged at each step. It is an algorithm that is sensitive to overlearning. Different examples can be generated from one data set [19]. Figure 1 illustrates the structure of the decision tree, which is a self-iterative structure. The random forest algorithm is more accurate than decision trees in estimating the error rate. Algorithm steps; • The data set to be used in the study is prepared, • Trees are created for the samples and the results are predicted, • Averaging for the regression problem • The process continues recursively according to the stopping criterion as the most important factors.

Figure 1. Structure of decision tree [20]. In this study, the Anaconda Jupyter interface was employed. The random forest regressor algorithm, one of the machine learning algorithms in scikit-learn, one of Python's open source libraries, was utilized because the problem at hand is a prediction problem. Decision Tree Algorithm Decision trees represent a supervised learning algorithm that automates the decision-making process by identifying the optimal solution from a set of alternatives. The decision tree algorithm is a nonparametric prediction model that can be used to address both regression and classification problems [21]. In the case of categorical final values, the problem is classified, whereas in the case of continuous values, it is a regression problem [22]. The algorithm begins by extracting examples from the data set and then subdividing them into subclasses. In essence, decision trees make predictions by dividing the data into branches according to their characteristics. This approach entails dividing the data set into a tree structure, with each node representing a specific value of a feature [23]. Leaf nodes contain continuous values, and the tree makes predictions based on the average of these values. The structure of the decision tree consists of a root node, an internal node, and leaves. The root node is the first node and contains the entire data set. The internal node represents the data set partitioned into multiple subsets [24]. The leaves are the final predictions. In this study, the decision tree algorithm is employed to estimate the production time. Gradient Boosting Algorithm Gradient Boosting Regressor is a powerful and flexible machine learning algorithm for regression problems. It is a decision tree-based algorithm. It is an iterative variant of sequentially organized tree models. It first runs the mod-

el, detects errors, then runs it again and so on iteratively. It continues in a stronger way by learning from the previous step [25]. Each new decision tree generated by the Gradient Boosting algorithm is based on the principle of minimizing the errors calculated in the previous tree [26]. This results in a more accurate outcome by correcting the errors in a sequential manner. Gradient Boosting employs gradient descent at each step to minimize the errors. The algorithm functions according to the principle of transforming learners with limited abilities into those with enhanced capabilities. It can be described as an ensemble algorithm. Steps of the algorithm; • Examine the compatibility of the equation coefficients with the data by defining the loss function, • Determining the state where the loss function is minimum by determining the fixed variable, • Calculation of errors, • Estimation for each observation, it is listed in four articles [27]. Performance Metrics Mean square error (MSE): The metric is employed in the context of regression problems and represents the average of the squares of the differences between actual and predicted values. As this value approaches zero, the performance of the model improves.

Root mean square error (RMSE): It is obtained by taking the square root of the root mean square error. The RMSE value shows the closeness of the estimates to the ac-

tual values. It is a second-order error metric that measures the magnitude of the error. RMSE is the standard deviation of the difference between actual and predicted values. An RMSE value close to zero means good performance [28].

Mean absolute error (MAE): The sum of the absolute values of the differences between the actual value and the predicted value.

Mean absolute percentage error (MAPE): It expresses the error between actual values and predicted values expressed as a percentage. The closer it is to 0%, the more meaningful the results [29].

R square (R2): A performance metric describing the accuracy of the model.

yi=Actual values yi=Estimated values N=Number of observations

Results And Discussion

Identification of Data This study was conducted using a data set obtained from a company engaged in the production of steel structural products. The company provided input values for 99 finished products of varying sizes, diameters, and weights, as well as the time required to produce them on one machine. The data set is presented in Table 1. The size, diameter, and weight were used as input data, while the production time was used as an output value for estimation purposes. Table 2 illustrates the data types and indicates that the data set does not contain null values. The analysis of data sets with empty values is a challenging endeavor. Table 3 presents the results of the statistical calculations. It was observed that the products with the lowest standard deviation values were weight values. The highest standard deviation was observed between production times. In this case, it can be concluded that there is not much difference between weight values, but there is more difference between production times.

the production times of products with comparable lengths is considerable. This is due to the influence of additional variables on the time required. Figure 4 illustrates the binary relationships between the variables. Figure 5 shows the distribution of the products produced in terms of size, diameter, weight and production times.

