ML models have demonstrated outstanding performance in diagnosing vehicle engine faults, primarily due to their ability to learn complex patterns from large-scale sensor data. This capability enables more accurate and automated fault detection compared with traditional rule-based methods.

Du and Wei [4] proposed a method for detecting vehicle engine faults using analytic hierarchy process (AHP) and neural network. They first used AHP to determine the optimal input variables for vehicle engine fault diagnosis. Then, they maximized the performance of the model by tuning the hyperparameters of the neural network through a gradual growth method. Nixon et al. [9] proposed an ML approach to diesel engine health prognostics using engine controller data. They employed a random forest-based fault diagnosis model and analyzed the correlations between the input variables. Li et al. [31] proposed a fault diagnosis for vehicle engine using an ensemble learning-based ML model. They combined the probability values derived from random forest, k-nearest neighbor, and XGBoost, and classified the fault type based on the highest value among the average probabilities. Akbalık et al. [2] used sound signals for detecting vehicle engine faults. They preprocessed the sound data using discrete wavelet transform and classified the fault types using an extreme learning machine.

However, these ML models are inherently limited in their ability to provide transparent explanations for their diagnostic decisions. In the context of vehicle engine fault diagnosis, it is essential not only to detect the presence of a fault but also to identify the specific input variables that contributed to the prediction, as such information supports appropriate repair and maintenance strategies. Conventional ML models function as black boxes, lacking interpretability regarding their internal reasoning processes. As a result, they fail to reveal which features played a decisive role in the diagnosis, thereby constraining their practical applicability in real-world maintenance scenarios where understanding the root cause of failure is critical.

With recent advancements in XAI, fault diagnosis models that incorporate XAI methods have gained increasing attention. These approaches aim to identify the key input variables that significantly contribute to fault occurrences. By providing such insights, they enhance the interpretability of the diagnostic process, enabling users to better understand and trust model predictions.

Hasan et al. [32] developed an explainable fault diagnosis model for bearing systems. They first used Boruta and Spearman’s rank correlation coefficient to select key input variables. Then, they constructed k-nearest neighbor-based diagnosis model and analyzed the feature contributions of the model using SHAP. Jang et al. [33] proposed an XAI method for fault diagnosis of industrial processes. They interpreted the diagnosis model based on adversarial autoencoder using SHAP. They performed hierarchical clustering based on the local SHAP values of faulty samples to intuitively examine the influence of input variables. Zereen et al. [23] proposed an XAI method for fault diagnosis using audio data which was collected from machine sensors. They extracted features of the audio data in terms of time and frequency. Then, they constructed a fault diagnosis model using logistic regression and analyzed its explainability using SHAP and LIME, comparing the two methods.

As research in this domain progresses, the focus has gradually shifted from solely improving diagnostic accuracy to enhancing model transparency and usability in real-world applications. The integration of XAI methods into fault diagnosis enables more informed decision-making by bridging the gap between model outputs and human understanding. Such developments not only support practical maintenance strategies but also lay the foundation for more interpretable diagnostic systems.

VLMs have recently emerged as powerful tools capable of jointly processing and reasoning over both visual and textual information. Although they were initially developed for multimodal applications such as image captioning and visual question answering, their use has expanded to areas involving structured data analysis and decision-support tasks across diverse domains. These models exhibit strong capabilities in extracting contextual indicators, generating human-readable explanations, and enhancing interpretability within complex data-driven systems.

Roberts et al. [34] proposed Image2Struct, a benchmark designed to evaluate VLMs in extracting structured representations from images. Their method enables automated and quantitative assessment by reconstructing image structures (e.g., LaTeX, HTML) and comparing them with the original data using multiple similarity metrics. Evaluations across diverse domains, including webpages, mathematical equations, and musical scores, revealed significant performance variations, demonstrating the benchmark’s effectiveness in distinguishing VLM capabilities. Xia et al. [35] evaluated VLMs on spreadsheet comprehension through self-supervised tasks involving optical character recognition (OCR), spatial reasoning, and format perception. While VLMs exhibited strong OCR performance, they struggled with issues such as cell omission, misalignment, and limited spatial understanding, indicating the need for further improvements. Chen et al. [36] proposed Ocean-OCR, a large-scale VLM that achieves outstanding performance in OCR tasks, including document understanding, scene text recognition, and handwriting interpretation. Leveraging a native resolution vision transformer and extensive OCR datasets, Ocean-OCR outperformed professional systems such as TextIn and PaddleOCR while maintaining robust general multimodal reasoning capabilities.

Although VLMs have demonstrated remarkable progress in analyzing both image and text data, their application to interpreting XAI outputs remains underexplored. In particular, few studies have investigated the use of VLMs to generate textual explanations for vehicle engine fault diagnosis models by integrating insights from multiple XAI methods. To bridge this gap, our method leverages VLMs to translate LIME and SHAP outputs from ML-based diagnostic models into coherent textual explanations, which are subsequently synthesized by an LLM into a unified diagnostic report. This translation process mitigates the cognitive burden of interpreting complex visual outputs and enhances user accessibility.

