Forecasting Return Quantity, Timing and Condition in Remanufacturing with Machine Learning: A Mixed-Methods Approach

Abstract

Remanufacturing plays a key role in the circular economy by reducing material consumption and extending product life cycles. However, a major challenge in remanufacturing is accurately forecasting the availability of cores, particularly regarding their quantity, timing, and condition. Although machine learning (ML) offers promising approaches for addressing this challenge, there is limited clarity on which influencing factors are most critical and which ML approaches are best suited to remanufacturing-specific forecasting tasks. This study addresses this gap through a mixed-method approach combining expert interviews with two systematic literature reviews. The interviews with professionals from remanufacturing companies identified key influencing factors affecting product returns, which were structured into an adapted Ishikawa diagram. In parallel, the literature reviews analyzed 125 peer-reviewed publications on ML-based forecasting in related domains—specifically, spare parts logistics and manufacturing quality prediction. The review categorized data sources into real-world, simulated, and benchmark datasets and examined commonly applied ML models, including traditional methods and deep learning architectures. The findings highlight transferable methodologies and critical gaps, particularly a lack of remanufacturing-specific datasets and integrated models. This study contributes a structured overview of ML forecasting in remanufacturing and outlines future research directions for enhancing predictive accuracy and practical applicability.

1. Introduction

The growing demand for raw materials, driven by rising production volumes and shorter product life cycles, has reinforced the urgency of transitioning toward a circular economy. In response, the European Union has prioritized reducing the use of primary materials by introducing the Industrial Clean Deal, which mandates that by 2023, 24% of the raw materials must be circular materials [1], while also promoting value retention strategies such as reuse, repair, and remanufacturing at the end of product life [2]. Among these, remanufacturing as a key process of the circular economy is an industrial process to return a used product, hereafter referred to as core, to “at least its original performance, with a warranty that is equivalent or better than that of the newly manufactured product” [3]. Research has shown that remanufacturing can reduce greenhouse gas emissions by 79–99%, energy usage by up to 55%, and material consumption by up to 50% compared to conventional new production [4,5,6,7,8]. In addition to its environmental benefits, remanufacturing also has notable economic advantages, as remanufactured products are typically sold at lower prices than newly manufactured equivalents [9].

The remanufacturing process typically consists of five steps: (1) disassembly, (2) cleaning, (3) inspection and sorting, (4) repair or replacement, and (5) reassembly, each of which is supported by quality control measures [6]. Therefore, the efficient execution of these steps depends mainly on accurate operational and tactical planning, particularly regarding the availability and condition of incoming cores. However, return flows are influenced by numerous interdependent factors, including customer usage behavior; product life cycle dynamics; business models; the core acquisition strategy, hereafter referred to as take-back strategy; return logistics; and market conditions [9]. These factors collectively contribute to significant variability in the quantity, timing, and condition of returned cores, creating a high degree of uncertainty in forecasting processes [10].

Traditional forecasting techniques, such as linear regression analysis (R), moving average (MA), and exponential smoothing (ES), typically based on deterministic or rule-based models, often struggle to capture these uncertainties with their nonlinear and heterogeneous patterns present in real-world remanufacturing systems [11,12]. For example, these techniques may fail to account for sudden spikes in returns due to product recalls, unusual usage patterns across customer segments, or variable core quality influenced by environmental conditions. As a result, these methods frequently lead to inaccurate predictions of core availability, causing disruptions and inefficiencies in production planning and execution [13,14].

In contrast, ML offers promising capabilities by enabling data-driven predictions based on large and diverse datasets. ML models, e.g., decision trees (DTs), support vector machines (SVMs), or artificial neural networks (ANNs), are capable of capturing nonlinear relationships, managing missing or noisy data, and learning from complex historical patterns, which makes them particularly suitable for forecasting return quantity, timing, and condition [14,15]. However, despite growing academic interest in this area, there is still limited understanding of which influencing factors are most critical and which ML models are best suited for forecasting in the specific context of remanufacturing.

Previous research has explored various forecasting methods in related contexts. Early studies, such as those by Goh and Varaprasad [16] and Kelle and Silver [17], proposed statistical models to estimate returns of reusable containers. While foundational, these works differ from the complexities inherent in remanufacturable products. Guide [11] highlighted this distinction, emphasizing the added uncertainty in return quantity and timing. Subsequent studies by Brito and van der Laan [12] examined the cost implications of different forecasting models under imperfect data, showing that more data does not always yield more accurate results.

Hybrid approaches have been developed to overcome data limitations. Marx-Gomez et al. [18] applied a combination of simulation, fuzzy logic, and neuro-fuzzy systems to predict returns of photocopiers. Clottey et al. [19] and Geda and Kwong [20] advanced time-lag-based forecasting using distributed lag models, with the latter incorporating Bayesian inference for parameter estimation. Additionally, methods like ARIMA and Holt–Winters have been used for seasonal forecasting [21], while neural networks have been applied to predict the return quantity of heavy machinery [15].

While forecasting the condition of returned cores has received comparatively less attention than quantity and timing in the remanufacturing literature, several studies have addressed this challenge, often through core quality assessment or grading frameworks. For instance, Ponte et al. [22] and Stamer and Sauer [23] explore methods for classifying the condition of returned products to support remanufacturing decisions. These approaches typically involve categorizing cores into predefined quality levels (e.g., usable, repairable, or scrap) based on visual inspection, historical usage data, and sensor-based condition indicators.