Figure 2. Distribution of material name. Visualization of the Data Set The Autoviz library, one of the open-source libraries in Python, was employed for the purpose of data visualization. Figure 2 illustrates the frequency graph of eight distinct product types. A total of 99 unique products belonging to eight different product types were produced. The most prevalent product type in this instance is Metal Shaft Semi-Finished Product. Figure 3 illustrates the distribution of size, diameter, and weight values according to production times. It can be observed that in certain instances, the discrepancy between

Case Study In this study, the Random Forest Regressor, Decision Tree Regressor, and Gradient Boosting Regressor algorithms were employed. The results are presented in Table 4. In addition to the three distinct algorithms employed, the gradient boosting regressor algorithm, which yields the optimal outcome, is utilized once more through cross-validation (CV). The objective here is to enhance performance. In CV, the data set is partitioned into multiple training and test sets, the model is trained and tested on these sets, and the results are averaged to obtain more reliable performance metrics. This process allows for the identification and rectification of issues such as overfitting and underfitting. The Gradient Boosting Regressor (CV) model demonstrated superior performance relative to the other models. In particular, the mean absolute percentage error (MAPE)

Figure 3. Scatter plot of each continuous variable vs target. Table 4. Results

MAE: Mean absolute error, MSE: Mean square error, RMSE: Root mean square error, MAPE: Mean absolute percentage error.

Figure 4. Pair-wise scatter plot of all continuous variables.

Figure 5. Distribution of products according to variables. value of 19.25% was notably lower than the other models, indicating that the model's predictions were highly accurate. The Random Forest Regressor and Decision Tree Regressor models exhibit high R2 values, yet simultaneously display high MAPE values. This indicates that the model is prone to making significant errors in certain instances. The Gradient Boosting Regressor model also demonstrates satisfactory performance, with a high R2 value and low MAE, MSE, and RMSE. However, the MAPE value is relatively high, indicating that the model may be prone to significant errors.

In conclusion, the Gradient Boosting Regressor (CV) model is the most effective in terms of overall performance. In the future, it would be beneficial to investigate and implement methods to reduce the MAPE values. The algorithm's accuracy, estimated at approximately 96.9%, was utilized to create a dataset comprising product features that had never been produced. The dataset was subjected to testing and prediction in order to ascertain the estimated time required for the production of the products in question. The results are presented in Table 5.

Conclusion

The estimation of production lead times is of significant importance for the purposes of planning and scheduling, cost control, resource management, inventory management, and customer satisfaction. In companies with a high product variety, the only way to determine the time required to produce each new order is to produce the product once. This necessitates the readjustment and recalibration of machines to facilitate the production of the new order. Consequently, communicating lead times to customers represents a significant challenge. It is of great importance to provide accurate lead times in order to ensure customer satisfaction. In this study, the estimated accuracy of production lead times for existing products was found to be approximately 96.9%. Nevertheless, forecasting the present circumstances will not assist the company. The company has already attempted this and has gained insight into the time required to produce existing products. The primary objective of this study is to predict the time required to produce a previously unproduced product order. The dataset provided by the company was subjected to analysis using machine learning algorithms, resulting in a highly accurate prediction. The study will enable the company to identify the production times of products with varying parameters. This will facilitate the estimation of production times. Upon analysis of the algorithms, it becomes evident that reliable predictions can be made. These predictions can be utilized by the company in a number of ways, including predicting delivery times for products ordered, notifying customers of delivery times, and creating production planning schedules. This study contributes to the use of machine learning algorithms in the estimation of production times. A dynamic scheduling can be created by optimizing the use of machines. These studies are of great importance to companies that wish to minimize costs and maximize customer satisfaction.

Data Availability Statement

The authors confirm that the data that supports the findings of this study are available within the article. Raw data that support the finding of this study are available from the corresponding author, upon reasonable request. Author’s Contributions Şeyma Duymaz: Modeling, Analyzing, Writing, Interpreting Results. Ali Fuat Güneri: Obtaining the data set, Review-editing.

Conflict of Interest

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Use of AI for Writing Assistance Not declared.

Ethics

There are no ethical issues with the publication of this manuscript.

Share and Cite