LLMs are advanced neural architectures trained on massive textual corpora, enabling strong capabilities in language comprehension and generation. Beyond producing coherent natural language, LLMs excel at synthesizing information from diverse sources into structured and comprehensive narratives. These capabilities make LLMs particularly well-suited for applications that require the integration of heterogeneous outputs into unified reports, thereby enhancing interpretability, knowledge accessibility, and practical decision support across domains.

Xie et al. [37] investigated whether LLMs can simulate human trust behavior. Their study revealed that LLMs generally exhibit trust behavior under the framework of Trust Games, which are widely recognized in behavioral economics. Tu et al. [38] evaluated the evolution of ChatGPT across time. They constructed ChatLog, an ever-updating dataset with large-scale records of long-form ChatGPT responses across 21 natural language processing (NLP) benchmarks. Their comprehensive evaluation demonstrated that most capabilities of ChatGPT have progressively improved over time, exhibiting a stepwise evolutionary pattern. Sui et al. [39] developed a benchmark to assess the capability of LLMs to interpret tabular data. Their study included 7 tasks, such as cell lookup and row retrieval, to evaluate the structural understanding capabilities of LLMs. They observed that performance varied depending on various input factors, including table input format, content order, role prompting, and partition marks.

Despite the advancements in ML-based fault diagnosis, the integration of LLMs for interpreting model predictions remains limited. In particular, their potential to consolidate heterogeneous explanation outputs into a unified and accessible format has not been extensively explored. In this paper, we propose a novel VLM-LLM integrated method in which VLMs first generate textual interpretations of XAI outputs such as LIME and SHAP, and LLMs subsequently synthesize these complementary explanations into a coherent diagnostic report. By combining the visual interpretability of XAI with the integrative narrative capabilities of LLMs, the proposed method alleviates the cognitive burden of interpretation and enhances decision support in vehicle engine fault diagnosis.

The ML model is formulated as a classification system to diagnose vehicle engine faults based on sensor data. Specifically, we employ the XGBoost classifier [30], a gradient boosting-based model known for its outstanding predictive performance in fault diagnosis tasks [40,41]. XGBoost constructs an ensemble of decision trees in a sequential manner, where each subsequent tree focuses on minimizing the residual errors of the preceding ones, thereby enhancing the overall model accuracy.

A notable advantage of XGBoost is its ability to effectively handle sparse data through sparsity-aware split findings. This mechanism enables robust learning even when the dataset contains incomplete or irregular sensor readings, which are common in real-world vehicle diagnostics. In addition, XGBoost supports both first- and second-order gradient optimization, allowing the model to capture not only the direction but also the curvature of the loss function. This feature improves convergence speed and enhances optimization stability, which is useful in complex fault pattern classification tasks. Another important strength is the inclusion of L1 and L2 regularization terms in the objective function. This helps control model complexity and reduce overfitting, thereby improving generalization to unseen data. Furthermore, XGBoost employs a weighted quantile sketch algorithm to efficiently select split points in large datasets. This reduces both memory consumption and computational cost, making the model suitable for processing large-scale vehicle sensor data.

To enhance the performance of vehicle engine fault diagnosis, we fine-tune key hyperparameters, including learning rate, tree depth, and the number of estimators using grid search. This tuning process is guided by validation loss, and early stopping is employed to prevent overfitting, thereby achieving a well-balanced trade-off between model complexity and generalization capability. As a result, the optimized XGBoost effectively discriminates among different engine fault types while maintaining high predictive reliability under varying input conditions.

Previous studies have validated the effectiveness of XGBoost in vehicle engine fault diagnosis. For instance, Tao et al. [42] applied XGBoost to detect misfire conditions in diesel engines, achieving outstanding accuracy compared to traditional ML models. Similarly, Pratap et al. [43] developed an AI-based vehicle fault diagnosis system that employed XGBoost to analyze two-wheeler vehicle signals, demonstrating strong robustness in real-world applications. These findings support our choice of XGBoost as the core classifier in the proposed method.

LIME analysis [17] is employed to improve the explainability of the ML-based vehicle engine fault diagnosis model. LIME approximates the local behavior of a complex predictive model by generating perturbed samples around a given input and observing the resulting changes in predictions. It then fits a simple and interpretable linear model to these samples to identify the features that most strongly influence the original output. For example, when a model diagnoses vehicle engine faults, LIME helps determine whether the classification was driven by variations in specific attribute values or by recurring patterns within the input data.

Existing studies have predominantly presented the outputs of LIME in visual formats such as bar charts and decision rule paths, as shown in Fig. 2. However, these visualizations require manual interpretation and conversion into textual explanations by the user, which can be both time-consuming and prone to subjective bias. To overcome these limitations, we employ a VLM to automatically transform the results of LIME analysis into structured and coherent textual explanations. This approach clarifies the model’s reasoning for each specific prediction by providing interpretable descriptions of how individual input features contribute to the diagnostic outcome.

In this context, the effectiveness of our method critically depends on the design of prompts used for the VLM [44]. Since the VLM generates responses by interpreting prompts through its pre-trained knowledge, the quality and relevance of the output can vary significantly depending on the structure, wording, and contextual detail of the input prompt. Therefore, the design of prompts with adequate contextual information is essential to ensure accurate, reliable, and informative textual explanations.

Due to the length, the structured prompt designed to guide the VLM in generating textual explanations for the LIME analysis is provided in Appendix A. The prompt is composed of four main components: a role description, a dataset description, a task specification, and an interpretation guideline.