More strategic, long-term forecasting approaches have also emerged in recent years, particularly in the context of electric vehicle (EV) batteries. Liang et al. [24] present a forecasting model that estimates the quantity and quality of returned EV batteries by integrating product sales, customer behavior, usage intensity, and expected service life. These variables are modeled as separate probability distributions and combined using a convolution-based approach to capture their interdependencies.

Similarly, Huster et al. [25] examine long-term return forecasting for remanufacturing capacity planning. Their study focuses on how different forecasting assumptions, such as variations in service time, return rates, and quality levels, affect planning outcomes. They highlight the need for scenario-based modeling to handle the inherent uncertainties in core returns and stress the value of quality-related forecasting in aligning remanufacturing capacity with actual return flows.

Despite the relevance of these studies, their isolated focus makes it difficult to generalize findings across remanufacturing contexts. It is still unclear which influencing factors are decisive and which ML models are particularly suitable for use in remanufacturing. Against this background, the following research questions (RQs) arise, which are addressed in this study:

• RQ1: Which data characteristics are commonly used to train machine learning models for forecasting the availability of returned cores in remanufacturing? This question identifies the types of input variables (features) commonly selected to develop predictive models. Understanding the nature and relevance of these features is essential for designing accurate and practically implementable models in real-world remanufacturing environments.
• RQ2: Which data sources are commonly used for training machine learning models for forecasting the availability of returned cores in remanufacturing? This includes exploring the origins of the datasets used to train and evaluate models. Clarifying the advantages and limitations of each data source helps assess their transferability to remanufacturing applications, where data availability is often limited or inconsistent.
• RQ3: Which supervised machine learning models are commonly applied to forecast return quantity, timing, and condition of cores? The aim is to analyze the types of supervised learning problems addressed in the literature and assess the prevalence and suitability of different ML algorithms.

Based on these research questions, this study aims to derive key insights and highlight open challenges associated with forecasting in remanufacturing. Moreover, an outlook on future research directions is provided to support the broader adoption of ML-based forecasting approaches in academia and industry.

The remainder of the study is structured as follows. Section 2 describes the methodology, including expert interviews used to identify relevant influencing factors and systematic literature reviews undertaken to evaluate applicable ML models. Section 3 presents the empirical and literature-based results of the three research questions. Section 4 discusses these findings in the context of remanufacturing challenges and opportunities. Section 5 concludes the study and provides suggestions for future research.

2. Materials and Methods

This study applies a mixed-method approach, combining an empirical–exploratory component with a systematic literature review. Semi-structured expert interviews were conducted with professionals from remanufacturing companies to gain practical insights into the influencing factors affecting the predictability of return quantity, timing, and condition. In addition, a systematic literature review was carried out to assess the current research on machine learning methods used for forecasting in remanufacturing and adjacent domains.

2.1. Expert Interviews

As a first step, expert interviews were used to gather domain-specific insights on factors influencing core returns, directly addressing RQ1. A qualitative empirical approach was adopted using semi-structured expert interviews to investigate these practical influencing factors. This method is well-suited for accessing context-specific, experience-based knowledge and is commonly used in exploratory research [26,27]. The semi-structured format enables consistency across interviews while also allowing for flexible, open-ended exploration of case-specific insights [28]. A stimulus-based format was employed to enhance validity: predefined influencing factors were presented to the participants for assessment and commentary.

These influencing factors were identified beforehand through a brainstorming session aimed at compiling a broad range of potentially relevant variables. This approach is instrumental in expert settings where time is limited but depth of response is essential [29]. Participants were also encouraged to suggest additional factors not covered in the predefined list.

Five interviews were conducted with professionals responsible for evaluating and processing returned cores, illustrated in

Table 1. Participants in expert interviews.

Expert	Function	Sector
E1	Product management	Commercial vehicle
E2	Quality management	Rail transport
E3	Business development	Passenger vehicle
E4	Product management	Commercial vehicle
E5	Technical project manager	Bike systems

. Participants were selected via purposive sampling to ensure relevant domain expertise. Data saturation was observed after the fourth interview, as recurring themes consistently emerged. The fifth interview confirmed the saturation of key insights, indicating that additional interviews were unlikely to yield substantially new information. All interviews were conducted online using video conferencing tools. Audio recordings were not made to create an open and informal environment. Instead, a third-party observer documented detailed, real-time notes that were structured in a result-oriented format. While this approach supported open dialogue, it also introduced potential limitations, such as observer bias or incomplete data capture. To mitigate these risks, the interviewees subsequently validated the notes by reviewing and confirming the interview notes, ensuring objectivity and corresponding quality.

2.2. Literature Review

A structured literature review was conducted to systematically assess the application of machine learning models for forecasting return quantity, timing, and condition in remanufacturing. Various models for literature review processes exist in the academic domain, such as those proposed by Bandara et al., Levy & Ellis, Li & Wu, and vom Brocke et al. [30,31,32,33]. While these models may vary in detail, they typically follow a three-phase structure: planning, conducting, and reporting [34]. The literature review conducted in this study is based on a generic process model proposed by Brendel et al., which synthesizes various established review models [35]. This model refines the traditional three-phase structure by dividing it into six systematic steps: goal definition, definition of review scope, search, analysis, synthesis, and discussion. The latter model is used for the literature review in this study and visualized in Figure 1.

