Approaches for Lifetime Prediction of Vehicle Traction Battery Systems During a Technical Inspection: A Systematic Review

Abstract

Creating trust in society for new technologies, such as a new types of powertrains, and making them marketable requires transparent, neutral, and independent technical verification. This is crucial for the acceptance and success of electrified vehicles in the used car markets. A key component of electric vehicles is the traction battery, whose current and future condition, particularly regarding aging, determines its residual value and safe operation. This review aims to identify and evaluate methods for predicting the lifetime of onboard traction batteries, focusing on their applicability in technical inspections. A systematic literature and patent review was conducted using targeted keywords, yielding 22 patents and 633 publications. From these, 150 distinct lifetime prediction methods were extracted and categorized into a four-level mind map. These methods are summarized, cited, and structured in detailed tables. The relationships between approaches are explained to clarify the current research landscape. Long Short-Term Memory, Convolutional Neural Networks, and Particle Filters were identified as the most frequently used techniques. However, no methods were found suitable for predicting the lifetime of traction batteries during technical vehicle inspections, which operate under short test durations, limited data access, and diverse real-world operating conditions. Most studies focused on cell-level testing and did not address complete battery systems in operational vehicles. This gap highlights the need for applied research and the development of practical methods to support battery assessment in real-world conditions. Advancing this field is essential to foster confidence in battery systems and enable a sustainable transition to electromobility.

1. Introduction

The automotive industry is currently undergoing a significant transformation towards more sustainable powertrain technologies, with battery electric vehicles (BEVs) playing a key role [1,2,3]. This development is driven by government policies, such as Regulation (EU) 2023/851 of the European Parliament, which mandates the ban on the sale of new combustion engine vehicles from 2035 in Europe [4]. In Germany, the aim is to have 15 million fully electric vehicles on the roads by 2030 [5]. This target has been jointly set by the German federal government, automotive manufacturers, and labor unions [5]. This transition to battery electric vehicles is already evident in the numbers of new BEV registrations in Germany, which has increased from 64,931 vehicles (1.80%) in 2019 to 380,609 vehicles (13.50%) in 2024 [3]. In parallel, the general energy transition toward renewable energy sources such as solar and wind power further accelerates the shift toward electromobility, as their intermittent and fluctuating generation increases the demand for electricity storage solutions. To address this challenge, the traction batteries in BEVs can, for example, improve the utilization and stability of renewable energy systems, and therefore naturally reinforce the transition to electric mobility [6].

Due to the increasing number of registered electrified vehicles, more and more EVs will undergo technical inspections (TIs). The term “technical inspection” is used as a generic term for different vehicle checks, such as vehicle appraisals, where experts for motor vehicles and road traffic assess and evaluate vehicles and vehicle damage [7,8]. The Institute for Expert Services (in German: Institut für Sachverständigenwesen e.V.) and VDI—The Association of German Engineers developed and published guidelines and principles [7,8]. A TI is also a vehicle check to ensure that the vehicle is safe, roadworthy, and operating in accordance with the applicable laws and regulations [9]. In Germany, for example, one of these laws is the Road Traffic Licensing Regulation, which regulates the performance of a Periodical Technical Inspection (PTI) by an officially recognized testing organization. In the following, the term TI is used for both types of vehicle checks.

Currently, in all these different TI checks for BEVs, the battery, which is part of the powertrain, is only visually inspected for condition and abnormalities [8,10]. A visual inspection does not provide reliable conclusions about the actual condition of the traction battery due to aging processes, such as the remaining capacity, internal resistance, and overall performance. These unknown factors lead to uncertainties in residual value determination during vehicle appraisal. Since the battery is the vehicle’s most expensive component, these uncertainties significantly impact the residual value assessment [11].

This can be seen, for example, in an analysis of the AutoScout24 online marketplace on used vehicles aged between two and four years with a mileage of less than 50,000 km. The comparable VW ID.3 and VW Golf models are evaluated against each other. The ID.3 has a significantly higher gross list price of approximately 10,000 € more than the Golf [12,13]. A comparison between the VW ID.3 and the VW Golf reveals that, as of April 2024, the average residual value of the electric vehicle was approximately 3700 € lower than that of the value-retaining VW Golf with an internal combustion engine (ICE) (VW ID.3: 22,986 € vs. VW Golf: 26,697 €) [14]. This comparison shows a much higher loss in residual value of the BEVs. This results in lower economic efficiency and a worse total cost of ownership (TCO) for electric vehicles.

These high losses in residual value and worse TCO for electrified vehicles after a short lifetime slow down the successful market access for electromobility. Car rental companies are also moving away from electric vehicles. Hertz initially planned to purchase 100,000 Tesla and 65,000 Polestar vehicles, aiming to electrify a quarter of its fleet by the end of 2024. However, the company has now reversed course, selling 20,000 electric vehicles from its current vehicle fleet and refocusing on internal combustion engine vehicles [15].

The overproportionate drop in the residual value of BEVs compared to ICE vehicles can be attributed to the lack of knowledge and experience with used BEVs among technicians, fleet managers, salespeople, or vehicle buyers, as well as the absence of suitable battery evaluation methods. This situation creates doubts and skepticism in society about battery electrified vehicles [2,16,17]. Recent studies indicate that traction batteries, which are the most expensive component of an electrified vehicle, perform better than their reputation suggests [18,19]. Public acceptance of electromobility could potentially improve if a reliable assessment of battery condition were integrated into technical inspections.

Within the literature, a wide range of approaches for assessing battery condition can be found under the generic term lifetime prediction methods. These methods determine the current state of the aging process and predict the remaining useful lifetime (RUL). Despite this, these approaches are still largely confined to academic research and have not yet found meaningful application in real-world traction battery systems. Existing products on the market can provide a snapshot of the current battery State of Health (SoH) or remaining available battery capacity. However, they do not offer reliable predictions of future performance or degradation [20,21].

Knowledge of the remaining useful life obtained through lifetime prediction during a technical inspection could help minimize uncertainties in residual value estimation, reduce social concerns and reservations, and ensure the safe operation of electric vehicles. Therefore, this review aims to identify methods for predicting the lifetime of traction batteries and evaluate their applicability in the context of technical inspections. This assessment aims to highlight the most promising research approaches and pinpoint remaining research gaps.

To achieve this aim, this work conducts a detailed and up-to-date review of patents and the literature. Several databases were searched using defined keywords, and the results were analyzed. Based on this, the currently researched methods for predicting the lifetime of lithium-ion batteries were identified. In a second step, the most promising methods were selected and discussed in terms of their applicability in TI.

The initial activities in the industry demonstrate a pressing need for active research on this topic. The German automotive testing organizations collaboratively created the Charta 2030, a publication outlining the necessary advancements in technical vehicle testing. The goal is to sustainably and safely introduce the rapidly advancing automation and electrification of vehicles to the road. Therefore, further developments and innovations in technical vehicle testing are crucial for ensuring safe electromobility and reliable traction batteries. A lifetime prediction of the traction battery aligns with the Charta 2030 strategy and its objectives [22].

2. Related Work

Several reviews already exist on the topic of lifetime prediction of lithium-ion batteries, which are briefly presented below and discussed in connection with this work.

In August 2021, Yao et al. published a review with 155 references, which deals with lithium-ion battery state of health estimation and prediction models [23]. Yao et al. explain degradation mechanisms, core definitions of SoH, and the development status of estimation and prediction approaches. These methods are grouped into three main categories: model-based, data-driven, and fusion-based approaches, which are further structured in a mind map. The authors identify four significant challenges for SoH estimation: difficult measurability, strong time dependence, irreversibility of degradation, and high system nonlinearity. Yao et al. analysis emphasizes the comparative strengths and weaknesses of each method category. The review also outlines future trends, including the integration of multiple algorithms, advanced feature extraction techniques, and the use of cloud platforms for enhanced processing and scalability. However, the study remains theoretical mainly, with no application scenarios involving real-world traction batteries or data acquisition under practical vehicle inspection conditions. Yao et al. do not address implementation constraints, vehicle appraisal considerations, or the specific needs of periodic technical inspections. It also focuses on SoH estimation, rather than predicting remaining useful life. In contrast, this work aims to evaluate lifetime prediction methods for traction batteries with a clear emphasis on practical feasibility, measurability during inspections, and integration into real-world inspection frameworks. Therefore, whereas the review by Yao et al. provides valuable methodological background, it does not offer a directly applicable solution for the objectives of this work.

In July 2022, Pang et al. published a review with 107 references, which deals with the prediction of the state of health and service life of lithium-ion batteries [24]. Pang et al. compared and analyzed the current prediction models for the remaining useful life of LIB, which are divided into mechanism-based models, semi-empirical models, and data-driven models. The fusion of the various approaches is also explained and discussed. Pang et al. discuss the advantages, limitations, technical challenges, and performance metrics of each model type, and include an evaluation based on six performance indicators. Hybrid modeling approaches and future directions, such as enhanced interpretability and model simplification, are also addressed. Whereas Pang et al. provide a thorough methodological overview, their study remains theoretical and does not consider real-world applications or the integration of prediction methods into technical inspection procedures. Moreover, no reference is made to the use of onboard battery data from electric vehicles. In contrast, this work evaluates explicitly lifetime prediction methods for traction batteries in the context of technical inspections, focusing on feasibility, measurability, and practical applicability. Thus, the review of Pang et al. serves as a valuable theoretical foundation but does not provide a solution aligned with the operational goals of this work.