The role description defines the function of the VLM as a domain expert in vehicle engine diagnostics, ML, and XAI. It instructs the model to analyze LIME-based explanations produced by a vehicle engine fault diagnosis model. The dataset description outlines the input and output variables used in the model. It also specifies the normalization method applied to the data and summarizes key statistical properties including mean, standard deviation, and median for each variable. The output variable is a categorical label representing the engine’s operational state, and the description includes the characteristics associated with each class. This information provides essential contextual grounding for the VLM’s interpretive reasoning. The task specification introduces the LIME visualization and presents its components, including prediction probabilities, decision rules, and local feature contributions. Lastly, the interpretation guideline provides detailed instructions for how the VLM should analyze the LIME output. It requests a structured explanation that addresses the relative influence of each variable, the prioritization of variables, the interpretation of variables with minimal contributions, and the overall decision-making logic.

In the prompt design, several key elements are presented in bold black fonts to emphasize critical components. These include the statistical summaries of the input data, the employed prediction model, the predicted class, the value of the highest prediction probability, and the LIME feature conditions with their contribution values. These elements are dynamically populated within the prompt according to the characteristics of each test instance. Moreover, numbered formats are utilized to represent major analysis steps that require sequential reasoning, whereas dash markers are used to denote supporting details or explanatory instructions within each step.

This structure helps the VLM generate context-specific explanations grounded in model reasoning. Building upon this structure, the overall prompt design is formulated to guide the VLM in generating technically accurate and analytically well-founded explanations. This enhances the interpretability of the diagnostic process, supporting its applicability in practical deployment contexts.

Similar to LIME analysis in Section 3.2, SHAP analysis [18] is employed to improve the explainability of the ML-based vehicle engine fault diagnosis model. Based on cooperative game theory, SHAP assigns each feature a Shapley value that quantifies its individual contribution to a specific prediction, thereby providing both local and global perspectives on the model’s decision-making process. For instance, in the context of vehicle engine fault diagnosis, SHAP can highlight which sensor readings or operational parameters strongly drive the model’s diagnostic outcome.

A widely used visualization for this purpose is SHAP waterfall plot, as shown in Fig. 3. This plot illustrates how the model’s output gradually evolves from a baseline value by sequentially adding the positive and negative contributions of individual features until the final output is reached. It provides an intuitive decomposition of feature-level effects for a given instance, clearly indicating which variables increase or decrease the model’s predicted outcome. However, similar to other SHAP visualizations, it requires manual interpretation, which may introduce subjective bias. To address this limitation, we employ a VLM to automatically transform SHAP waterfall plots into structured and coherent textual explanations, thereby improving interpretability and enhancing accessibility of the diagnostic reasoning process.

As with LIME, the effectiveness of employing a VLM to interpret SHAP outputs largely depends on the design of prompts. Since the VLM generates textual explanations based on its pre-trained knowledge, variations in the phrasing, structure, and contextual completeness of prompts can influence the clarity, faithfulness, and reliability of the generated output. Accordingly, providing well-structured prompts enriched with sufficient contextual information is essential to ensure accurate, consistent, and meaningful interpretations of SHAP results.

The structured prompt for VLM-based SHAP analysis was adapted from the corresponding prompt designed for LIME analysis, with targeted modifications to reflect the characteristics of SHAP interpretation. The complete prompt is provided in Appendix A due to its length. Similar to the prompt for LIME analysis, it is composed of four main components: a role description, a dataset description, a task specification, and an interpretation guideline.

The role description was retained with minor terminology updates to reference SHAP, and the dataset description remained unchanged since both tasks rely on the same input and output variables. The task specification was revised to capture the distinctive characteristics of SHAP waterfall plots, highlighting the stepwise decomposition of the prediction from the baseline value to the final output through sequential positive and negative feature contributions, along with their magnitudes and color-coded representations. Finally, the interpretation guideline instructs the VLM to generate a structured and rigorous explanation by identifying key positive and negative features, describing the cumulative contribution process, discussing low-impact variables, and analyzing the decision logic underlying the model’s prediction.

This design ensures that the generated explanations remain locally focused and resistant to overgeneralization, thereby producing clear, reliable, and contextually grounded explanations suitable for diagnostic reporting. In doing so, the SHAP prompt complements the LIME-based approach, providing a consistent yet distinct framework for generating VLM-driven explanations. Together, these prompt structures establish a robust foundation for integrating multiple XAI methods into a unified interpretability pipeline.

Although VLM-based analyses of LIME and SHAP individually provide valuable textual explanations, their outputs often highlight different aspects of the model’s decision-making process. LIME focuses on local surrogate approximations derived from feature perturbations, while SHAP provides game-theoretic attributions that quantify cumulative feature contributions. As a result, relying on a single method may constrain interpretability and introduce bias. To overcome this limitation, we employ an LLM to synthesize the complementary insights obtained from both VLM-based LIME and SHAP analyses into a unified and coherent diagnostic report. This integration enables users to leverage the strengths of each explanation method while mitigating their respective limitations.

To achieve this objective, we designed a structured prompt for unified report generation. The complete prompt is provided in Appendix A due to its length. The prompt follows the same four-part structure as those used in the VLM-based analysis stages: a role description, a dataset description, a task specification, and an interpretation guideline.