Following the model proposed by Brendel et al. [35], two separate literature reviews were conducted to identify relevant machine learning models applied in core forecasting in remanufacturing. One review focused specifically on models used to forecast return quantity and timing, while the other addressed models related to the prediction of core condition.

As part of the preparation, the aim of both literature reviews is first defined, the research area is narrowed down, and the research question is precisely formulated [35]. Both reviews aimed to identify and critically evaluate peer-reviewed publications applying supervised machine learning techniques to forecasting in remanufacturing or related domains. The analysis concentrates on the types of ML models employed, the nature of the data sources, and the potential transferability of existing approaches to remanufacturing contexts.

Given the limited number of direct applications of ML in remanufacturing, a transfer strategy was adopted:

• For return quantity and timing, the scope was extended to include publications on spare parts demand forecasting due to structural similarities such as uncertain return flows, failure rates, and non-regular demand patterns.
• For condition prediction, the review focused on studies in manufacturing that apply ML to quality, defect, and fault prediction, based on the conceptual overlap with assessing the condition of returned parts.

This approach enabled a broader, yet targeted, exploration of machine learning applications with potential relevance to remanufacturing.

According to Banker and Kauffman, the focus of the research area for these literature reviews is placed on the research direction of decision support and design science [37], aligning the reviews with RQ2 and RQ3 defined in the “Introduction” section.

2.2.1. Forecasting Return Quantity and Timing

The search was conducted in February 2025 across four major academic databases: IEEE Xplore, ScienceDirect, Scopus, and Wiley Online Library. Boolean operators were used to construct precise title-based queries. Due to the scarcity of literature directly focused on remanufacturing, the search included publications on spare parts demand forecasting, given the similarities in logistical uncertainty and forecasting requirements. The exact query is illustrated in Figure 2.

During the literature search, the previously identified journals and databases were systematically searched for relevant publications [35]. The initial search resulted in a data set of 194 publications. The publications found were then subjected to a formal and content analysis, sorting out unsuitable results using defined exclusion criteria. During the formal screening process, 54 duplicates were eliminated, and 37 publications based on predefined criteria (e.g., language, time horizon, and scientific nature) were excluded. The remaining 103 entries underwent a content-based relevance assessment, regarding RQ2 and RQ3, following the approach by Booth et al. [38], involving a three-stage screening of titles, abstracts, and full texts. This process led to the exclusion of 31 publications: five during the title screening, 19 during the abstract review, and seven after full-text evaluation, resulting in a final dataset of 72 studies. The selection process is documented in Figure 3, based on the “Preferred Reporting Items for Systematic reviews and Meta-Analyses” (PRISMA) statement [39].

In the synthesis phase, the selected publications were analyzed in relation to the types of ML models employed, the data set source, and the context of their application. A qualitative content analysis was conducted to identify recurring modeling strategies and trends relevant to forecasting in remanufacturing. These findings are presented in detail in Section 3.

2.2.2. Forecasting the Condition of Cores

A second literature review was conducted to investigate supervised learning methods for predicting the condition of cores, focusing on approaches analogous to remanufacturing. This review followed the same six-step structure as the previous one, ensuring methodological consistency.

As before, the review targeted peer-reviewed articles published after 2015 in English or German. The databases and search logic remained unchanged. However, due to the lack of remanufacturing-specific studies on condition prediction, the scope was extended to include publications on quality, defect, and fault prediction in manufacturing. This methodological transfer assumes that analogous challenges in production environments can inform condition prediction upon return. The specific search query is depicted in Figure 4.

The search yielded 250 publications. Following a similar screening approach, 74 duplicates and 42 formally non-compliant publications were excluded. The remaining 134 entries were assessed in three steps—titles, abstracts, and full texts—leading to the exclusion of 81 studies. In total, 53 publications were included in the analysis. The selection process is visualized in Figure 5 using the PRISMA statement [39].

In the synthesis and discussion phases, these publications were analyzed to identify the types of machine learning models used, the learning tasks (e.g., classification or regression), and the context of their application. The results form the basis for a structured evaluation of condition forecasting methods in Section 3.

3. Results

This chapter presents the findings from the expert interviews and systematic literature reviews conducted in this study. The objective is to identify key influencing factors affecting the return quantity, timing, and condition of cores in remanufacturing, and to evaluate which machine learning models are most commonly applied for predictive tasks in this domain.

3.1. Identification of Influencing Factors of Return Quantity, Timing, and Condition of Cores

To systematically identify and categorize the factors influencing return quantity, timing, and condition in remanufacturing, insights from expert interviews were consolidated using a cause-and-effect diagram, commonly known as an Ishikawa or fishbone diagram. This visualization method, originally developed by Kaoru Ishikawa, is particularly well-suited for revealing complex interdependencies between variables [40].

The Ishikawa diagram groups the influencing factors into five adapted categories: People, Material, Environment, Returns, and Product, which are derived from the traditional 5-M method (Man, Machine, Material, Method, and Mother Nature) [41]. This structure was modified to better reflect the specific context of remanufacturing, as shown in Figure 6. The resulting categorization, particularly the distribution of attributes across categories, helps to identify underrepresented areas and supports a more structured and comprehensive brainstorming of additional influencing factors.

The People category includes all human-related influences, encompassing both customers and users on the one hand, and production and logistics personnel on the other. The customer’s behavior directly affects the return timing and condition of a product through several steps: the purchase decision, the intensity and duration of product use, and ultimately, the decision to return it [24]. Improper use, delayed return, or failure to return the part altogether can drastically impact the predictability of returns. In addition, production employees also play a critical role. Errors during manufacturing, handling, or transport can negatively influence a product’s lifespan and return condition. Interview participants repeatedly highlighted the importance of staff training and procedural adherence in minimizing unplanned wear or damage prior to the return of cores.