In July 2022, Zhao et al. published a review with 34 references that addresses RUL prediction methods for lithium-ion batteries [25]. The authors provide a structured classification of prediction approaches into model-based, data-driven, and hybrid methods, each with detailed explanations of strengths, limitations, and example applications. The publication outlines typical challenges in RUL prediction, such as high model complexity, limited generalizability under variable operating conditions, and the difficulty of extending models from single cells to battery packs. While these challenges are well acknowledged, the discussion remains largely conceptual and lacks evaluation of real-world applications. In contrast, this work focuses specifically on assessing the feasibility of lifetime prediction methods in the context of technical inspections of traction batteries. Furthermore, the relatively limited number of 34 references suggests that the review may not fully capture the scope of recent developments, especially when compared to this work’s extensive patent and literature research. As a result, Zhao et al. review does not provide a sufficient basis for evaluating inspection-relevant prediction methods and therefore does not meet the practical objectives of this work.

In September 2022, Kafadarova et al. published a review with 28 references that addresses the state of health battery diagnostics methods: current status and future challenges [26]. The publication categorizes SoH prediction methods into three main groups (model-based, data-driven, and fusion approaches) and provides a basic summary of six submethods within these categories. While the structure offers a helpful overview, it remains limited in methodological depth and scope. Compared to the broader and more detailed classification presented in this work, Kafadarova et al. list only four submethods per group and do not include more recent or specialized techniques. Moreover, the review misses a discussion on real-world battery data, onboard applicability, or inspection-based limitations, which are essential to the objectives of this work. With only 28 cited references, the reviewed publication presents a narrow perspective that may overlook significant developments, especially when contrasted with the extensive patent and literature research conducted for this work. As a result, the review does not provide the necessary foundation for evaluating lifetime prediction methods in the context of technical inspections of traction batteries.

In September 2022, Ansari et al. published a review with 147 references that addresses the remaining useful life prediction for LIB: A comprehensive review of methods, key factors, issues, and future outlook [27]. The aim is to provide a critical discussion and analysis of various aspects, such as methods, classifications, characteristics, advantages, disadvantages, and research gaps in the RUL prediction of lithium-ion battery storage systems. Ansari et al. utilized similar platforms for literature research, excluding Cell Press and Springer Link, and conducted no patent research. The 147 reviewed and analyzed publications were published between 2014 and 2021, which means that recent developments beyond that period, particularly in machine learning, hybrid modeling, and edge deployment, are not included, limiting the review’s timeliness. The authors describe battery degradation mechanisms and provide a structured mind map that summarizes 43 prediction approaches, including several fusion methods. However, this review does not evaluate the applicability of these methods to real-world battery systems in electric vehicles. The only references to real data pertain to public data sets from NASA and MIT, which are not representative of typical traction battery behavior under real driving conditions. Furthermore, although the publication acknowledges the relevance of SoH estimation in electric vehicles, this is only briefly discussed with limited source support and without an analysis of its integration into inspection or maintenance routines. In contrast, this work conducts extended patent and literature research, explicitly targeting methods for lifetime prediction of traction batteries and their feasibility in the context of technical inspections. Therefore, whereas Ansari et al. offer a broad methodological overview, the review lacks the TI application focus necessary to meet the objectives of this work.

In October 2022, Elmahallawy et al. published a review with 184 references that offers a comprehensive review of lithium-ion battery modeling, state of health, and remaining useful lifetime prediction [28]. Elmahallawy et al. focused on the health status of lithium-ion traction batteries and their operational safety. The main goals are to present different battery models to replicate dynamic battery behavior and to organize them in a mind map. For this purpose, various battery modeling techniques were compared. Physically-based and data-driven approaches have been proposed and discussed in the literature, including their respective advantages and disadvantages. The review provides a general classification of prediction methods and battery models, emphasizing how they reflect aging behavior, thermal characteristics, and operational stress under real-world driving conditions. Although the NASA data set is used as the primary benchmark, the publication does refer to battery management system (BMS) data and its role in real-world applications. A proposed flow chart outlines a diagnosis and maintenance process for lithium-ion batteries in electric vehicles based on machine learning approaches and BMS integration. However, this framework is tailored toward onboard diagnostics and continuous monitoring via BMS, rather than discrete and externally conducted assessments like those performed during technical inspections. The study does not address practical limitations related to inspection scenarios. Furthermore, while the review is extensive in its scope, it does not include patents and focuses primarily on academic publications. Compared to this work, Elmahallawy et al. do not assess method maturity or implementation feasibility in a regulatory or inspection-related context. The review does not offer an evaluation of method suitability under the limited data and time conditions typical for vehicle inspections. Therefore, although the publication provides a broad overview of modeling strategies and lifetime estimation concepts, it does not fulfill the objectives of this work.

In March 2023, Zhao et al. published a review with 105 references that addresses state estimation and remaining useful life prediction methods for LIB [29]. The review begins by clearly defining key concepts such as SoH and RUL, followed by a detailed presentation of existing prediction approaches. These methods are systematically categorized in a comprehensive mind map and analyzed with respect to their methodological properties. The review distinguishes between data-driven, model-based, and hybrid approaches and discusses their respective advantages and disadvantages. Particularly noteworthy is the structured comparison of methods and the identification of research gaps, especially regarding the applicability of RUL prediction in real-world battery systems and real-time scenarios. However, the publication focuses on algorithmic aspects and does not examine the practical feasibility of implementing these methods in regulated processes such as technical inspections. Critical factors like measurement accessibility, data availability, robustness under real operating conditions, and compliance with regulatory requirements are only marginally addressed. Furthermore, the specific challenges associated with traction batteries, especially concerning vehicle inspections, are not considered. In contrast, this work explicitly focuses on lifetime prediction in the context of technical inspections. Based on an extended patent and literature analysis, this work not only examines the underlying methods but also evaluates them with regarding their practical applicability within inspection procedures. Criteria such as data accessibility, result accuracy, requirements for input data, and integration into established inspection workflows are systematically assessed. While Zhao et al. provide a valuable methodological overview, this work offers a more application-oriented perspective and contributes to bridging the gap between research and regulatory practice. Additionally, the inclusion of patent literature provides further insights into industrial relevance and innovation dynamics in the field of lifetime prediction.

In August 2023, Khandelwal et al. published a review with 35 references that addresses RUL prediction of LIB and emphasizes SoH estimation, error metrics, and method comparisons across machine learning and filtering-based approaches [30]. Khandelwal et al. present a general RUL prediction flowchart and evaluate various data sets and algorithms for SoH and RUL estimation, including techniques such as recurrent neural networks (RNNs), particle filters, and ensemble learning. While the work highlights methodological diversity and outlines common challenges in RUL prediction, it remains highly academic and limited in scope. The data sets examined are not based on real-world vehicle applications but rather on laboratory-controlled conditions or standard benchmarks like NASA data, which limits their transferability to practical use cases. Moreover, the review does not specifically focus on lifetime prediction in the context of technical inspections, nor does it address how prediction methods could be adapted for use with on-board traction battery data. It also lacks a discussion on key operational constraints such as measurement accessibility, robustness, and interpretability within automotive inspection frameworks. In contrast, this work expands on these aspects by incorporating comprehensive patent and literature research. This broader basis supports a more detailed evaluation of lifetime prediction methods, particularly regarding their applicability for traction batteries during real-world technical inspections. As a result, the review by Khandelwal et al. does not fulfill the objectives of the current review, which seeks a practical and inspection-relevant perspective on lifetime prediction.

In May 2024, Artelt et al. published a review with 70 references that addresses hybrid approaches and datasets for RUL prediction, emphasizing the integration of data-driven and physics-based methods [31]. The review systematically categorizes modeling strategies across various application domains and identifies a growing trend in combining Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), and other machine learning techniques with physical degradation models. Compared to this work, which seeks to evaluate the applicability of lifetime prediction methods for traction batteries within the context of technical inspections, the review by Artelt et al. differs in several key aspects. First, the scope of Artelt et al. remains broad across multiple domains and systems, including bearings, aircraft engines, and industrial actuators, while the current study focuses specifically on traction batteries used in electric vehicles. Second, Artelt et al. emphasize the technical configuration of hybrid models and the types of publicly available datasets employed, but they do so without critically evaluating their practical applicability in operational or regulatory inspection contexts, which is the primary objective of this study. Additionally, this work incorporates extensive patent and literature research, significantly broadening the evaluative basis compared to the 36 full-text hybrid modeling studies examined in Artelt et al. This comprehensive source base enables a deeper assessment of model robustness, data requirements, and interpretability. These factors are particularly relevant in inspection and certification environments where explainability and measurement feasibility are critical. Therefore, while both reviews overlap methodologically in their examination of data-driven and hybrid RUL models, they differ substantially in purpose and application orientation. Artelt et al. provide a valuable synthesis of modeling techniques and dataset usage, whereas the current review contributes a practical, inspection-oriented perspective that has not yet been adequately addressed in the literature.