The role description defines the objective of the LLM as generating a consolidated diagnostic report that integrates VLM-based LIME and SHAP explanations. The dataset description provides context regarding the engine behavior and emission-related variables, their normalization, and the diagnostic output classes. This ensures that the LLM remains grounded in the domain-specific characteristics of the data. The task specification instructs the LLM to merge the two explanations into a single coherent narrative by consolidating overlapping evidence, reconciling differences, and preserving technical rigor. This step encourages the model to produce a holistic interpretation of each prediction instance rather than a simple presentation of LIME and SHAP outputs. Lastly, the interpretation guideline describes the structure of the final report. It requires the LLM to summarize the predicted fault class, highlight consistent feature contributions, incorporate method-specific insights, and present an integrated decision logic. The guideline further constrains the explanation to local interpretation, thereby ensuring clarity and reliability.

By explicitly encoding these elements, the unified report generation prompt ensures that the resulting explanations are not only coherent and comprehensive but also precise and academically rigorous, thereby making them well-suited for diagnostic reporting.

To validate the effectiveness of the proposed method, we utilize the publicly available EngineFaultDB dataset [45]. EngineFaultDB is a comprehensive dataset designed for vehicle engine fault classification, comprising 55,999 entries across 14 distinct variables. The dataset was constructed using real-world data collected from the C14NE spark ignition engine [46] during vehicle acceleration. It contains measurements from both normal operations and multiple fault scenarios.

Table 1 presents the input variables of the EngineFaultDB dataset. These input variables are designed to capture both the dynamic behavior of the engine during vehicle acceleration and its corresponding emission characteristics. The dataset includes key engine parameters including manifold absolute pressure (MAP), throttle position sensor (TPS), force, power, revolutions per minute (RPM), fuel consumption (measured in l/h and l/100 km), and speed. In addition, it incorporates exhaust gas emission variables, including carbon monoxide (CO), hydrocarbons (HC), carbon dioxide (CO₂), oxygen (O₂), lambda, and air-fuel ratio (AFR). These variables were measured using the non-dispersive gas analyzer (NGA) 6000 to ensure high precision data acquisition. All input variables are represented as continuous numerical values. Table 2 further summarizes the distribution of these variables by reporting their mean, standard deviation and quartiles, thereby providing a comprehensive statistical context for the subsequent analysis.

The proposed method classifies vehicle engine fault into 4 categories: no fault, rich mixture, lean mixture, and low voltage, as detailed in Table 3. The no fault class represents standard operating conditions without any detectable anomalies. The rich mixture class corresponds to an excessive fuel-to-air ratio, leading to increased carbon monoxide emissions due to incomplete combustion. The lean mixture class indicates insufficient fuel supply, which results in elevated oxygen levels and unstable engine performance. Lastly, the low voltage class is related to electrical irregularities in the ignition or fuel systems, causing reduced combustion efficiency and abnormal sensor readings. The dataset comprises 16,000 normal samples, 10,998 rich mixture samples, 15,000 lean mixture samples, and 14,001 low voltage samples. It provides a diverse and representative foundation for evaluating the performance of the proposed vehicle engine fault diagnosis method.

The performance of the proposed method was evaluated through comparative experiments with multiple ML-based classification models. All input variables were normalized to the range [0, 1] using min-max normalization [47]. The dataset was partitioned into training, validation, and test sets in proportions of 70%, 10%, and 20%, respectively. A total of 13 ML models were employed for the experiments, including logistic regression [48], k-nearest neighbor [49], multi-layer perceptron (MLP) classifier [50], random forest [51], decision tree [52], extratree classifier [53], AdaBoost [54], linear discriminant analysis [55], quadratic discriminant analysis [56], gradient boosting [57], LightGBM [58], CatBoost [59], and XGBoost [30].

In Sections 3.2–3.4, this study employed OpenAI GPT-4o [60]. GPT-4o is a multimodal LLM that can process and generate outputs across text and image inputs within a unified architecture. It has demonstrated high performance in a wide range of tasks, including natural language understanding, complex reasoning, knowledge retrieval, and cross-modal information processing. The model has been successfully applied to various domains, including scientific research and real-time human-computer interaction. In this study, GPT-4o was utilized to generate interpretable textual explanations and to support diagnostic reasoning in vehicle engine fault classification.

To quantitatively evaluate the performance of the vehicle engine fault diagnosis model, we employed 4 widely used classification metrics: precision, recall, f1-score, and accuracy [61]. Precision quantifies the proportion of true positive predictions among all positive predictions made by the model. Recall measures the ability of the model to correctly identify positive cases from all actual positives. The F1 score, which is the harmonic mean of precision and recall, provides a balanced evaluation metric when both precision and recall are critical. Accuracy represents the overall ratio of correctly classified instances to the total number of samples, serving as a general measure of classification performance. The mathematical definitions of these metrics are presented in Eqs. (1)–(4). (1)Precision=TruePositiveTruePositive+FalsePositive (2)Recall=TruePositiveTruePositive+FalseNegative (3)F1−score=2×Precision×RecallPrecision+Recall (4)Accuracy=TruePositive+TrueNegativeTruePositive+FalsePositive+FalseNegative+TrueNegative