The Material category refers to the quality, durability, and sourcing of the input materials used in the original product. Poor material quality or variability can reduce a part’s reliability, leading to premature failure, affecting the return condition, and potentially triggering a customer-initiated return. Thus, while material aspects are primarily associated with product condition, they can also indirectly influence return timing and volume through their impact on failure rates.

The Environment category encompasses external factors, including climatic conditions, market dynamics, and broader macroeconomic trends. Experts emphasized that environmental conditions (e.g., temperature, humidity, geographic location) significantly affect product degradation and customer usage behavior. From a market perspective, fluctuations in sales of new products directly impact the availability and return rate of corresponding cores. As pointed out by multiple interviewees, there is often a causal correlation between the timing of new product sales and the timing of returns [42,43].

The Returns category captures the design and structure of the take-back system used by a company. Experts identified a wide range of take-back models in practice, including [44,45]:

• Ownership-based take-back;
• Contractual take-back agreements (e.g., service contracts, contract repairs);
• 1:1 returns;
• Returns tied to discounts on remanufactured goods;
• Purchase-based returns;
• Voluntary returns.

Each of these systems entails different levels of predictability. According to Agrawal and Singh, companies that maintain ownership of the product, e.g., by offering product–service systems, or that have contractual obligations for returns, are generally more capable of accurately forecasting return timing and quantity [46]. Interview insights strongly supported this, particularly from remanufacturers with integrated service contracts.

The Product category includes the inherent characteristics, such as design complexity, application area, and expected service life. These features determine the wear profile, directly influencing the return condition. Components with longer life cycles or modular design tend to be returned in more predictable intervals and often in better condition. As several experts noted, a product failure reflects degradation in condition and is frequently the trigger for the return itself, making it both a consequence and a cause.

A central insight from the expert interviews is that these categories do not operate in isolation. Instead, they form a network of interacting variables. For example, poor material quality (Material) might lead to failure (Product), but the impact on return timing may depend on usage intensity (People) and environmental exposure (Environment). Similarly, while a take-back model (Returns) determines whether and when a product is returned, the condition upon return is shaped by all preceding factors. This interdependency underlines the complexity of forecasting in remanufacturing and highlights the necessity for data-driven approaches to capture such multi-factorial dynamics.

3.2. Machine Learning Methods for Forecasting Return Quantity and Timing of Cores

This section presents the literature review results concerning ML methods used to forecast the quantity and timing of returned cores. Due to the limited number of studies focusing directly on remanufacturing, the review was extended to include spare parts demand forecasting, given the structural similarities between the two domains. These include uncertain return flows, component life cycles, failure rates, and variable demand patterns, all of which are difficult to predict. The review emphasizes data sources and the applied machine learning models, as these elements are relatively domain-agnostic and thus transferable across industries.

3.2.1. Data Set Sources

In the context of this literature review, three primary sources of data were identified, from which the analyzed publications obtained their datasets for model training:

• Simulated data generated through computational models;
• Freely available benchmark datasets, often created in the context of competitions or for comparative research purposes;
• Real-world data.

Table 2. Primary source of process data to train models.

Data Source	Publications	Percentage [%]
Simulation	[20,24,47,48,49,50,51,52]	11
Benchmark	[53,54,55,56,57]	7
Real-world data	[15,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115]	82

provides an overview of the reviewed publications according to their primary data source. While only a small proportion of studies relied on simulation data (11%) or benchmark data (7%), most publications utilized real-world data for training and validating their models.

The following provides a detailed discussion of each data source category used for model development, training, and validation.

Simulation: Eight publications [20,24,47,48,49,50,51,52] employed synthetic data generated through simulations of manufacturing or remanufacturing processes. Simulation was especially prevalent in studies addressing remanufacturing, where access to operational data is often restricted due to confidentiality or the complexity of reverse logistics [20,24,48,49,50]. Several studies used Monte Carlo simulations to replicate return flows by sampling from statistical distributions that represent uncertainty in variables such as return timing, product lifespans, failure rates, and external disruptions [20,48,49]. In addition to validation, simulation environments often served as controlled testbeds for tuning model hyperparameters and exploring different configurations [20,49]. Jung et al. explicitly cited data scarcity as a motivation for using simulation, emphasizing its role as a practical workaround in the early stages of model development [47].

Benchmark: Five studies utilized publicly available datasets as a training basis for ML models [53,54,55,56,57]. These benchmark datasets supported comparability across studies and reproducibility of results. Notably, three studies [53,54,55] used maintenance data from the Army DELIIS system, which includes detailed records of spare parts failures and maintenance histories for military vehicles such as tanks and aircraft. These datasets enabled the application of various ML techniques, including text mining, time series analysis, and neural networks. Chandriah et al. [56] employed Norway’s public vehicle registration data to forecast spare parts demand using deep learning, while Pawar and Tiple [57] adapted a nontraditional dataset, the Vietnam War Bombing Operations dataset, for predictive modeling, highlighting creative strategies for compensating for data scarcity.