In May 2024, Patrizi et al. published a review with 72 references that addresses degradation models and RUL prediction for testing design and predictive maintenance of LIB [32]. Patrizi et al. categorize and explain both stochastic and data-driven methods, such as general path models, Wiener and ARIMA processes, Kalman filtering, and a wide range of machine learning (ML) and deep learning (DL) approaches for modeling capacity degradation and estimating the remaining useful life (RUL) of lithium-ion batteries. Compared to this work, which aims to identify methods for predicting the lifetime of traction batteries and evaluate their applicability within the context of technical inspections, this publication conceptually overlaps in its focus on capacity degradation and RUL estimation. However, it likely distinguishes itself by specifically focusing on traction batteries and the practical applicability of lifetime prediction models in regulatory or periodic technical inspections for vehicles, which is not addressed in the review. In contrast, the reviewed article offers a broader perspective on degradation modeling in general lithium-ion battery contexts, with a strong emphasis on experimental test design and predictive maintenance strategies. It does not concentrate on technical inspections or regulatory contexts, nor does it include patent or literature survey methods as part of its methodology. While it presents valuable insights into degradation modeling, particularly exponential modeling of capacity fade and the use of advanced ML techniques such as transformers and LSTMs, it does not assess these methods in terms of their feasibility for real-world deployment in inspection scenarios. The discussed publication, by contrast, provides technical depth in modeling approaches but has a narrower scope and lacks extensive empirical evaluation or inspection relevance.

In June 2024, Reza et al. published a review with 185 references that addresses the recent advancements in RUL prediction for LIB in electric vehicle applications [33]. The publication examines a wide range of modeling approaches, including both physics-based and data-driven methods, with a strong emphasis on deep learning architectures. Additionally, the review discusses influencing factors such as battery aging mechanisms, feature extraction, data preprocessing, hyperparameter tuning, and the role of battery management systems (BMS) in enabling accurate RUL estimation. Particular attention is given to challenges such as thermal runaway, relaxation effects, material selection, and algorithmic complexity. While this work shares a common focus with the present review publication by Reza et al. in terms of identifying effective methods for lifetime prediction of lithium-ion batteries, its scope differs substantially. Reza et al. primarily address general EV applications, aiming to support safety, performance, and maintenance planning. In contrast, this work explicitly evaluates the applicability of RUL prediction methods within the context of technical inspections. This includes a specific focus on measurability, interpretability, and the integration potential of predictive models within inspection frameworks. Furthermore, while the reviewed publication offers broad coverage of methodological developments, it lacks a critical assessment of how such models can be translated into actionable tools for inspection routines. The current work fills this gap by assessing methods not only in terms of their predictive performance but also regarding their feasibility, data requirements, and robustness in real-world inspection scenarios. Hence, the two reviews are complementary: Reza et al. provide a foundational overview of state-of-the-art methods, whereas the present study offers a practical perspective oriented toward technical implementation and operational decision-making.

In November 2024, Zhang and Li published a review with 127 references that deals with SoH estimation for LIB in hybrid electric vehicles, categorizing approaches as experimental, model-based, and data-driven approaches [34]. Zhang and Li address capacity as a key degradation indicator, emphasizing performance monitoring under dynamic HEV conditions. In contrast, this work focuses on capacity prediction methods for traction batteries and evaluates their applicability within technical inspection frameworks. Unlike the reviewed publication, which does not consider regulatory or inspection-related implementation, this review integrates insights from extensive literature and patent research to assess method robustness, data requirements, and real-world feasibility. Thus, Zhang and Li offer a complementary technical perspective, whereas this work provides a broader, application-oriented evaluation.

As the related work shows, none of the existing review publications have specifically evaluated lifetime prediction methods for traction batteries in the context of technical inspections. Moreover, no previous reviews have performed a patent and literature analysis on this scale. The challenges outlined in the introduction, such as the development of battery lifetime and the determination of residual value, highlight the need to connect lifetime prediction with technical inspection.

3. Methodology for the Patent and the Literature Research

The patent and literature research was conducted in accordance with the preferred reporting items for systematic reviews and meta-analyses (PRISMA) 2020 guidelines, which is shown in Figure 1. To find all documents related to the topic under investigation, various platforms, such as the patent database (see Section 3.2), several journals, and scientific platforms, were utilized. All documents analyzed in this patent and literature research includes all available publications up to the end of 2024 and was evaluated through March 2025. Seven keywords were defined for patent and literature research and connected by Boolean operators. A detailed description is provided in the individual subsections.

3.1. Selected Keywords

The selected keywords were based on existing literature and common industry terminology to enable a broad yet targeted search query in the field of battery lifetime prediction. The keywords also take into account possible variations in terminology across different sources. This approach helps ensure that a wide range of relevant documents is covered without losing focus on the topic.

The different keywords are always a combination based on the words “Lithium-Ion Battery” and various expressions on the topic of “Lifetime Prediction.” This ensures that it is not just a matter of determining such a value, but strictly predicting it. The following seven add-ons were used:

• Lithium-Ion Battery Remaining Useful Life
• Lithium-Ion Battery Aging Prediction
• Lithium-Ion Battery Lifetime Prediction
• Lithium-Ion Battery State of Health Prediction
• Lithium-Ion Battery Life Forecasting
• Lithium-Ion Battery Capacity Degradation Prediction
• Lithium-Ion Battery Internal Resistance Prediction

The Boolean operators “AND” and “OR” were used as an additional tool to refine research results.

3.2. Patent Research Method

Patent research was conducted using the German Patent and Trademark Office’s database, specifically the German Patent Information System (DEPATIS) [35], which includes global patent publications. Therefore, other databases, such as those of the European Patent Office (ESPACEnet), the World Intellectual Property Organization (WIPO), the China Intellectual Property Right (CNIPAnet), or the database of the United States Patent and Trademark Office (USPTO), were not used.

At DEPATIS, the advanced search function was utilized, with the predefined keywords applied to the “full-text” and “title” fields. The results for each keyword could be exported as .csv files. Afterwards, all entries from the seven exported files were merged into a summary to eliminate duplicates. In the next step, the titles of the remaining patents were checked, and all entries that did not fit the topic were deleted. The resulting patents were manually inserted into a library in the literature management program, JabRef 5.12 [36].

Although the International Patent Classification (IPC) system could have been used as a basis for the patent research [37], as shown in

Table 1. List of the meanings for the individual IPC components and their corresponding results.

Symbol	Title
G	Section of Physics
01	Class of measuring
R	Investigation of electric and electronic properties
31	Testing electric properties
392	Individual testing apparatus for electrical devices and supervision of individual cells or group of cells within a battery
G01R 31/392	11,348 hits
G	Section of Physics
06	Class for Computing, Calculating, Counting
F	Electric digital data processing
17	Digital computing or data processing equipment or methods
50	Adapted for specific functions, like information retrieval, database structures and file system structures
G06F 17/50	198,894 hits

, this approach would not have been applicable because in both examples, the IPC-based search procedure returned so many results that it is not possible to identify and filter the relevant patents.

Table 1 presents the IPCs G01R 31/392 and G06F 17/5, which were also identified in the DEPATIS patent research. G01R 31/392 pertains to patents concerned with methods of individual testing related to individual cells or groups of cells within batteries. G06F 17/50 pertains to patents focused on digital computing or data processing equipment or methods specifically tailored for information retrieval, as well as database structures or file system structures for these purposes [37].

3.3. Literature Research Method

The literature review process is outlined and detailed step by step in Figure 2.

In the first step, literature research was conducted on the platforms provided by major publishing houses, as outlined by [38].

Table 2. List of chosen publishing houses and the databases used in them.

Platforms	Used Databases
1. IEEE	IEEE Xplore [39]
2. Elsevier	ScienceDirect [40]
3. Energies	CellPress [41]
4. Springer Link	Springer Nature Link [42]
5. MDPI	MDPI own database [43]
6. Wiley Online Library	Wiley own database [44]

shows the platforms used and the corresponding databases.

Each platform had unique settings for the search procedure to specifically filter suitable documents, resulting in more relevant and targeted hits. The advanced search function was utilized for this, and the keywords were searched in the “abstract”, “keyword”, and “title” fields.

On all platforms, except Springer Nature Link, the results for each keyword can be exported as .bib files. These .bib files are imported into the libraries of the literature management program, JabRef 5.12 [36], for further processing. The results from Springer Nature Link can be exported as .csv files. In these .csv files, the Digital Object Identifier (DOI) is manually extracted from each entry and inserted into the corresponding JabRef library.

According to Figure 2, all duplicates were deleted in a second step. Since each search query was conducted using seven defined keywords with similar meanings, there were partial overlaps in the results. For this purpose, a separate library was created in JabRef for each search query. Afterward, all 42 individual libraries, composed of seven keywords from six publishers, were merged, and all duplicates were removed.