To evaluate the quality of the textual explanations generated in Sections 3.2–3.4, we adopted a dual evaluation framework that combines VLM-as-a-Judge [62] and human expert assessment [21]. The VLM-as-a-Judge employs a VLM to automatically assess the quality of generated explanations according to predefined criteria. In this study, OpenAI GPT-4o was used as the evaluation model due to its strong multimodal reasoning capabilities, while LLaMA 3.2-Vision was additionally employed as an independent evaluator to mitigate potential model-specific bias [63]. By aggregating evaluation outcomes across these two VLMs, we aim to enhance the robustness of the automated assessment. This automated approach ensures scalability and reproducibility, whereas human expert assessment provides domain-grounded insights that capture practical interpretability in real diagnostic contexts. By employing both evaluation methods under the same conditions, we leverage the complementary strengths of automation and expert judgment, thereby improving the robustness and reliability of the overall evaluation.

Both evaluations were conducted using a 5-point Likert scale [64] to ensure comparability, where 1 indicates strongly disagree and 5 indicates strongly agree. The evaluation criteria employed in this study are summarized in Table 4. The prompt for the VLM-as-a-Judge is presented in Table A4. The comprehensibility assesses whether the explanation is expressed in a clear and easily interpretable manner, while the faithfulness evaluates the degree to which the explanation accurately reflects the underlying visualization and its diagnostic meaning. The diagnostic relevance examines the alignment of the explanation with established knowledge in vehicle engine fault diagnosis, and the explanatory value evaluates the usefulness of the explanation in enhancing user understanding beyond visual analysis. Finally, the reliability captures the perceived credibility and trustworthiness of the explanation from the perspective of domain experts. Collectively, these criteria enable a comprehensive evaluation of both the technical fidelity and the practical interpretability of the generated reports, ensuring a balanced assessment of explanation quality.

Based on the experimental settings described in Section 4.2, we evaluated the diagnostic performance of 13 ML models for vehicle engine fault diagnosis. Table 5 presents the results in terms of precision, recall, f1-score, and accuracy. For each metric, the best performing model is highlighted in bold, and the second-best model is underlined.

Among all evaluated models, the XGBoost-based fault diagnosis model achieved the highest performance across all evaluation metrics, with an overall accuracy of 75.55%. LightGBM, another gradient boosting-based ensemble model, ranked second with comparably strong results. These findings highlight the effectiveness of boosting-based ensemble methods in capturing complex, non-linear patterns inherent in vehicle engine fault data. In contrast, traditional statistical models such as logistic regression and linear discriminant analysis exhibited substantially lower performance, reflecting their limited capacity to model high-dimensional and non-linear sensor relationships.

It is worth noting that the EngineFaultDB dataset represents a challenging four-class classification problem involving noisy sensor measurements and overlapping fault characteristics. Under such realistic conditions, moderate accuracy levels are commonly observed, and perfect classification performance is difficult to achieve. In this context, the proposed method does not aim to provide deterministic or definitive fault judgments. Instead, it focuses on enhancing transparency by explaining how the diagnostic model arrives at a given prediction.

Overall, these results indicate that modern ensemble models, particularly XGBoost, provide robust diagnostic performance, while the interpretability method proposed in this study complements the classifier by exposing its reasoning process. This combination supports informed decision-making by domain experts, especially in scenarios where model predictions may involve inherent uncertainty.

We evaluated the effectiveness of a VLM in generating textual explanations for LIME analysis results obtained from the vehicle engine fault diagnosis model. Fig. 4 presents the LIME visualization of the XGBoost classifier, and Table 6 provides the corresponding VLM-based textual interpretation. In this instance, the model classified the instance as a lean mixture fault with a prediction probability of 84%.

The explanation identified extremely low CO emissions as the most influential variable in classifying the instance as a lean mixture condition. This observation is consistent with established principles in combustion theory, where excess oxygen leads to reduced CO output. The second most influential factor was fuel consumption l/100 km, supporting the model’s interpretation of reduced fuel injection, which is characteristic of lean operation. Additional contributing features included elevated vehicle speed and HC, which the VLM interpreted as compensatory responses to limited fuel delivery.

The analysis identified features with marginal contribution, such as power output and RPM, which served as supplementary indicators of engine activity but were not decisive for the lean mixture classification. It also noted the absence of variables such as MAP, TPS, Lambda, and AFR, whose values likely remained within normal operating ranges and thus offered limited discriminatory power in this instance.

The overall explanation revealed a coherent inference of the model’s decision-making process. It demonstrated that the classification was based on the convergence of low emissions, high mechanical output, and low fuel consumption. The VLM effectively captured this logic through an interpretable narrative, clarifying how each input variable contributed to the final prediction. This supports the feasibility of combining LIME with VLM-based interpretation to improve transparency in vehicle engine fault diagnosis and to facilitate the communication of model reasoning in a user-accessible format. Additional results of the VLM-based LIME analysis can be found in the Supplementary Materials.

We further employed a VLM to generate textual explanations for the SHAP waterfall plot results of the vehicle engine fault diagnosis model. Fig. 5 illustrates the SHAP waterfall plot corresponding to the same instance analyzed in Section 5.2, and Table 7 presents the VLM-based interpretation of this result.

