نوع مقاله : مقاله پژوهشی

نویسندگان

• فائزه زمانی ¹
• احمد ابراهیمی ²
• رویا سلطانی ³
• بابک فرهنگ مقدم ⁴

¹ دانشجوی دکتری رشته مدیریت صنعتی، واحد علوم و تحقیقات ،دانشگاه آزاد اسلامی ، تهران، ایران

² استادیارگروه مدیریت صنعتی و تکنولوژی،واحد علوم و تحقیقات ، دانشگاه آزاد اسلامی ، تهران، ایران ، نویسنده مسئول: Ahmad.Ebrahimi@srbiau.ac.ir

³ استادیارگروه مهندسی صنایع، دانشگاه خاتم، تهران، ایران

⁴ دانشیار موسسه مطالعات برنامه‌ریزی و مدیریت، تهران، ایران

چکیده

هدف این پژوهش بررسی عوامل موثر در پیش‌بینی زمان انتظار و ایجاد مدل پیش‌بینانه زمان انتظار سفارشات کانبان به جهت بهبود پایداری و تاب‌آوری در زنجیره تامین ناب می‌باشد. برای دستیابی به این هدف، مطالعه از روش داده‌کاوی پیروی می‌کند، مجموعه داده‌ها شامل 103023 مشاهده، ازسیستم کانبان واکسترانت زنجیره تامین با رعایت الزامات شاخص‌های کیفیت دیتاست در بازه 6/1402 تا 11/ 1402 استخراج شده است. ابتدا شاخص‌های موثر بر زمان انتظار سفارشات استخراج شده است و به جهت بهبود عملکرد و دقت پیش‌بینی، از فرآیندکاوی جهت شناسایی متغیرهای پرتکرار و تاثیرگذار در واریانت‌های اصلی و سپس در مرحله برازش مدل، از رویکرد تحلیل گام‌به‌گام تلفیقی جهت انتخاب ویژگی‌ها و از تنظیم پارامتر رویکردهای رگرسیونی ناپارامتریک استفاده شده است. مدل پیش‌بینانه با استفاده از مدل‌های رگرسیونی خطی چندمتغیره، چندمتغیره دارای انحنا، لاسو، الاستیک‌نت، درخت تصمیم تقویتی، جنگل تصادفی بوت‌استرپ، k- نزدیک‌ترین همسایه، شبکه عصبی تقویتی برازش داده شده است. عملکرد مدل‌های رگرسیونی برازش شده با استفاده از شاخص‌های ارزیابی R^2 ، RASE و اعتبارسنجی نتایج و مدل تایید شده است. نتایج نشان داد که عوامل لجستیکی در زمان انتظار سفارشات موثر بوده و الگوریتم شبکه عصبی تقویت شده بهترین مدل در پیش‌بینی زمان انتظار سفارشات با دقت 96 درصد و با خطا 84/5 است. سپس قابلیت پیش‌بینی مدل برای دیتاهای جدید در سیستم صدور سفارشات کانبان به کار گرفته شده است، نتایج و بهبودهای حاصل از بهره‌گیری قابلیت‌های داده‌کاوی در سیستم کانبان همگی بیان‌گر تاثیر معنی‌دار ترکیب ابزار ناب و یادگیری ماشین به جهت توانمندسازی و تاب‌آوری زنجیره تامین ناب می‌باشد.

کلیدواژه‌ها

• یادگیری ماشین
• رگرسیون
• مدیریت زنجیره تامین ناب
• کانبان
• زمان انتظار

موضوعات

• هوش مصنوعی وکاربرد آن در مدیریت

عنوان مقاله [English]

Predicting the Lead time of auto parts orders in the supply chain using machine learning

نویسندگان [English]

• Faezeh Zamani ¹
• Ahmad Ebrahimi ²
• Roya Soltani ³
• Babak Farhang Moghaddam ⁴

¹ Ph.D. Student of Industrial Management Group, Department of Economics and Management, Science and Research Branch, Islamic Azad University, Tehran, Iran

² .Assistant Professor of Industrial Management & Technology Group, Department of Economics and Management, Science and Research Branch, Islamic Azad University, Tehran, Corresponding Author: IranAhmad.Ebrahimi@srbiau.ac.ir

³ Assistant Professor of Industrial Engineering Group, Department of Engineering, Khatam University, Tehran, Iran

⁴ Associate Professor, Institute Management and Planning Studies, , Associate Professor, Institute Management and Planning Studies, Tehran, IranIran

چکیده [English]

This research aims to investigate the effective factors in predicting lead time (LT) and create a predictive model of LT to improve sustainability and resilience for Kanban orders in the lean supply chain (LSC). The study follows the data mining (DM) method, and the dataset includes 103023 observations from the Kanban system, which were extracted in compliance with the requirements of the dataset quality indicators in the period 1402/6 to 1402/11. First, indicators affecting the LT of orders were extracted. Process mining was used to identify influential variables in high-variance processes to improve performance and accuracy. A stepwise analysis approach was used to select features for the model fitting stage. Also, tuning the parameters of non-parametric approaches was used. The predictive model uses Multiple Linear Regression, Multiple with curvature, Lasso, Elastic Net, Boosted Decision Tree, Bootstrap Random Forest, K-Nearest Neighbor, and Boosted Multi-Layer Perceptron. The performance of the fitted regression models has been confirmed using R^2, RASE, and validation of the results and model. The results showed that the logistical features are effective in LT, and the Boosted Multi-Layer Perceptron is the best for predicting orders' LT with an accuracy of 96% and an error of 5.84. Using the model's predictive capability for new data in the Kanban system, the results obtained within four months have been used. The improvements from using DM capabilities in the Kanban system all express the significant impact of combining lean and machine learning (ML) tools to empower and resilient Lean Supply Chain Management (LSCM).