Real-world data: A total of 60 of the 72 reviewed publications [15,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115] used empirical datasets from actual operations, making this the dominant data source (82%). These datasets often include historical time series tracking usage, failures, maintenance events, inventory levels, and sales. The data were typically sourced from sectors such as military logistics, automotive aftermarket, or heavy machinery industries, where the need for accurate return forecasts is especially pronounced, reflecting a broad applicability across different domains. While real-world data increases model relevance and external validity, it also presents practical challenges, including missing values, inconsistent measurements, and noise—issues that must be addressed during preprocessing and model training.

3.2.2. Machine Learning Methods

After assessing the data sources, the review turned to the machine learning models applied to forecast return quantity and timing. The analysis considered both the types of models employed and the experimental strategies used to evaluate their effectiveness.

Overall, 34% of publications (24 out of 71) evaluated only a single model or different variants of the same model, such as alternative architectures or hyperparameter settings. The remaining 66% (47 publications) compared multiple models in their experiments. In the following, the term “prime model” refers to the best-performing model or the primary focus of a publication, while “baseline model” denote those used for performance comparison.

Table 3. Overview of the ten most frequently used models.

	Traditional Statistical Models					Traditional ML Models			Deep Learning Models
Publications *	MA	ARIMA	ES	SBA	CR	R	SVM	RF	ANN	LSTM
Saraswati et al. [15]									x
Tsiliyannis [48]		x
Zhu et al. [51]				x	x
Van der Auweraer & Boute [52]			x	x
Kim et al. [53]	x					x	x	x	x	x
Choi & Suh [54]						x	x	x	x
Kim [55]		x	x			x	x		x
Chandriah & Naraganahalli [56]			x	x	x					x
Chien et al. [58]		x		x
Guo et al. [59]			x						x
do Rego & de Mesquita [60]	x			x
Dombi et al. [61]		x	x						x
Amirkolaii et al. [62]	x		x	x	x				x
Ifraz et al. [64]						x	x	x	x
Mobarakeh et al. [65]	x		x	x	x
AlAlaween et al. [66]						x			x
Babaveisi et al. [68]				x	x		x
Boukhtouta & Jentsch [69]	x	x					x
Dali & Chengcheng [70]		x	x							x
Han et al. [71]							x		x
Lee & Kim [73]	x		x				x
Liu et al. [74]							x		x
Ma et al. [75]									x	x
Melo et al. [76]	x				x				x
Pawar & Tiple [78]						x	x	x	x
Wang et al. [81]							x		x
Wu & Bian [82]			x
Xing & Shi [83]					x		x		x
Cao & Li [85]	x		x						x
Carmo et al. [86]										x
Caserta & D’Angelo [87]					x			x	x
Ding et al. [88]	x		x			x
Fan et al. [89]		x		x			x			x
Guimaraes et al. [90]		x
Hong et al. [91]		x			x			x		x
Hu et al. [93]									x
Huang et al. [94]		x	x				x		x
Jiang et al. [96]	x		x	x	x		x		x	x
Kacmary et al. [97]		x				x
Kim et al. [98]	x					x	x	x	x	x
Li et al. [99]	x	x	x		x				x
Liu [100]			x
Lucht et al. [101]									x
Ma & Kim [102]		x
Mao et al. [103]									x
Özbay et al. [104]	x		x
Qiu et al. [105]									x
Ren & Zhang [106]		x
Rosienkiewicz [107]	x	x	x	x			x		x
Tsao et al. [109]	x		x	x	x	x	x	x	x
Tsao et al. [110]	x		x	x	x	x	x	x	x
Vaitkus et al. [111]	x		x				x		x
Vasumathi & Saradha [112]					x
Xu et al. [113]							x	x	x
Yang et al. [114]						x		x
Zhu et al. [115]								x
Total	17	15	21	13	15	12	21	12	31	9

ANN = Artificial Neural Network, LSTM = Long-Short-Term-Memory, MA = Moving Average, ARIMA = Autoregressive Integrated Moving Average, ES = Exponential Smoothing, SBA = Syntetos–Boylan Approximation, R = Regression, SVM = Support Vector Machine, CR = Croston. * The following publications did not apply any of the most frequently applied models and are therefore not included in the table: Refs. [20,24,47,49,50,57,63,67,72,77,79,80,84,92,95,108].

presents the ten most frequently used ML models across the reviewed publications.

Traditional statistical models: Classical forecasting techniques—including Moving Average (MA), Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ES), Syntetos–Boylan Approximation (SBA), and Croston (CR)—were widely used, primarily as baseline models [65,89,107,109]. These models are valued for their simplicity and interpretability, and they perform well on stable time series data with consistent trends or seasonality. However, their limitations become apparent when faced with highly volatile, nonlinear, or intermittent return patterns, which are common in core returns for remanufacturing. Consequently, they are often used to benchmark more flexible ML-based models.

Traditional machine learning models: Algorithms such as Regression (R), Support Vector Machines (SVMs), and Random Forests (RFs) were frequently applied due to their ability to capture nonlinear dependencies and handle diverse input features. SVMs and RFs, in particular, demonstrated robustness against overfitting and performed well even with smaller or incomplete datasets, making them especially relevant in contexts where high-quality data is scarce or inconsistent [69,96,113]. These models also support interpretability and can handle categorical and continuous input variables.

Deep learning models: Advanced models such as Artificial Neural Networks (ANNs) and Long Short-Term Memory (LSTM) networks were increasingly employed in studies dealing with larger, more complex datasets [56,66,103]. LSTMs in particular excel at modeling temporal dependencies, making them ideal for time-series forecasting tasks in remanufacturing. However, their high data and computational requirements can present obstacles in industrial settings where training data may be limited or noisy [56,70].