The explanation highlighted abnormal emission indicators as the dominant factors influencing the classification. In particular, CO emerged as the most decisive feature, reflecting a combustion imbalance characteristic of lean operation. Supporting contributions were observed from elevated engine revolutions, increased fuel consumption per hour, and moderate variations in CO₂ and HC levels, all reinforcing the inference of excess air relative to fuel injection. Vehicle speed was further identified as an operational variable that strengthened the model’s confidence in diagnosing a lean mixture under load conditions.

At the same time, the VLM identified counter-signals such as torque and oxygen concentration, which partially mitigated the lean diagnosis but were ultimately insufficient to outweigh the stronger emission- and operation-driven evidence. Variables including MAP, AFR, and Lambda were characterized as having negligible impact, suggesting that their values remained within normal operating ranges for this instance.

Overall, the SHAP-based explanation revealed a stepwise contribution process in which dominant emission-related features progressively elevated the prediction toward the lean mixture class, while weaker counteracting signals exerted only limited influence. The VLM effectively captured this reasoning in narrative form, demonstrating the feasibility of converting SHAP attributions into coherent, human-interpretable explanations. This approach complements the LIME-based analysis by providing a magnitude-oriented perspective, thereby enhancing the transparency and trustworthiness of the diagnostic model. Additional results of the VLM-based SHAP analysis can be found in the Supplementary Materials.

To demonstrate the feasibility of integrating multiple interpretability methods, we employed an LLM as a report generator to consolidate VLM-based explanations derived from both LIME and SHAP analyses. While Sections 5.2 and 5.3 presented instance-level narratives independently generated from LIME and SHAP visualizations, this section highlights how the LLM synthesizes these complementary perspectives into a single unified diagnostic report.

As shown in Table 8, the generated report first confirmed a consistent classification outcome, identifying the instance as a lean mixture condition with high confidence across both interpretability methods. The LLM subsequently aligned overlapping evidence, emphasizing that emission-related features, particularly low carbon monoxide, moderate CO₂, and slightly elevated HC, served as strong and coherent indicators of lean combustion. In parallel, operational metrics such as high RPM and vehicle speed, combined with the dual fuel consumption pattern, low distance-normalized and high hourly consumption, were identified as mutually reinforcing indicators.

The LLM also reconciled method-specific differences between LIME and SHAP. LIME emphasized interpretable threshold-based rules, such as ranges of CO and fuel consumption, while SHAP quantified additive feature contributions in a stepwise manner. In cases of apparent discrepancies, such as the interpretation of CO directionality, the LLM clarified that both methods ultimately supported the conclusion that lower than typical CO values increased the probability of lean operation. Moreover, the report distinguished between features omitted in LIME due to near-normal values and those assigned minimal weights by SHAP, thereby ensuring consistent representation of low impact evidence across methods.

Through this synthesis, the LLM generated an integrated decision logic that contextualized the prediction as the result of convergent signals from emissions, operational load, and fuel efficiency, moderated by weak counter-signals such as torque and oxygen levels. The resulting explanation formed a coherent narrative that remained diagnostically faithful to the model’s reasoning process while enhancing accessibility for human interpretation.

This experiment illustrates the potential of LLM-based report generators to serve as interpreters that unify heterogeneous explanation modalities. By combining the threshold-oriented clarity of LIME with the magnitude-based decomposition of SHAP, the method enhances transparency, mitigates the risk of conflicting interpretations, and provides domain experts with a more robust and interpretable account of ML–driven fault diagnosis.

To assess the explanatory quality and user satisfaction of the proposed method, we employed a dual evaluation method that combined VLM-as-a-Judge-based assessment with a user study incorporating quantitative ratings and qualitative feedback. The evaluation employed a five-point Likert scale survey administered to 10 domain experts with professional experience in vehicle engine systems. The study compared three explanation settings: VLM-based LIME analysis, VLM-based SHAP analysis, and the proposed method. The evaluation aimed to assess five key criteria, including comprehensibility, faithfulness, domain relevance, explanatory value, and reliability. In addition, open-ended responses were collected to identify strengths and areas for improvement.

As shown in Tables 9 and 10, the VLM-as-a-Judge evaluation results indicate that the proposed method achieved the highest scores across all evaluation criteria under both GPT-4o and LLaMA 3.2-Vision. In particular, improvements in comprehensibility and faithfulness highlighted the effectiveness of synthesizing LIME- and SHAP-based explanations into a unified narrative that enhances clarity and alignment with the model’s diagnostic reasoning. Although LLaMA 3.2-Vision produced slightly more conservative absolute scores than GPT-4o, the relative performance trends remained consistent. In both evaluations, the proposed method outperformed standalone LIME and SHAP analyses, especially in domain relevance and explanatory value, suggesting that the observed gains are robust to evaluator-specific bias. Reliability scores remained lower than other criteria, reflecting the inherent challenge of establishing trust in AI-driven diagnostics. Nevertheless, the proposed method demonstrated consistent improvements in reliability over individual explanation approaches. Overall, the agreement between two independent VLM evaluators provided strong evidence for the robustness and generalizability of the proposed method.