In the third step, the remaining documents were filtered according to the scope of this work. Therefore, the title and abstract of all entries in the merged JabRef library were reviewed to determine whether the publication is relevant to LP of LIBs. All irrelevant documents were excluded from further evaluation.

In the fourth step, the remaining documents in the JabRef library were categorized based on the approaches used for lifetime prediction.

4. Results

This section first presents the quantitative results of the patent and literature research. Based on these findings, promising general approaches for lifetime prediction are identified and discussed.

4.1. Patent Research Results

The patent research identified 29 results, of which 4 were duplicates. Figure 3 shows the distribution of these 25 patents concerning the publication time over the last 15 years. Three patents from the initial 25 did not align with the topic related to the lifetime prediction of lithium-ion batteries and were therefore removed.

4.2. Literature Research Results

After conducting the literature research, 9948 publications were found. In the second step, all duplicates were removed, resulting in 4038 documents as listed in

Table 3. Publication overview at the individual publishers with the different keywords and how many of these were duplicates.

Publisher	Total Number	Total Number Without Duplicates
1. IEEE Xplore	2521	1243 (−51%)
2. ScienceDirect	1844	532 (−71%)
3. CellPress	3045	1272 (−58%)
4. Springer Nature Link	1706	530 (−69%)
5. MDPI	537	329 (−39%)
6. Wiley Online Library	295	139 (−53%)
Overall	9948	4038 (−59%)

.

Figure 4 illustrates the distribution of publication quantities, excluding duplicates, over time across various databases. A three-dimensional bar chart was created to quickly show how the number of publications has increased in recent years for all publisher and how the topic of lithium-ion battery lifetime prediction was not widely addressed in research before 2010.

In the third step of the literature research, the remaining 4038 results were manually checked, and all documents that were not relevant to the LP of LIB were excluded. The following examples illustrate the filtering process. Many publications refer to the State of Charge instead of the State of Health, or the application pertains to areas unrelated to mobility, such as photovoltaic systems or batteries for portable use. These documents were excluded, and only mobility-related entries were retained for further analysis.

Some publications discuss fuel cells and battery cells with new electrode materials, electrolytes, or coatings. Numerous documents were found that not only dealt with lithium-ion batteries but also with all-solid-state, sodium-ion, zinc-ion, or lithium-air batteries, as noted in the keywords from Section 3.1. This review focuses exclusively on state-of-the-art lithium-ion batteries and not on future battery technologies, which is why all publications regarding other battery technologies were excluded.

Additional exclusion criteria included health, temperature and maintenance management, as well as estimation or cost reduction topics.

The number of final publications for LP method categorization totals 633 after this filtering process.

4.3. Evaluation of Approaches and Methods

The remaining documents were sorted to determine which approach was used for each of the 22 patents and 633 publications to predict the lifetime of lithium-ion batteries. The 150 different approaches were synthesized and tabulated to create a mind map, as shown in Figure 5 and according to [23,25,27,28,29,45,46,47,48,49].

The mind map is divided into four levels from left to right. On the far left, in level 0, is the central main element Lifetime Prediction Methods, which branches into the main categories Model-Based and Data-Driven in level 1. Each of these two main categories branches off into subcategories further to the right, in Level 2. Model-Based is divided into Mechanistic/Electrochemical Model, Equivalent Circuit Model, Empirical Model, and Thermal Model. Data-Driven branches out into Machine Learning, Statistical Approach, Stochastic Approach, Intelligent Algorithm, and Time Series Analysis. In the remaining level 3+ on the right-hand side, the numbers in the ellipses indicate how many approaches or methods are hidden behind the elements from level 2. A complete overview of all approaches can be found in the Tables in the Appendix B (

Table A2. List of publications addressing various variants of clustered Model-Based methods.

Level 1	Level 2	Level 3	Source
Model-based	Mechanistic/ Electrochemical Model	- Battery Aging Mathematical Model	[72,73]
		- Incremental Capacity Analysis (ICA)	[65,74,75,76,77,78,79]
		- Differential Thermal Voltammetry Signal Analysis (DTV)	[80,81]
		- Electrochemical Impedance Spectroscopy (EIS)	[82,83,84,85,86,87,88,89,90,91]
		- Pseudo 2-dimensional Model (P2D)	[50,53,64,92,93]
		- Single Particle Model (SPM)	[54,94]
	Equivalent Circuit Model (ECM)	- Integral-Order Model (IOM)	[95,96,97,98,99,100,101,102,103,104]
		- Fractional-Order Model (FOM)	[105]
		- Lumped Parameter Model (LPM)	[106]
	Empirical Model	- Empirical Prediction Model	[52,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125]
		- Operating Condition-based Degradation Equation (OCDE)	[126,127,128]
	Thermal Model	- Reduced-Order Electrochemical Thermal Model (ROTM)	[129]
		- Electro-Thermal-Aging Model (ETA)	[130]
		- 3D Thermal Model	[50]

,

Table A3. List of publications addressing various variants of clustered Data-Driven methods (part 1).

Level 1	Level 2	Level 3	Source
Data-Driven	Machine Learning	- Vector Regression (VR)	[Details in Table A5]
		- Vector Machine (VM)	[Details in Table A6]
		- Gaussian Process (GP)	[Details in Table A7]
		- Artificial Neural Network (ANN)	[63,83,86,97,98,101,131,132,133,134,135,136,137,138,139]
		- Neural Network (NN)	[Details in Table A9 and Table A10]
		- Boosting Algorithm	[58,82,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156]
		- Fuzzy Logic	[157,158,159,160,161]
		- Regression Modeling	[Details in Table A8]
Data-Driven	Statistical Approach	- Auto Regressive Moving Average (ARMA)	[162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179]
		- Entropy Analysis	[169,170,180,181,182,183,184,185]
		- Ornstein-Uhlenbeck Process (OUP)	[186,187]
		- Grey Model (GM)	[188,189,190,191,192,193,194,195,196,197,198,199,200,201,202,203,204,205,206,207,208,209]
		- Weibull Process (WP)	[210,211]
		- Rain Flow Counting Algorithm (RFCA)	[164]
		- Moving Average Filter (MAF)	[212]
		- Vector Autoregressive Model (VAR)	[213,214]
		- Dempster-Shafer Theory (DSF)	[215]
		- Functional Principal Component Analysis (FPCA)	[216]
		- Non-Linear Least Squares (NLLS)	[217,218]
	Stochastical Approach	- Particle Filter (PF)	[Details in Table A11]
		- Kalman Filter (KF)	[Details in Table A12]
		- Bayesian Model	[80,81,110,151,219,220,221,222,223,224,225,226,227,228,229,230,231,232,233,234,235,236,237,238,239,240,241]
		- Monte Carlo Simulation	[79,81,101,103,110,165,186,202,227,234,242,243,244,245,246,247,248,249,250,251,252,253]
		- Wiener Process (WP)	[186,187,223,229,234,254,255,256,257,258,259,260,261,262,263,264]
		- Markov Chain (MC)	[193,201,251,258,265,266,267]
		- Brownian Motion (BM)	[265,268,269,270,271,272]
		- Cauchy Process	[268]

,

Table A4. List of publications addressing various variants of clustered Data-Driven methods (part 2).

Level 1	Level 2	Level 3	Source
Data-Driven	Intelligent Algorithm	- Particle Swarm Optimization (PSO)	[58,99,185,202,212,273,274,275,276,277,278,279,280,281,282,283,284,285,286,287,288,289,290,291]
		- Tuna Warm Optimization (TWO)	[292]
		- Swarm Intelligence Optimization (SIO)	[162]
		- Artificial Bee Colony Optimization (ABC)	[52,273,293,294,295]
		- Fruit Fly Optimization (FFO)	[270,296]
		- Whale Optimization Algorithm (WOA)	[162,273,297,298,299,300,301,302,303]
		- Sparrow Search Algorithm (SSA)	[60,138,162,304,305,306,307,308,309,310,311,312,313,314,315]
		- Harris Hawks Optimization (HHO)	[150]
		- Pelican Optimization Algorithm (POA)	[316]
		- Jellyfish Optimization (JFO)	[317]
		- Dung Beetle Optimization (DBO)	[318,319,320]
		- Adam Optimization Algorithm	[321,322]
		- Cuckoo Search Optimization Algorithm	[101,323,324,325,326]
		- Grey Wolf Optimization (GWO)	[156,205,327,328,329,330,331,332,333,334]
		- Artificial Fish Swarm Algorithm (AFSA)	[196,335]
		- Northern Goshwak Optimization (NGO)	[55]
		- Successive Variational Mode Decomposition (SVMD)	[292,336]
		- Feature Vector Selection (FVS)	[337,338]
		- Optimal Graph Entropy (OGE)	[184]
		- Genetic Algorithm (GA)	[99,178,339,340,341]
		- Genetic Algorithm Ant Algorithm (GAAA)	[342]
		- Beetle Antenae Search (BAS)	[108]
		- Multi-Objective Arithmetic Optimization Algorithm (MOAOA)	[343]
		- Self-Adaptive Differential Evolution (SADE)	[344]
		- Teaching-Learning Based Optimization (TLBO)	[345]
		- Jumping Spider Optimization Algorithm (JSOA)	[346]
Data-Driven	Time Series analysis	- Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN)	[103,176,177,270,274,275,290,299,311,330,331,347,348,349,350,351,352,353,354,355,356,357,358,359,360,361]
		- Variational Mode Decomposition (VMD)	[55,138,162,163,169,230,292,297,309,319,320,325,334,336,362,363,364,365,366,367,368,369]
		- Empirical Mode Decomposition (EMD)	[168,171,173,174,201,237,287,358,364,370,371,372,373,374,375,376,377,378,379,380,381,382,383,384,385,386,387,388,389,390,391]
		- Multi-Attenuation-Mode Decomposition Model (MAMDM)	[392]
		- Discrete Wavelet Transformation (DWT)	[287,364,393,394]
		- Wavelet Packet Decomposition (WPD)	[395]
		- Gamma Process	[396]
		- Aging Density Function Model (ADFM)	[397]
		- Granger Causality Test (GC)	[213]

,

Table A5. List of publications addressing various variants of Vector Regressions (VRs).