Introduction

The main problem in this research is identifying the factors that effectively predict the LT of orders in the LSC, choosing the best ML algorithm for predicting the exact LT, and how process mining can effectively identify the most repeatable variables in the main variants and investigate how DM can reduce waste in LSC.
Despite classification studies on risk, disruption, and delay prediction in the literature, to our knowledge, fewer articles were found regarding the use of DM to predict the accurate LT of orders in the LSC with logistical features. Also, according to researchers, DM is considered a tool to overcome the limitations of lean tools and strengthen their performance. However, the studies corresponding to the executive case did not observe the results and improvements from the ML application in predicting the LT of orders.
Therefore, in this research, in terms of innovation, 1) machine learning has been used to accurately predict the LT of Kanban orders, considering logistical factors, 2) Process mining has been used in the identification stage of influential variables, 3) The results and improvements obtained from predicting the LT of orders regarding risk reduction and sustainability improvement have been examined and compared.
Research Question(s)
The main questions in this research are specified as follows:

What factors affect LT's prediction in the lean supply chain?
How do we predict the LT in the lean supply chain?
How can DM effectively reduce waste in the lean supply chain?
Literature Review

Regarding the issue's importance and urgency, transparency and accurate prediction of the LT have reduced risk and improved sustainability and resilience in the LSC. These effects are significant in both theoretical and operational dimensions, such as reducing logistic costs, safety stock, working capital, stoppage, level of inventories, storage cost, energy consumption, and risk. After reviewing the literature, the most relevant articles in the field of ML are listed in Table2.

Methodology

This research is practical from the objective point of view, and from the data point of view, it is quantitative. This study includes four main processes: 1) reviewing the literature and data collection, 2) research method and pre-processing, 3) model construction, and 4) model evaluation and results (Jayanti, 2022 & Wasesa). First, influential variables were extracted by reviewing the literature. Then, the dataset was extracted from the Kanban system in compliance with the requirements of the data set quality indicators from 6/1402 to 11/1402. Then, process mining was used to identify the features with the most repeatability in the main variants, and finally, influential variables were extracted through brainstorming. An integrated stepwise analysis approach has been used to select features. The predictive model uses MLR, curvature, Lasso, Elastic Net, Boosted DT, Bootstrap RF, KNN, and Boosted Multi-Layer Perceptron. The parameters of non-parametric approaches are tuned to improve forecasting performance and accuracy. In this research, evaluation and validation are the main criteria for evaluating the model's predictive power, and error and accuracy indices have been used together. Therefore, the performance of the fitted regression models using R^2 and RASE evaluation indices and validation of the results and the model are confirmed.

Results

After fitting the regression models, for each row of test data, predict the LT and compare it with the actual values of the LT; then, to identify the best model, R^2, RASE, and model comparison approaches are used.
The results show that the Boosted Multi-Layer Perceptron, with one hidden layer, five activation functions, and a learning rate of 0.1, has the highest accuracy at 96% and the lowest root average square error at 5.84, compared to other fitted models.

Discussion and Conclusion

The obtained results show that the identified independent variables are related to customer factors (safety stock), manufacturer factors (inspection status, quality paint), logistic factors (vehicle, distance), part factors (name, part-expert), and order factors (number of holidays, Kanban issue date) are effective on the LT. As the selected model in this research, the regression model of the Boosted Multi-Layer Perceptron has the highest R^2 and the lowest RASE criteria. Process mining is practical and helpful in identifying the main variants. By using the model's predictive capability for new data in the Kanban order issuing system within four months, the improvements all express the significant impact of combining lean tools and ML to empower LSCM. The practical implications of this research can guide managers in implementing practices with lean tools, improving sustainability, eliminating waste, and being more competitive in the current challenging business environment. Academics can benefit from the present study because it provides ML practices that can be further tested and validated.
This research generalizes and develops the use of DM as a decision-making support tool in predicting the LT to overcome the limitations of lean tools, and it can improve the efficiency and stability of the LSC and reduce the risk. While this research provides valuable insights, it also has limitations, including the lack of data on influential variables identified in the literature. In implementing this research, there are suggestions for future research that examine factors such as production capacity, weather, and location conditions and deep learning to fit more reliable and accurate results and investigate prescriptive analyses to optimize the LT of orders based on the fitted regression models, the design of the experiment and using the profiler's capabilities.
Keywords: Machine Learning, Regression, Lean Supply Chain Management, Kanban, Lead Time.
 

کلیدواژه‌ها [English]

• Machine Learning
• Regression
• Lean Supply Chain Management
• Kanban
• Lead Time