Table 11 presents the results of the user study. Although the absolute scores were lower than those obtained in the VLM-as-a-Judge evaluation, the relative trends remained consistent. The proposed method again outperformed both standalone LIME and SHAP analyses across all evaluation criteria. Participants particularly appreciated improvements in faithfulness and explanatory value, underscoring the effectiveness of the unified report in conveying diagnostic reasoning. Nevertheless, the comparatively lower scores for reliability across all methods highlight the persistent challenge of fostering user confidence in automated explanations within safety-critical diagnostic environments.

Table 12 summarizes the open-ended feedback. Positive remarks highlighted the accessibility of the unified report, its clear and structured format, and the sense of trust fostered by presenting consistent evidence across methods. Participants noted that the integration of visual indicators and textual reasoning enhanced decision-making and communication in practical diagnostic contexts. On the other hand, suggestions for improvement emphasized the need for more domain-specific diagnostic guidance (e.g., tailored recommendations for sensor replacement or fuel system checks), improved handling of discrepancies between LIME and SHAP outputs, and options to adjust report length and complexity for different user groups. Participants also recommended emphasizing threshold deviations more explicitly to enhance interpretability for less experienced users.

Taken together, these findings confirm that the proposed method delivers superior explanatory performance by enhancing clarity, consistency, and user relevance compared to single-method explanations. At the same time, the qualitative feedback highlights directions for refinement, particularly in domain customization, flexible presentation, and reliability enhancement. Addressing these areas is expected to further improve the usability and practical adoption of the proposed method in real-world diagnostic scenarios.

Although LIME and SHAP are jointly integrated into a unified diagnostic report, differences in their attribution mechanisms may lead to variations in feature importance and ranking [65]. To move beyond purely descriptive reconciliation and to enhance cross-XAI reliability, we conducted a quantitative consistency and discrepancy analysis between LIME- and SHAP-based local explanations using an LLM.

As summarized in Table A5, a structured analysis prompt was designed to objectively compare feature-level attributions produced by LIME and SHAP for the same diagnostic instance. Overlapping features were first identified, and their importance values were independently normalized to account for scale differences between LIME coefficients and SHAP values. Based on these normalized representations, four quantitative metrics were computed: feature overlap ratio [66], Spearman rank correlation [67], directional consistency score [68], and normalized importance divergence [69].

Table 13 presents the analysis results for the same lean mixture instance previously analyzed in Table 8. Strong structural agreement was observed, with nine overlapping features out of eleven unique features, yielding a feature overlap ratio of 0.818. Both methods consistently assigned the highest relative importance to CO, confirming agreement on the dominant role of emission-related indicators. Other shared features, including fuel consumption metrics, vehicle speed, RPM, and exhaust gas variables, further reinforced this common explanatory basis.

Directional consistency analysis revealed perfect semantic alignment, with a directional consistency score of 1.00, indicating that all overlapping features contributed in the same direction to the predicted class. In contrast, rank- and magnitude-based metrics showed moderate divergence, with a Spearman rank correlation of 0.48 and a normalized importance divergence of 0.236. These differences reflect methodological distinctions, LIME’s local sensitivity vs. SHAP’s additive marginal contributions, rather than contradictory diagnostic reasoning.

Overall, the results demonstrate strong qualitative consistency and moderate quantitative alignment between LIME and SHAP. The two methods should therefore be interpreted as complementary, and explicitly quantifying their agreement and divergence strengthens the reliability of the proposed method.

Recent studies have shown that both VLMs and LLMs are susceptible to prompt sensitivity and may occasionally produce misleading or hallucinated outputs [70,71]. To address this concern, we conducted an explicit hallucination analysis on the generated explanations and introduced an additional verification mechanism to impose domain-specific constraints on the explanation process.

First, we examined all VLM and LLM-generated explanations used in the experiments to identify potential hallucinations, defined as statements that are inconsistent with the underlying XAI visualizations, model outputs, or domain knowledge of vehicle engine fault diagnosis. This examination was conducted through manual and qualitative inspection of the evaluated instances, in which the generalized textual explanations were systematically compared against the corresponding LIME or SHAP visualizations and diagnostic evidence.

Within the scope of the instances analyzed in this study, no hallucinated explanations were observed. In particular, all referenced features, contribution directions, and diagnostic interpretations were grounded in the corresponding XAI results and aligned with well-known combustion and engine operation principles. These observations indicate that the structured prompt design adopted in this study effectively constrained the generative behavior of the models under the evaluated conditions.

Nevertheless, to further mitigate the risk of erroneous explanations, we introduce an additional verification stage based on a domain-aware review agent [72]. The prompt of this agent is presented in Table A6. Specifically, explanations generated by the VLMs for LIME and SHAP interpretation and the unified reports produced by the LLMs are subsequently re-evaluated by an independent verification agent. This agent assesses factual consistency with the original XAI outputs, checks adherence to domain-specific constraints, and flags unsupported claims or speculative reasoning.

By incorporating this verification stage, the proposed method enforces an explicit separation between explanation generation and explanation validation. This design reduces reliance on a single generative pass and enhances robustness against prompt sensitivity and latent hallucination risks. Although the present study did not observe hallucinated outputs, the verification mechanism provides an additional safeguard and establishes a scalable foundation for deploying the proposed method in safety-critical and energy-related diagnostic systems.