Level 4	Source
- Support Vector Regression (SVR)	[52,88,142,151,215,237,266,294,316,319,328,329,337,349,372,375,398,399,400,401,402,403,404,405,406,407,408,409,410,411,412,413,414,415,416,417,418,419]
- Least Absolute Shrinkage and Selection Operator (LASSO)	[149,216]
- Relevance Vector Regression (RVR)	[193,374]

,

Table A6. List of publications addressing various variants of Vector Machines (VMs).

Level 4	Source
- Multi-Kernel Relevance Vector Machine (MKRVM)	[205,284,285,298,300,420]
- Least Squares-Support Vector Machine (LS-SVM)	[116,156,159,176,202,270,282]
- Support Vector Machine (SVM)	[131,195,221,225,315,401,421,422,423,424,425,426,427,428]
- Radial Basis Function (RBF)	[402]
- Relevance Vector Machine (RVM)	[56,196,214,253,278,280,308,338,371,373,399,421,429,430,431,432]
- Sparse Bayesian Learning (SBL)	[190,433]
- Dynamic Grey Related Vector Machine (DGRVM)	[208]

,

Table A7. List of publications addressing various variants of Gaussian Processes (GPs).

Level 4	Source
- Gaussian Process Regression (GPR)	[74,78,83,149,160,161,207,225,240,242,284,289,296,334,347,354,358,365,382,390,403,414,417,426,434,435,436,437,438,439,440,441,442,443,444,445,446,447,448,449,450,451,452,453,454,455,456,457,458,459]
- Neural Gaussian Process (NGP)	[201,298,460,461]
- Gaussian Mixture Regression (GMR)	[198,462]
- Gaussian Sine Function, Levenberg-Marquardt (GS-LM)	[363]
- Multioutput (Concolved) Gaussian Process (MCGP)	[463,464,465]
- Multi-Model Gaussian Process (MMGP)	[119]

,

Table A8. List of publications addressing various variants of Regression Modelings.

Level 4	Source
- Multivariate Adaptive Regression Spline (MARS)	[466]
- Random Forest Regression (RFR)	[83,149,247,400,467,468]
- Random Forest Model (RF)	[141,150,214,425,450,469,470,471]
- Ridge Regression Algorithm (RRA)	[247,400]
- Linear Regression (LR)	[131,137,142,165,172,214,379,401,425,467,472]
- Linear Quantile Regression (LQR)	[172,473]
- Quantile Regression Random Forest (QRRF)	[473]
- Total Least Squares Regression Model (TLS)	[474]
- Double Exponential Capacity Degradation Model (DECDM)	[429,439,475]
- k-Nearest Neighbor Algorithm (k-NN)	[131,141,145,401,476]
- Regressive Multiple-source Domain Adaption (RMDA)	[477]

,

Table A9. List of publications addressing various variants of Neural Networks (NNs) (part 1).

Level 4	Source
- Autoencoder	[62,419,443,462,478,479,480,481,482,483,484,485,486,487,488,489,490,491,492,493,494,495]
- Gated Recurrent Unit (GRU)	[57,68,79,103,175,177,188,197,202,242,275,290,322,325,331,356,366,386,496,497,498,499,500,501,502,503,504,505,506,507,508,509]
- Convolutional Neural Network (CNN)	[55,59,60,63,68,127,155,175,220,228,235,238,351,352,355,356,362,383,386,387,388,395,428,439,457,470,476,478,485,493,502,503,505,508,510,511,512,513,514,515,516,517,518,519,520,521,522,523,524,525,526,527,528,529,530,531,532,533,534,535,536,537,538,539,540,541,542]
- Channel Attention Mechanism (CA)	[537,543,544]
- Adaptive Feature Separable Convolution (AFSC)	[545]
- Radial Basis Function Neural Network (RBF)	[108,305,402]
- Nonlinear Auto Regressive Neural Network (NARNN)	[546,547]
- Nonlinear Auto Regression Recurrent Neural Network (NAR-RNN)	[548]
- Bimodal Fusion Regression Network (BFRN)	[549]
- Temporal Convolutional Network (TCN)	[55,94,283,311,320,357,501,550]
- Long Short-Term Memory (LSTM)	[53,62,67,68,70,80,99,101,132,157,162,182,184,203,210,213,214,220,228,231,234,235,237,238,243,245,249,256,268,270,273,274,292,295,297,304,307,309,311,313,321,322,350,351,353,355,359,360,362,363,364,365,366,369,376,377,380,381,383,409,410,423,424,431,434,436,446,447,461,472,478,485,490,493,498,504,505,507,510,513,515,518,522,523,527,530,531,535,536,539,544,545,548,551,552,553,554,555,556,557,558,559,560,561,562,563,564,565,566,567,568,569,570,571,572,573,574,575,576,577,578,579,580,581,582,583,584,585,586,587,588,589,590,591]
- Mogrifier Long Short-Term Memory (MLSTM)	[138]
- Stacked Long Short-Term Memory (SLSTM)	[147,592]
- Bi Long Short-Term Memory (BiLSTM)	[55,59,60,94,230,289,303,318,327,352,358,360,375,391,398,503,511,516,520,528,529,541,543,593,594,595,596,597,598,599]

,

Table A10. List of publications addressing various variants of Neural Networks (NNs) (part 2).

Level 4	Source
- Adaptive Dropout Long Short-Term Memory (ADLSTM)	[227]
- Deep Neural Network (DNN)	[53,55,81,93,115,131,348,376,481,498,501,525,559,588,600,601,602,603]
- Quantum Reservoir Enhanced Deep Neural Network (QREDNN)	[489]
- Deep Learning Neural Network (DLNN)	[97,302,340,573]
- Deep Learning Regression Model (DLRM)	[604,605]
- Deep Cross Network (DCN)	[550]
- Feed-Forward Neural Network (FFNN)	[130,133,185,204,281,368,421,454,487,606,607,608,609,610]
- Cascaded Forward Neural Network (CFNN)	[317]
- Lightweight Neural Network (LNN)	[100]
- Extreme Learning Machine (ELM)	[254,278,286,291,299,305,306,310,330,332,342,343,378,611,612,613,614]
- Random Vector Functional Link Network (RVFLN)	[546]
- Wavelet Neural Network (WNN)	[393,615]
- Recurrent Neural Network (RNN)	[57,68,184,245,277,321,424,428,498,500,507,526,538,554,557,563,616,617]
- Bayesian Neural Network (BNN)	[181,397,618,619]
- Elman Neural Network (ENN)	[88,334,380,383,620]
- Broad Learning System (BLS)	[373,561]
- Temporal Transformer Network (TTN)	[111,479,621]
- Graph Convolutional Network (GCN)	[622]
- (Monotonic) Echo State Network (MESN)	[293,344,565,623]
- Deep Residual Shrinkage Network (DRSN)	[560]
- Deep Belief Network (DBN)	[56,567]
- (Denoising) Transformer-based Neural Network (DTNN)	[486,624]
- Flexible Parallel Neural Network (FPNN)	[625]
- Spatio-Temporal Multimodal Attention Network (ST-MAN)	[626]

,

Table A11. List of publications addressing various variants of Particle Filters (PFs).

Level 4	Source
- Particle Filter (PF)	[51,52,54,84,92,104,107,108,109,110,112,118,120,154,157,163,189,204,215,217,226,253,263,276,288,308,323,324,326,345,346,354,370,383,390,404,405,406,407,437,482,572,581,601,608,619,627,628,629,630,631,632,633,634,635,636,637,638,639,640,641,642,643,644,645]
- Unscented Particle Filter (UPF)	[115,158,200,251,255,257,279,293,347,393,399,615,646,647,648,649]
- Regularized Particle Filter (RPF)	[171,650]
- Auxiliary Particle Filter (APF)	[475]
- Second-Order Central Difference Particle Filter (SCDPF)	[651]
- Converted Sampling Particle Filter (CSPF)	[127]
- Improved Mutated Particle Filter (IMPF)	[117]
- Inheritance Particle Filter (IPF)	[339]
- Grey Particle Filter (GPF)	[652]
- Enhanced Mutated Particle Filter (EMPF)	[653]
- Linear Optimization Resampling Particle Filter (LORPF)	[199]
- Spherical Cubature Particle Filter (SCPF)	[654]
- Double Exponential Empirical Particle Filter (DEEPF)	[655]
- Gauss-Hermite Particle Filter (GHPF)	[656]

and

Table A12. List of publications addressing various variants of Kalman Filters (KFs).