3.3. Machine Learning Methods for Forecasting the Condition of Cores

Building on the return quantity and timing findings, this section examines machine learning methods applied to forecasting the condition of returned cores. Due to a lack of studies specifically addressing condition prediction in remanufacturing, the review draws on adjacent research in quality prediction within manufacturing processes. The underlying rationale is that the core condition upon return can be analogously modeled using quality-related indicators derived from production data.

Within the reviewed literature, two primary methodological approaches were identified: classification and regression, each corresponding to the nature of the predicted outcome. Classification methods, used in 21 of the 53 reviewed studies, aim to predict discrete quality classes (e.g., pass/fail, scrap/non-scrap). Regression methods, applied in 32 studies, focus on forecasting continuous condition metrics such as wear level, dimensional deviations, or surface roughness. The following sections provide a detailed examination of data sources and model types used in both approaches, with a view toward their relevance and transferability to remanufacturing.

3.3.1. Data Set Sources

As with forecasting return quantity and timing, three primary data sources were identified in the reviewed literature on condition prediction: real-world production data, simulated data, and public benchmark datasets. These sources mirror the structure and characteristics discussed in the previous section.

Table 4. Primary source of process data to train models.

Data Source	Publications	Percentage [%]
Simulation	[116,117]	3
Benchmark	[118,119,120,121,122,123,124,125,126,127,128,129]	23
Real-world data	[130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168]	74

provides an overview of the reviewed publications classified according to these data source types.

Simulation: Two publications [116,117] used simulation-based data to train and validate models under controlled conditions, particularly when real-world data was scarce. For instance, Alenezi et al. [116] implemented a physics-informed, weakly supervised approach incorporating domain knowledge into model training. Similarly, Wang et al. [117] utilized graph-based modeling to emulate variability in mass-customized production settings. In both cases, simulation facilitated the modeling of complex manufacturing dynamics without sufficient empirical data.

Benchmark: Twelve studies [118,119,120,121,122,123,124,125,126,127,128,129] employed publicly available datasets to benchmark model performance. These datasets often feature high dimensionality and class imbalance, making them ideal for testing the robustness of hybrid or ensemble methods. For example, Bai et al. [126] and Zhou et al. [118] evaluated models under realistic industrial constraints, while others, like Psarommatis et al. [120] and Kao et al. [119], focused on multistage or zero-defect manufacturing processes, using datasets that reflect sequential quality control points and sparse defect labels.

Real-world data: The majority of studies, 39 out of 53, relied on empirical datasets collected from operational manufacturing environments [130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168]. These covered a broad range of applications, including additive manufacturing [133,156,167], semiconductor production [135,142], metal casting [141], and bioprocessing [164]. The datasets often included time-series sensor readings, process parameters, and post-production quality labels.

Although real-world data offers strong external validity and practical relevance, it also introduces challenges such as missing values, inconsistent sampling, and class imbalance. These issues were addressed using advanced preprocessing techniques, transfer learning, or ensemble modeling [123,150,154]. Several studies also explored the integration of edge computing [147,153] and explainable AI methods [144] to enhance real-time applicability and transparency in decision-making.

3.3.2. Machine Learning Methods

The analysis of machine learning methods for condition prediction revealed various techniques, reflecting the complexity and variety of manufacturing quality data. As with return forecasting, approximately one-third of the studies applied a single model or internal model variants, while most conducted comparative analyses of multiple approaches.

Table 5. Overview of the ten most frequently used models.

	Regression					Classification
Publications *	R	SVM	RF	ANN	LSTM	DT	RF	XGB	SVM	ANN
Alenezi et al. [116]	x	x	x
Wang et al. [117]			x
Psarommatis et al. [120]	x				x
Bai et al. [121]		x		x
Caihong et al. [122]				x
Liu et al. [123]		x
Zhang et al. [125]	x	x	x	x
Bai et al. [126]		x		x
Bai et al. [127]		x
Demirel et al. [128]			x
Deng et al. [129]		x		x
Chien et al. [130]
Wang et al. [131]		x	x	x
Zhang et al. [133]	x		x		x
Wang et al. [134]	x	x		x
Wang et al. [136]			x	x
Nikita et al. [137]		x	x	x
Schorr et al. [138]			x
Kobayashi et al. [139]	x		x
Sun et al. [140]	x		x		x
Yang et al. [141]		x		x	x
Bai et al. [143]		x		x
Jiang & Yen [145]				x	x
Ju et al. [146]		x	x
Trappey & Chien [150]	x
Beckschulte et al. [152]				x
Kim et al. [159]	x	x	x	x
Li et al. [161]		x			x
Tercan & Meisen [164]	x	x	x	x
Xiao et al. [167]	x			x
Zhou et al. [118]									x
Kao et al. [119]						x			x	x
Shim et al. [132]							x			x
Huynh [135]										x
Forsberg at al. [144]						x	x	x
Lee et al. [147]									x	x
Matzka [148]						x			x	x
Mohammadi & Wang [149]									x
Yuan et al. [151]						x	x
Chen et al. [153]						x	x	x	x
Deuse et al. [155]						x			x
Huang et al. [156]									x	x
Jun et al. [157]						x	x		x	x
Jung et al. [158]							x	x	x
Lee et al. [160]						x	x		x	x
Olowe et al. [162]						x	x	x	x
Sankhye & Hu [163]							x	x
Tian et al. [165]								x
Zhang et al. [168]						x	x
Total	11	16	14	16	6	10	10	6	12	8

ANN = Artificial Neural Network, LSTM = Long-Short-Term-Memory, XGB = Extreme Gradient Boosting, R = Regression, SVM = Support Vector Machine, DT = Decision Tree, RF = Random Forest. * The following publications did not apply any of the most frequently applied models and are therefore not included in the table: Refs. [124,130,142,154,166].

summarizes the ten most commonly used ML models identified across the reviewed publications.