While the proposed DRIVE focuses on fault classification and explainable report generation, its outputs are directly applicable to real-world automotive maintenance and troubleshooting workflows [3,4]. To illustrate this practical relevance, we discuss how the unified diagnostic report supports technician decision-making using a representative lean mixture case.

As shown in Table 8, the unified report consolidates LIME and SHAP-based explanations into a structured and instance-level diagnostic narrative that highlights dominant signals, operating context, and counter evidence. From a maintenance perspective, this report bridges model inference and component-level inspection. For example, the consistent identification of low CO levels combined with moderate CO₂ and elevated HC emissions directs technicians toward air-dominant combustion issues, guiding inspection priorities such as intake air leaks, airflow sensor calibration, or fuel delivery constraints.

The report further contextualizes the fault under dynamic operating conditions by linking elevated RPM and vehicle speed with a dual fuel-consumption pattern. This helps technicians distinguish transient driving effects from persistent system-level faults, thereby narrowing troubleshooting pathways and reducing unnecessary component replacement. Explicit identification of counter signals, such as torque and oxygen levels, also supports ruling out competing fault hypotheses.

Overall, the stepwise and structured nature of the unified report aligns well with automotive diagnostic workflows, reducing the cognitive burden associated with interpreting raw XAI visualizations. Although an end-to-end deployment is beyond the scope of this study, the presented case demonstrates that DRIVE can function as an effective decision-support layer in practical automotive maintenance pipelines.

Real-world vehicle diagnostics often impose constraints on inference latency and computational resources, particularly in scenarios requiring near-real-time decision support [73,74]. It should be noted that the primary objective of DRIVE is not real-time engine control, but interpretable diagnostic reasoning and decision support for maintenance, troubleshooting, and post-event analysis.

From a computational standpoint, DRIVE follows a modular architecture with distinct latency characteristics. The core fault classification is performed using XGBoost, whose inference time is on the order of milliseconds under embedded hardware assumptions and is well suited for deployment on edge devices or in-vehicle diagnostic units. This enables timely detection of potential fault states without reliance on external computation.

In contrast, the explainability and report generation stages, including LIME and SHAP analysis, VLM-based interpretation, and LLM-based synthesis, incur higher computational overhead. These components are intended for near-real-time or offline diagnostic contexts, such as fleet maintenance systems and engineering analysis platforms, where response times of several seconds are acceptable and interpretability is prioritized over strict latency constraints.

The modular design of DRIVE allows flexible deployment strategies. For example, a hierarchical configuration can be adopted in which fault classification runs on the vehicle or edge device, while XAI computation and VLM–LLM reasoning are executed on a gateway, on-premise server, or cloud infrastructure [75,76]. Future work will focus on latency-aware optimization, lightweight explanation generation, and partial on-device deployment to further improve engineering feasibility in energy and safety-critical automotive systems [77,78].

The experimental evaluation in this study focused on LIME and SHAP as representative instance-level XAI techniques, as the primary objective of DRIVE is to enhance local interpretability by translating instance-specific visual explanations into coherent textual reasoning and unified diagnostic reports. Nevertheless, other widely used XAI methods, such as partial dependence plots (PDP) [19] and individual conditional expectation (ICE) [79] plots, offer complementary explanatory perspectives.

PDP and ICE are effective in capturing global and semi-local relationships between input variables and model predictions. PDP illustrates the average marginal effect of a feature across the dataset, while ICE reveals instance-level heterogeneity in model responses. Fig. 6 presents representative PDP and ICE visualizations for key engine-related variables, demonstrating how changes in sensor values influence predicted fault probabilities.

Although PDP and ICE were not integrated into DRIVE, these results indicate that the proposed method is not limited to LIME and SHAP. In principle, PDP and ICE visualizations can be interpreted using the same VLM-based translation mechanism and subsequently synthesized by the LLM together with LIME- and SHAP-based explanations. Incorporating additional XAI techniques, including PDP, ICE, and counterfactual explanations, therefore represents a natural extension of DRIVE and will be explored in future work to further enhance interpretability and practical decision support in vehicle engine fault diagnosis and energy-related systems [33].

In the current implementation, GPT-4o was employed as the VLM to interpret LIME and SHAP visualizations and generate explanatory reports, owing to its strong multimodal reasoning capability. Nevertheless, reliance on a single VLM raises questions regarding the generalizability of the proposed method when alternative models are used.

Importantly, DRIVE is designed to be VLM-agnostic. The VLM is strictly constrained to translating structured XAI visualizations into textual explanations under predefined prompts and domain-specific guidelines. Consequently, the core diagnostic evidence derived from the ML model and XAI methods remains unchanged regardless of the selected VLM. Replacing GPT-4o with other VLMs, such as LLaMA 3.2-Vision [63], is therefore expected to mainly affect linguistic style and conservativeness, rather than the underlying explanatory content or diagnostic conclusions [80].

Auxiliary evaluations with alternative VLMs indicate that, while absolute Likert-scale scores may vary slightly, the relative performance trends and qualitative interpretations remain consistent under identical prompt structures. This suggests that the explanatory effectiveness of DRIVE is primarily driven by structured XAI integration and prompt design rather than by VLM-specific characteristics. Future work will include systematic benchmarking of multiple VLMs to further assess robustness, faithfulness, and hallucination resistance in safety-critical diagnostic settings [81].