Level 4	Source
- Kalman Filter (KF)	[52,271,276,432,435,452,630]
- Unscented Kalman Filter (UKF)	[50,106,113,158,185,217,371,374,399,408,541,596,606,609,646,657,658,659]
- Extended Kalman Filter (EKF)	[95,102,114,118,154,383,635,641,650,658]

).

To improve the lifetime prediction process, it is possible to combine methods. This is called the Fusion Approach, which can be divided into Model-Based and Data-Driven Fusion [50,51,52,53,54] or Multiple Data Driven Fusion [55,56,57,58,59]. In this work, the Fusion Approach is not mentioned in level 1 of the mind map (Figure 5) as a third part because the focus is placed on the methods rather than on the combination of several approaches from the areas of Model-Based and Data-Driven.

The next step is to determine how often an approach has been used. Figure 6 shows the distribution of machine learning methods in relation to the time of publication over the last 15 years. The use of machine learning methods has increased more than sevenfold in recent years and has therefore become especially important in this area of research. The individual methods that belong to machine learning are shown in Table A3 and more in detail in Table A5, Table A6, Table A7, Table A8, Table A9 and Table A10. For this reason, the various Machine Learning methods are presented in more detail in Figure 7.

Figure 5 illustrates that Stochastic Approaches exert a consistent and slightly increasing influence on the selection of lifetime prediction methods. For a more thorough evaluation, Figure 8 presents the various methods, which are concealed in the mind map under stochastic approaches at level 3+, separately.

An upward trend in Intelligent Algorithms is also visible over the period under review in Figure 6. These algorithms are utilized in conjunction with other methods to enhance model accuracy, improve convergence, and mitigate local optima caused by manual parameter tuning, leading to more precise lifetime predictions [60]. This combination is referred to as multiple data-driven fusion in the realm of fusion approaches. The integration of Intelligent Algorithms, detailed in Table A4, with data-driven models is advantageous for battery lifetime predictions due to the aforementioned benefits. When used synergistically, these methods yield more accurate and efficient predictions by leveraging the strengths of each algorithm. Data-driven models, particularly neural networks, are fundamentally adept at handling complex non-linear relationships, improving battery life predictions. Combined with the optimization capabilities of intelligent algorithms, this could lead to highly accurate forecasts.

Figure 7 illustrates the individual components of the Machine Learning domain, where the rapid rise of Neural Networks is notably evident. The significant growth in the field of machine learning, as presented in Figure 6, in recent years is primarily attributed to the adoption of neural networks. In light of this remarkable increase in Neural Networks for battery lifetime prediction, these methods are being further explored. An overview of the 38 different variations and extensions that correspond to level 4 of the mind map is available in Table A9 and Table A10.

In the Stochastic Approaches, the Particle Filter dominates this area, as can be seen from the height of the bar charts in Figure 8. In the literature research, the area of particle filters is discussed in more detail in Table A11, which lists 14 different variations and extensions.

The evaluation of the most commonly used approaches in Figure 6 includes Machine Learning, Stochastic Approaches, and Intelligent Algorithms. Based on this, the top three methods with the most applications out of 22 patents and 633 publications are in the areas of neural networks, intelligent algorithms, and particle filters.

Table 4. Overview of the top 3 methods used in the research in the respective categories.

Category	Top 3 Methods
Neural networks based on Table A9 and Table A10	1. Long Short-Term Memory (LSTM): 134 times
	2. Convolutional Neural Network (CNN): 67 times
	3. Gated Recurrent Unit (GRU): 32 times
Particle filter based on Table A11	1. Particle Filter (PF): 65 times
	2. Unscented Particle Filter (UPF): 16 times
	3. Regularized Particle Filter (RPF): 2 times
Intelligent algorithms based on Table A4	1. Particle Swarm Optimization (PSO): 24 times
	2. Sparrow Search Algorithm (SSA): 15 times
	3. Grey Wolf Optimization (GWO): 10 times

provides an overview of the frequency of methods used in the top three categories. Additionally, Figure 9 illustrates how often these methods were employed relative to the review period. It can be observed that Particle Filters were used less frequently, while Long Short-Term Memory and Convolutional Neural Network became more prominent. Conversely, the use of Intelligent Algorithms remained fairly steady.

Section 4.2 “Literature Research Results” describes the sorting process of the 4038 publications. There is also a category for publications related to mobility applications. In this sorting process, 37 publications were identified that fit into the area of lifetime prediction of lithium-ion batteries and also describe an application on or with electric vehicles.

Table A13. Overview of the 27 identified publications that refer to the area of application in mobility and the approaches used for lifetime prediction.

Authors, Publication and Source	Used Approach
Haifeng et al., October 2009, [95]	Equivalent Circuit Model + Kalman Filter
Sarasketa-Zabala et al., November 2013, [121]	Empirical Model
Han et al., October 2014, [180]	Statistical Approach
May et al., October 2017, [98]	Equivalent Circuit Model + Artificial Neural Network
Lipu et al., December 2018, [46]	Review
Ibanez et al., November 2019, [96]	Equivalent Circuit Model
Verma et al., January 2020, [129]	Thermal Model
Deng et al., November 2020, [226]	Particle Filter
Audin et al., December 2021, [62]	Long Short-Term Memory
Khodadadi Sadabadi et al., January 2021, [54]	Mechanistic/Electrochemical Model + Particle Filter
Yao et al., August 2021, [23]	Review
Gong et al., October 2021, [68]	Long Short-Term Memory
Wang et al., December 2021, [165]	Regression Modeling + Auto Regressive Moving Average + Monte Carlo Simulation
Zhang et al., February 2022, [512]	Convolutional Neural Network
Zhang et al., April 2022, [660]	Review
Xu et al., August 2022, [564]	Long Short-Term Memory
Feng et al., October 2022, [450]	Gaussian Process + Regression Modeling
Deng et al., November 2022, [110]	Particle Filter
Elmahallawy et al., November 2022, [28]	Review
Pradeep et al., November 2022, [425]	Vector Machine + Regression Modeling
Suresh et al., November 2022, [63]	Artificial Neural Network + Convolutional Neural Network
Sharma et al., December 2022, [661]	Review
Von Bülow et al., January 2023, [662]	Review
Ha et al., June 2023, [467]	Regression Modeling
Liu et al., September 2023, [424]	Vector Machine + Long Short-Term Memory
Liang et al., October 2023, [469]	Regression Modeling
Wang et al., December 2023, [101]	Equivalent Circuit Model + Long Short-Term Memory + Monte Carlo Simulation + Cuckoo Search Optimization Algorithm
Von Bülow et al., February 2024, [554]	Long Short-Term Memory
Li et al., May 2024, [209]	Grey Model
Mishra et al., June 2024, [65]	Incremental Capacity Analysis
Reza et al., June 2024, [33]	Review
Alsuwian et al., July 2024, [663]	Review
Kang et al., July 2024, [64]	Electrochemical Circuit Model
Yifan et al., October 2024, [605]	Deep Learning Regression Model
Singh and Reddy, October 2024, [67]	Long Short-Term Memory
Zhang et al., November 2024, [34]	Review
Ansari et al., December 2024, [664]	Review
Sang et al., December 2024, [665]	Review

in the Appendix C shows these publications and the LP approaches used. The top three most frequently utilized approaches in this specific category of publications are listed in

Table 5. Overview of the methods used in the specific category of publications with application in mobility.

Category	Most Used Methods
Neural networks based on Table A9 and Table A10	Long Short-Term Memory (LSTM): 6 times Convolutional Neural Network (CNN): 2 times
Particle filter based on Table A11	Particle Filter (PF): 2 times
Intelligent algorithms based on Table A4	Cuckoo Search Optimization Algorithm: 1 time
Additionally used: Model-based based on Table A2	Equivalent Circuit Model (ECM): 4 times
Machine learning based on Table A3	Regression Modeling (RM): 3 times Artifical Neural Network (ANN): 2 times Vector Machine (VM): 2 times

.

The approaches used in the publications regarding their application in mobility (see Table A13) and the evaluation results in Table 4 agree on the use of neural networks, specifically Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) methods, as well as particle filters utilizing Particle Filter (PF) without modifications.

Based on these findings, Long Short-Term Memory, Convolutional Neural Network, and Particle Filter are discussed in order to develop a concept for lifetime prediction of onboard traction batteries during a technical inspection.