Regression: A total of 32 publications employed regression models like linear or multiple regression to predict continuous quality characteristics. Among traditional approaches, regression models remain a common baseline due to its simplicity and interpretability, appearing in 12 studies [116,120,125,133,134,139,140,150,159,164,167]. While often outperformed by more advanced methods, they are still valuable for low-variance scenarios and for establishing benchmark accuracy levels.

SVM and Random Forest (RF) represent this category’s most frequently applied traditional machine learning models. SVM was used in 16 studies [116,121,123,125,126,127,129,131,134,137,141,143,146,159,161,164], demonstrating strong performance in modeling nonlinear relationships, particularly with smaller datasets. RF appeared in 14 studies, offering ensemble-based robustness against overfitting and interpretability through feature importance ranking [116,117,125,128,131,133,136,137,138,139,140,146,159,164].

Deep learning models are also increasingly applied in regression tasks. ANNs were used in 16 publications [121,122,125,126,129,131,134,136,137,141,143,145,152,159,164,167], typically in scenarios with sufficient training data and complex feature interactions. More advanced models like LSTM appeared in six studies [120,133,140,141,145,161], especially where sequential dependencies or time-based process characteristics were central in determining final product quality.

Classification: Of the 53 reviewed publications, 21 studies applied classification techniques to predict discrete quality outcomes, such as defect classes, pass/fail status, or categorical condition levels. SVM emerged as the most frequently used classification method, adopted in 12 studies [118,119,147,148,149,153,155,156,157,158,160,162]. SVM’s ability to construct hyperplanes in high-dimensional spaces makes it particularly effective for separating class boundaries in imbalanced or noisy datasets—a common feature of industrial defect data.

RF also plays a central role in classification, appearing in ten publications [132,144,151,153,157,158,160,162,163,168]. Its ensemble structure allows it to generalize well, even in highly variable production environments. Other widely used methods include Decision Trees (DTs) [119,144,148,151,153,155,157,160,162,168] and Extreme Gradient Boosting (XGB) [144,153,158,162,163,165], which offer strong performance and scalability, especially in cases with limited or imbalanced labeled data.

Neural networks continue to gain traction in classification tasks as well. Although less prevalent than in regression, ANNs were still used in eight studies for categorical quality prediction [119,132,135,147,148,156,157,160]. These models are particularly favored when quality classification is based on complex sensor signals or multivariate process data.

In summary, the reviewed literature on quality prediction in manufacturing demonstrates a broad spectrum of machine learning techniques, each tailored to the nature of the data and the forecasting objective, whether categorical or continuous. This methodological diversity underscores the field’s maturity and provides valuable insights for transferring these approaches to remanufacturing, where similar challenges arise in predicting the condition of cores. The following discussion will synthesize these findings and reflect on their implications for future research and industrial applications in remanufacturing.

4. Discussion

This study aimed to identify key influencing factors and suitable ML approaches for forecasting return quantity, timing, and condition of cores in remanufacturing. The findings offer practical and academic perspectives from expert interviews and two systematic literature reviews. The following discussion synthesizes these insights, focusing on their transferability to the remanufacturing context and outlining current limitations and future research needs.

4.1. Transferability of Influencing Factors and Data Sources

The expert interviews revealed that the predictability of returns is influenced by a diverse set of interrelated factors, spanning product design, material quality, customer behavior, usage context, and take-back strategies. These factors were systematically categorized using an adapted Ishikawa diagram to reflect their multi-causal nature. Particularly noteworthy is the cascading influence of user behavior and product failure on return timing and condition, highlighting the need for data that captures technical, behavioral, and contextual variables. However, it should be noted that all interviewees came from the transportation sector, including automotive, rail, and bicycle industries, which may introduce a sector-specific bias in the identified influencing factors and their perceived importance.

Once these influencing factors are identified, the next step involves locating and extracting the relevant data from the company’s IT systems. This requires assessing where the data is stored, evaluating whether its granularity is sufficient, and determining how it can be accessed for model development. Not all identified factors are equally relevant, as some may have their influence overestimated or underestimated. Therefore, applying feature selection techniques is essential to isolate the most impactful variables, ensuring that the final set of model inputs contributes meaningfully to forecasting accuracy.

The literature reviews confirmed that machine learning models for forecasting return timing, quantity, and condition typically rely on three types of data: real-world operational data, simulated datasets, and publicly available benchmark datasets. This combination mirrors the data landscape in remanufacturing, where proprietary constraints, inconsistent data quality, and limited access to labeled datasets present ongoing challenges. While real-world data offers the highest fidelity for deployment, simulated and benchmark datasets remain crucial for early experimentation, model comparison, and addressing edge cases, particularly in underexplored remanufacturing scenarios.

However, the influencing factors identified in adjacent research domains, such as spare parts demand forecasting or quality and defect prediction in manufacturing, often differ from those relevant in remanufacturing. These domains focus on structured, product-centric variables within stable, linear supply chains. In contrast, remanufacturing involves greater variability and uncertainty, driven by diverse customer usage behaviors, inconsistent core conditions, and complex reverse logistics systems. As such, many of the influencing factors in remanufacturing are context-specific and underrepresented in the reviewed literature. This highlights the importance of conducting expert interviews to capture practice-oriented knowledge that complements and extends the literature findings.

Nonetheless, despite these differences in influencing factors, the underlying structure of machine learning models identified in the literature remains applicable. These model types provide a transferable foundation that can be adapted to the specific characteristics and requirements of remanufacturing use cases.

4.2. Applicability of Machine Learning Models in Remanufacturing

From a methodological perspective, the literature review showed that a broad spectrum of ML models has been applied in both forecasting domains, ranging from classical statistical approaches to advanced neural networks. For remanufacturing, where data availability can be sparse or inconsistent, traditional models like regression, SVMs, and RFs are particularly promising due to their robustness and interpretability.

In contrast, deep learning models such as LSTM and ANN offer advantages in capturing nonlinear dependencies and sequential behavior, which are beneficial in modeling degradation patterns or time-series returns but require extensive and high-quality datasets that are often unavailable in remanufacturing.

An important insight from the review is the alignment between model type and prediction target. For return quantity and timing, time-series forecasting methods and regression models dominate. At the same time, condition forecasting employs regression and classification techniques depending on whether the outcome is categorical or continuous. This flexible approach is directly transferable to remanufacturing, where cores may be assessed on graded scales or binary condition outcomes (e.g., non-scrap vs. scrap).

However, it is important to acknowledge the limitations inherent in transferring findings from adjacent domains, such as spare parts demand forecasting and quality or fault prediction in manufacturing. While these fields share structural similarities with remanufacturing, such as irregular demand patterns, failure prediction, and data-driven decision-making, they typically operate in more controlled, linear supply chains with better data availability and standardization. Remanufacturing, by contrast, involves greater uncertainty due to variable product usage histories, inconsistent core condition, and more complex reverse logistics. These differences may affect model performance and generalizability. Therefore, while cross-domain insights provide valuable starting points, their application in remanufacturing requires careful contextual adaptation and further empirical validation.

4.3. Challenges and Research Gaps

Despite methodological advancements, several gaps remain. First, there is a lack of direct application and validation of ML models within actual remanufacturing systems. Most current research is derived from adjacent domains like spare parts logistics and manufacturing, limiting contextual relevance. Second, data heterogeneity, including sensor types, process structures, and quality indicators, poses a significant challenge for model generalization. Third, behavioral and contextual data, repeatedly cited by experts as critical, are underrepresented in existing datasets.

To address these limitations, future research should prioritize the development of remanufacturing-specific datasets, including integrated behavioral, operational, and environmental variables. Moreover, applying advanced learning techniques such as multi-task learning, transfer learning, and hybrid modeling offers a promising approach for capturing the dynamics of remanufacturing. Multi-task learning, for instance, enables simultaneous prediction of multiple return characteristics, e.g., timing, quantity, and condition, by leveraging shared features across tasks to improve overall performance. Transfer learning allows knowledge gained from related domains or pre-trained models to be applied to remanufacturing scenarios, which is particularly beneficial in cases of limited or imbalanced data. Hybrid modeling combines data-driven machine learning with domain knowledge, such as physical laws or expert-defined rules, thus improving model robustness and interpretability.

5. Conclusions

This work investigates how machine learning can support forecasting return quantity, timing, and condition of used parts in remanufacturing. A mixed-method approach was employed, combining expert interviews with two systematic literature reviews. The expert interviews provided practical insights into the real-world factors affecting product returns. At the same time, the literature reviews explored existing machine learning methods used in related fields—specifically, spare parts logistics and manufacturing quality prediction.

The expert interviews revealed that a wide range of interconnected factors influence the predictability of returns. These include product and material properties, human behavior, environmental conditions, and the structure of the take-back system. The results were synthesized into an adapted Ishikawa diagram to visualize these multi-causal relationships and underline their relevance for forecasting models.

The literature reviews complemented these findings by analyzing 125 publications—72 related to return quantity and timing, and 53 to condition forecasting. While only a few studies explicitly addressed remanufacturing, many contributions from adjacent fields could be transferred. Data sources used for model training were primarily real-world datasets, but simulations and benchmark data also played a significant role, especially where access to proprietary data was limited.

Across both forecasting domains, a wide array of machine learning models was identified. Traditional models such as Regression, SVMs, and RFs were commonly used due to their robustness and interpretability. More complex models like ANNs and LSTMs were favored in data-rich environments, particularly when capturing nonlinear or temporal patterns. In the condition prediction literature, a distinction was made between classification and regression approaches, depending on whether the quality outcome was discrete or continuous.

This work contributes to the growing body of research in remanufacturing by systematically identifying applicable ML models and data requirements and proposing a structured framework for understanding the factors influencing product returns. However, while the work outlines suitable influencing factors and ML models, real-world application in remanufacturing is still challenged by the limited availability of historical data. Additionally, the review results have not yet been tested or benchmarked in industrial contexts. Without empirical validation in operational environments, the reliability and practicality of these findings regarding influencing factors and ML models remain uncertain. Future research should therefore focus on real-world implementation and the development of remanufacturing-specific datasets to strengthen practical applicability.