4.3.1. Long Short-Term Memory

Long Short-Term Memory is specifically designed to analyze time-dependent data. Its unique features, such as handling long-term dependencies and addressing the vanishing gradient problem with three control gates (input, forget, and output), make it particularly well-suited for modeling battery degradation and predicting lifecycle trends. LSTM is widely utilized to develop degradation models that accurately estimate the remaining service life of batteries. The algorithm’s ability to incorporate both historical data and online operational data enhances prediction accuracy. Transfer learning methods further refine these models, allowing for improved estimation across various battery types, operating conditions, and production batches. LSTM is also frequently combined with other techniques such as Empirical Mode Decomposition (EMD), auto-encoders, or Convolutional Neural Networks (CNN) to optimize predictions. Various battery parameters, including charge/discharge curves, voltage, current, and capacity, are used as input data to create highly precise models [23,46,61,62].

4.3.2. Convolutional Neural Network

Convolutional Neural Networks (CNNs), a specialized subtype of neural networks (NNs), are commonly employed for data analysis in grid-like architectures. They have proven to be highly effective in tasks such as numerical regression and machine vision. A defining feature of CNNs is their ability to implement a specialized linear operation, known as convolution, in at least one layer to process data. This technique is versatile: 1D grids can represent time-series data, such as energy usage, while more complex data, like images, are represented in 2D grids. The core functionality of CNNs lies in feature extraction. Convolution and subsampling are performed in cascaded stages, and the extracted features are fed into fully connected (FC) layers, which function similarly to standard artificial neural networks (ANNs). A CNN model comprises four primary components. Training a CNN from scratch can be challenging due to the need for a large dataset and its time-consuming nature. Advantages of CNNs include their proficiency in nonlinear mappings, superior performance in numeric prediction with a 1D setup, and their capacity to simplify categorization, eliminating the need for generated features. Core components of a CNN include the convolution layer, pooling layer, activation function, and fully connected layer [23,63].

4.3.3. Particle Filter

The Particle Filter (PF) method, despite its simple principle and unrestricted model noise, requires a large number of training samples and involves significant computational demands. It leverages the sampling importance resampling algorithm and a set of weighted particles to approximate the probability density function (PDF). This method has been utilized to estimate key battery parameters, such as state of charge and maximum available power, using non-linear dynamic models [23,46].

5. Applicability of Lifetime Prediction Approaches in Technical Inspections

Evaluating the applicability of lifetime prediction methods in the context of technical inspections requires a prior understanding of their procedures, framework conditions, and applicable requirements.

The term technical inspection refers to a general category of vehicle assessments, which include vehicle appraisals and standardized roadworthiness tests. In vehicle appraisals, qualified experts evaluate the condition of the vehicle, document existing and prior damage—whether already repaired or still unrepaired, and derive estimates for accident-related repair costs. This includes calculations for repair duration, possible compensation for value loss, and an assessment of the vehicle’s diminished value. The vehicle’s overall residual value is also determined based on its technical state and visual condition [7,8].

In the context of regulatory inspections, technical inspections such as the periodical technical inspection aim to ensure that vehicles remain safe, roadworthy, and compliant with environmental and legal standards. In Germany, this is governed by the Road Traffic Licensing Regulation (StVZO) and is carried out by authorized testing organizations [10]. Deficiencies are classified as minor, major, or dangerous. Early detection of safety or emissions-related defects helps prevent accidents and reduce environmental impact [9].

Based on this understanding, the identified LP approaches can then be evaluated in terms of how they may contribute to or support these inspection processes. This section aims to highlight the most promising approaches for further research and to identify remaining research gaps.

5.1. Technical Inspection Procedure

To further develop these technical inspections and integrate a new testing method, it is necessary to understand the processes involved and consider the current conditions of the test equipment. The information and procedures described below come from research and surveys conducted at TUEV SUED Division Mobility. The vehicle is located at a test station where various components are also tested using different tools. For this purpose, there is an inspection catalog for the components and systems that includes the corresponding inspection criteria in the vehicle appraisal or PTI area [10]. Currently, there are no test criteria or procedures for predicting the lifetime of onboard traction batteries. The following areas are included: identification of the vehicle, braking equipment, steering, visibility, lighting equipment, parts of the electrical system, axles, wheels, tires, suspension, chassis, chassis attachments, other equipment (e.g., warning triangle, first aid kit), and nuisance.

The process of a technical inspection follows a standardized procedure and takes approximately 30 min to complete.

At the beginning, the vehicle is identified by comparing the vehicle identification number (VIN), license plate number, and vehicle documents. This information is used to register the vehicle in the inspection system and can be retrieved from the diagnostic software for vehicle-specific test items.

Next, an external visual inspection is carried out. This includes checking the condition and functionality of the bodywork, lighting systems, windows, wheels, and tires for possible damage. An interior inspection follows, testing the function of seat belts, control elements, indicator lights, windshield wipers, and the horn. As part of the lighting inspection, all lighting components are also checked. A key part of the inspection is the functional check of safety-related systems such as steering, brakes, axles, and suspension components. To objectively evaluate braking performance, the vehicle is tested on a brake test bench. During this test, braking force, distribution, and balance are measured and analyzed using a diagnostic adapter.

Modern vehicles also utilize electronic diagnostic systems at several points during the inspection. By using the standardized onboard diagnostic 2 (OBD2) interface, inspectors can access the various control units in the vehicle. Suitable diagnostic adapters support the inspection systems by extracting data, for example, during the brake test, lighting check, and emissions inspection.

The emissions test, which is only necessary for ICE vehicles at PTI, ensures compliance with legally defined emission limits. It is carried out either by direct measurement using sensors or by reading the vehicle’s emission data via the OBD2 interface. During vehicle appraisals, a visual and functional inspection of the engine is conducted regarding noises, leaks, and diagnostic trouble codes. The lifetime prediction can be integrated here instead of the emissions and engine test to avoid any recognizable changes to the fixed TI process.

Finally, an underbody inspection is performed to check for leaks, corrosion, brake lines, structural components, and the condition of the exhaust system for ICE vehicles, or a visual inspection of the battery for BEVs. Once all inspections have been completed, the defects found are classified, the vehicle value is determined, and an inspection report is drawn up.

Based on the technical inspection process described above and the usual computational performance of electronic devices, the following key requirements for the lifetime prediction of traction batteries emerge as essential for various applications, according to [24].

5.2. Applicable Lifetime Prediction Approaches for Technical Inspection

In the future, testing organizations will expand their product portfolio regarding the lifetime prediction of onboard traction batteries and integrate it into existing services. These products and services will further develop technical inspections [22]. To find an existing LP approach that fits with the further developed TIs, the entries in the final literature library referring to applications in electric vehicles in the title and abstract were filtered out and discussed. The processes and requirements of these LP approaches were compared and discussed concerning the boundary conditions during a TI from

Table 6. Overview of the boundary conditions to be fulfilled for the integration of an LP for LIB in the TI.

Considered Topics	Boundary Conditions
Testing restrictions:	No error codes permitted in the display
	No test drive
	Non-destructive
Usable vehicle interfaces:	OBD2 interface, charging socket
Duration for data gathering:	10 min
Duration for calculation:	15 min
Available test equipment:	Laptop, tablet, cloud, diagnostic adapter, Wifi
Accuracy of the result:	±5%
Computational power laptop
CPU:	4C/8T, 1.9–4.2 GHz (i7-8650U)
RAM:	16 GB
Computational power tablet
CPU:	8C, 2.0–2.4 GHz (Exynos 1380)
RAM:	8 GB
Computational power cloud
CPU:	4–32 vCPUs, approx. 2.8–3.4 GHz
RAM:	32–128 GB

.

Kang et al. published in [64] the State of Health prediction of a used 85 kWh battery from an electric bus. To achieve this, Kang et al. disassembled the battery pack to obtain the individual cells. The disassembled pouch cells were used to measure the DC resistance and the available capacity at a 1C rate and 25 °C. A cycle performance prediction equation based on an equivalent circuit model (ECM) was derived to describe cycling performance and enable accurate SoH prediction. To do this, 250 cycles were carried out with the individual cells, and the ECM was trained on this data. This trained model can provide a SoH prediction with an accuracy of less than 5% for the 250 cycles using data from 20 cycles. Kang et al. approach will meet the accuracy requirements from Table 6, but cannot be integrated into a TI because it assumes the possibility of carrying out only 20 full cycles if an ECM is already trained in the background for the specific battery. Furthermore, removing the traction battery during TI is not permitted; additionally, not all inspection sites are equipped with intelligent DC charging stations that maintain the charging power at 1C, and no inspection station can maintain the necessary temperature conditions.

Reza et al. published in [33] recent advances in RUL prediction of LIB in electric vehicle applications, but conducted the implementation procedures on a test bench platform rather than in operating vehicles. Reza et al. point out that many RUL prediction approaches exist, yet each method suffers from a lack of data from real-world BEV applications. Additionally, Reza et al. describe the particle filter as a possible RUL prediction approach, noting that this requires a powerful computer processor. However, no exact value is given for a powerful computer processor, making it impossible to evaluate with the values from Table 6. Reza et al. discuss some issues important for integration into real-world applications, but do not provide solutions.

Mishra et al. presented in [65] an automated machine learning (AutoML) approach for RUL prediction of LIB in electric vehicles. The AutoML methods were trained with data sets from 4 batteries within the publicly available NASA test bench data set, not with real vehicle data. Mishra et al. disclosed the computational power used, for instance, an Intel Xeon processor with 2.5 GHz. This information allows for comparison on whether the existing devices from Table 6 would be sufficient for this AutoML. Mishra et al. plan to refine the proposed model with data from real-world usage scenarios in typical urban driving conditions, but do not provide detailed information or refer to the application of RUL prediction during a technical inspection.

Dineva published in [66] an evaluation of advances in battery health prediction for electric vehicles. Dineva highlights the limitations of current methods for RUL prediction, particularly the challenges of transferring laboratory-based models to real-world applications and the high computational demands of advanced approaches. Dineva introduces, among other things, a Long Short-Term Memory network that improves prediction accuracy but increases computational complexity, posing difficulties for real-time implementation. No detailed information is provided on the increased computing power and the challenges of real-time application, making it impossible to evaluate whether the boundary conditions from Table 6 would fit. Dineva does not refer to an application during an inspection but focuses solely on the application in electric vehicles.

Sing and Reddy published in [67] a deep learning architecture for predicting electric vehicle battery capacity degradation using the Internet of Things (IoT). They present an LSTM network for training, testing, and using in the deep learning architecture, which estimates the remaining useful life of an electric vehicle battery pack. For training and testing of the LSTM approach, a battery module was used under constant conditions in a laboratory rather than in a vehicle application. Sing and Reddy describe IoT and its environment in very general terms but do not elaborate on what would be necessary for practical application. There is also no information on additional boundary conditions that could be used to evaluate an application in a technical inspection.

Gong et al. published in [68] a data-driven LIB RUL prediction based on the operating data of actual electric vehicles. The new energy vehicle national monitoring and management platform in China provided the battery data, following the national standard GB/T 32960-2016 [69]. Gong et al. evaluated, based on real-world data, various approaches such as CNN, LSTM, and GRU for remaining useful life prediction. LSTM, CNN, and GRU are also the most frequently found LP methods in this work and align with the chosen approaches of Gong et al. Still, the focus of Gong et al. is on the results of the different approaches rather than the operating conditions during a technical inspection. The RUL prediction process is not discussed, and no further information on the process is provided.

Based on the discussed publication, various promising approaches for predicting the lifetime of LIB during a TI have emerged. However, none of the discussed publications offer an applicable process for use in a technical inspection. While all the documents found provide extensive information about real-world battery data or mention it as future work, this information has not yet been implemented. Therefore, it is currently impossible to apply the results of the patent and literature research to identify methods for predicting the lifetime of traction batteries or to evaluate their applicability within the context of technical inspections, which is the aim of this review.

Further work is necessary to fill the existing research gap and to define the framework conditions that are truly required and feasible for integrating a lifetime prediction approach into an inspection procedure. This entails clarifying how much battery data is needed, whether in partial or full cycles, and determining the duration for collecting this data in order for prediction models to yield reliable results. Additionally, it is crucial to evaluate whether the computing power available at inspection stations, such as that provided by laptops, tablets, or cloud computing services, is adequate.

Another key aspect is specifying which battery data should be utilized. This includes defining the type, resolution, and quality of the data. Without clear standards, models cannot operate reliably. The data transfer process also needs consideration, both from the vehicle to local devices like laptops or tablets and, if necessary, to cloud systems.

Environmental factors must be considered when transitioning laboratory conditions to real-world inspection settings. Laboratory measurements are typically conducted under controlled conditions, such as constant temperature, humidity, and defined charging or discharging power. However, inspection stations operate under varying conditions. These differences can impact the accuracy of prediction models and must be taken into account during model development and application.

5.3. Possible Procedure for Technical Inspection with Lifetime Prediction

A possible procedure for a TI to predict the onboard traction battery lifetime is explained below, according to [70]. Figure 10 illustrates the process from data collection to the resulting lifetime prediction value. The individual steps, “Data Gathering”, “Data Transportation”, and “Data Processing” are described. The results from the patent and literature research are utilized to develop an overview of each process step.

Starting with the top of Figure 10 “Data Gathering”, where the collection of battery related data is shown. Battery related data are divided into two different kinds of necessary data. Historical data, which are stored in the vehicle control units, such as the battery management system (BMS), can be read out with a diagnostic device via the OBD2 interface, for example. This data is important for minimizing the battery capacity prediction error because it allows for conclusions about the battery’s usage in the past [71]. Dynamic real-time battery data provides insights into the current condition and can also be generated through a charging hub or a test drive, for example. In summary, the first step is to collect data, which is then processed further.

However, significant challenges arise in practical applications. Communication with the BMS is not standardized across vehicle manufacturers, and the availability, the amount of data per unit of time (i.e., the sampling frequency), and the quality of the data vary considerably. Many relevant parameters are not freely accessible on the CAN bus, meaning that a non-discriminatory and manufacturer-independent access to the required battery data is currently not guaranteed. As a result, data collection during inspections is heavily dependent on manufacturer authorizations or reverse engineering measures, which currently limits the feasibility and reproducibility of standardized test procedures.

The next step is to transfer the data from the vehicle to the lifetime prediction model. A vehicle communication interface (VCI) is required to receive the data via a vehicle interface, such as the OBD2 interface, for example. Afterwards, the data is sent to a cloud-based data platform. The entire method could be located on this data platform. Yao et al. describe the development of cloud platform technologies and also the communication between VCI and the platform as a challenge in the LP process [23].

In practice, additional challenges arise from the different possible transmission pathways. The VCI can send data either via a local WLAN connection at the inspection site or through a SIM-card-based mobile data connection. Both approaches introduce operational constraints: a stable WLAN connection must be available at all times, or, alternatively, sufficient mobile network coverage must be ensured for the SIM-based communication. Interruptions in either communication channel can delay or prevent data transmission, thus affecting the reliability and efficiency of the entire lifetime prediction process. This also includes the implementation of robust cybersecurity measures, such as encrypted data transmission, secure authentication, and protection against unauthorized access, to ensure that sensitive vehicle and battery data is transmitted securely and cannot be intercepted or manipulated.

The third and last section “Data Processing” shows the four steps of data processing at the bottom of Figure 10. “Data Management” includes the optimization of input data, starting with the translation and conversion of data from hexadecimal code to readable numbers. Afterwards, the data is cleaned, and unnecessary information is removed. Intelligent algorithms, which are listed in Table A4, could be a great approach for these tasks. The data is filtered in the “Data Filtering” area. By selecting the best and most relevant data, the risk of misinterpretation is reduced. Data analysis can be handled by stochastic approaches mentioned in Table A3, such as the particle filter and its various modifications in Table A11. Before there is a result, the prepared data is processed in the “Data Modeling” area. Models recognize patterns and correlations in the datasets and can provide a lifetime prediction based on this. Machine learning approaches, shown in Table A3 and more in detail in Table A5, Table A6, Table A7, Table A8, Table A9 and Table A10, can be particularly suitable for this task.

6. Conclusions

Lifetime prediction of traction batteries plays a crucial role in alleviating societal concerns about electric vehicles, ensuring their safe operation, and facilitating the successful market adoption of electromobility. In this context, the present review offers a structured overview of existing methods and approaches for lithium-ion battery lifetime prediction.

The examination of a broad range of scientific literature and patents demonstrates the availability of numerous well-established methods for lifetime prediction. An extensive analysis reveals various proven methods, including model-based, data-driven, and fusion approaches, which reflect both the high degree of innovation and the scientific significance of the topic. The patent and literature research utilized defined keywords across various databases and publishers, followed by a filtering process to ensure a focused selection of the most relevant contributions. Therefore, while the results of this work may not encompass all available methods, they likely identify the best for the intended use case.

A more detailed examination indicates that most existing methods are designed for use during routine operation or in laboratory settings. However, there is a notable lack of approaches specifically tailored for practical use during technical inspections, especially under constraints of limited time and access in an onboard vehicle context. The transition from lab-based methods to real-world applications presents significant challenges, as highlighted by this work’s discussions. Addressing this gap is crucial, as it directly affects the feasibility of objective battery assessments during vehicle inspections. Such assessments are essential not only for safety evaluations but also for determining the battery’s residual value, which is a key factor in appraising the value of used battery electric vehicles.

Future research should focus on developing reliable, efficient, and easily integrated methods that enable meaningful lifetime prediction during technical inspections. Specifically, future approaches should be designed to work with limited data availability from the OBD2 interface and by using less than 30 min of generating and evaluating traction battery data during the inspection, relying exclusively on historical BMS operating data without the need for invasive measurements, or relying on data from the vehicle manucfacturer´s backend. These solutions should also be able to be integrated into commercially available diagnostic devices so as not to further increase the number of devices inspectors have to use. This approach can contribute to the advancement of standardized testing procedures and foster greater confidence in electric mobility. Additionally, establishing standardized definitions and metrics, such as the State of Health (SoH), is vital to ensure transparent and comparable results across various studies and applications. A common understanding of these key performance indicators will further support the objective evaluation of battery systems and promote harmonization across research and industry.