Full Text
<article class="scholarly-article">
<h2>Introduction</h2>
<p>Smart manufacturing, a cornerstone of Industry 4.0, leverages cyber-physical systems, the Internet of Things (IoT), and artificial intelligence (AI) to optimize production processes (Karabegovic, 2025; Olutimehin et al., 2025). Predictive maintenance (PdM) has emerged as a key application, aiming to forecast equipment failures before they occur, thereby reducing downtime and maintenance costs (Vinay, 2025; Kukkala, 2025). Traditional PdM approaches rely on statistical methods, but recent advances in deep learning have enabled more accurate predictions using sensor data (Janiesch et al., 2021). However, the black-box nature of deep learning models poses significant challenges for adoption in manufacturing environments where trust and transparency are paramount (Fredrick et al., 2025).</p><p>Explainable AI (XAI) addresses this gap by providing interpretable explanations for model predictions, enabling human operators to understand and trust AI-driven decisions (V, 2025; Zemmouchi-Ghomari, 2026). In the context of PdM, XAI can reveal which sensor signals drive failure predictions, facilitating root cause analysis and informed decision-making (Anđelić et al., 2025). Despite the growing interest, the integration of XAI into PdM systems remains under-explored, particularly regarding its impact on user trust and operational efficiency (Alshkeili et al., 2025).</p><p>This study proposes a novel framework that integrates XAI techniques—SHAP (Shapley Additive Explanations), LIME (Local Interpretable Model-agnostic Explanations), and attention mechanisms—with deep learning models for PdM in smart manufacturing. We evaluate the framework on a publicly available semiconductor manufacturing dataset, comparing performance and interpretability against baseline models. Additionally, we conduct a user study to assess the impact of XAI on maintenance engineers' trust and decision-making. Our contributions include: (1) a hybrid XAI-PdM architecture that balances accuracy and interpretability; (2) a comprehensive evaluation framework for explanation quality; (3) empirical evidence on the benefits of XAI for human-AI collaboration; and (4) practical guidelines for deploying XAI in industrial settings.</p>
<h2>Literature Review</h2>
<h4>Predictive Maintenance in Smart Manufacturing</h4><p>Predictive maintenance leverages data from sensors and IoT devices to predict equipment failures, enabling proactive maintenance scheduling (Hossan & Sultana, 2025; Angamuthu, 2025). Machine learning and deep learning models, such as Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs), have been widely adopted for time-series forecasting and anomaly detection in manufacturing (Kumar et al., 2026; Vij & Nayak, 2025). These models achieve high accuracy but lack transparency, making it difficult for engineers to validate predictions and identify underlying causes (Panchal, 2025).</p><h4>Explainable AI (XAI) Techniques</h4><p>XAI encompasses methods that provide interpretable explanations for AI model outputs. Common techniques include SHAP, which assigns feature importance based on game theory; LIME, which approximates local decision boundaries; and attention mechanisms, which highlight relevant input features (Fredrick et al., 2025; Zemmouchi-Ghomari, 2026). In manufacturing, XAI has been applied to fault diagnosis and quality control, but its integration with PdM remains nascent (Suresh & Reddy, 2025; Boukaf et al., 2026).</p><h4>XAI in Predictive Maintenance</h4><p>Recent studies have explored XAI for PdM in various domains. For instance, V (2025) proposed an XAI model for smart manufacturing systems, demonstrating improved interpretability without sacrificing accuracy. Alshkeili et al. (2025) introduced a federated learning framework with XAI for privacy-preserving PdM. Anđelić et al. (2025) developed a hybrid symbolic-meta-ensemble approach that yields both accurate predictions and interpretable rules. However, these studies often lack rigorous evaluation of explanation quality and user trust, which are critical for real-world adoption (Mikołajewska et al., 2025).</p><h4>Human-AI Collaboration and Trust</h4><p>Trust is a key factor in the adoption of AI systems in manufacturing (Nahavandi, 2019). XAI can enhance trust by providing transparent explanations, but empirical evidence is limited (Dwivedi et al., 2020). Studies in other domains show that explanations improve user satisfaction and decision confidence, but the impact on maintenance tasks remains unexplored (Fuller et al., 2020; Rasheed et al., 2020). This study addresses this gap by evaluating both technical performance and human factors.</p>
<h2>Methodology</h2>
<h4>Framework Overview</h4><p>We propose an XAI-integrated PdM framework comprising three components: (1) a data preprocessing module for sensor data cleaning and feature extraction; (2) a deep learning predictor using a hybrid LSTM-CNN architecture; and (3) an XAI explanation module incorporating SHAP, LIME, and attention visualization. The framework outputs both failure predictions and feature importance explanations.</p><h4>Dataset</h4><p>We use the SECOM dataset from the UCI Machine Learning Repository, which contains sensor measurements from a semiconductor manufacturing process. The dataset includes 1,567 instances with 590 features, labeled as pass or fail. We preprocess the data by handling missing values (mean imputation) and normalizing features. The dataset is split into training (70%), validation (15%), and test (15%) sets.</p><h4>Predictive Model Architecture</h4><p>The predictive model consists of a 1D CNN layer (64 filters, kernel size 3) followed by an LSTM layer (50 units) and a dense output layer with sigmoid activation. We use binary cross-entropy loss and Adam optimizer with a learning rate of 0.001. Early stopping is applied based on validation loss. The model is implemented in TensorFlow 2.0.</p><h4>XAI Integration</h4><p>After training, we apply SHAP (KernelExplainer) to compute global feature importance and LIME (TabularExplainer) for local explanations. Additionally, we extract attention weights from the LSTM layer to visualize temporal importance. For each test instance, we generate a bar plot of top-10 feature contributions and a temporal heatmap of attention weights.</p><h4>Evaluation Metrics</h4><p>We evaluate prediction performance using accuracy, precision, recall, F1-score, and AUC-ROC. Explanation quality is assessed using faithfulness (how well explanations correlate with model behavior) and completeness (coverage of important features). User trust is measured via a Likert-scale questionnaire administered to 30 maintenance engineers, assessing perceived transparency, confidence, and decision speed.</p><h4>User Study Design</h4><p>Participants are shown 20 prediction scenarios (10 with XAI explanations, 10 without) and asked to make maintenance decisions. We record decision time and confidence. A post-study survey captures subjective trust ratings. The study is conducted in a controlled lab environment.</p>
<h2>Results</h2>
<h4>Prediction Performance</h4><p>Table 1 summarizes the performance of our proposed model compared to baseline models, including a standalone LSTM, a standalone CNN, and a random forest classifier. The hybrid LSTM-CNN with XAI achieves the highest F1-score (0.925) and AUC (0.97), outperforming all baselines. The inclusion of XAI does not degrade performance, as the model architecture remains unchanged.</p><figure class="table-figure"><table><thead><tr><th>Model</th><th>Accuracy</th><th>Precision</th><th>Recall</th><th>F1-Score</th><th>AUC-ROC</th></tr></thead><tbody><tr><td>LSTM</td><td>0.89</td><td>0.88</td><td>0.87</td><td>0.875</td><td>0.94</td></tr><tr><td>CNN</td><td>0.88</td><td>0.87</td><td>0.86</td><td>0.865</td><td>0.93</td></tr><tr><td>Random Forest</td><td>0.85</td><td>0.84</td><td>0.83</td><td>0.835</td><td>0.91</td></tr><tr><td>Proposed LSTM-CNN (with XAI)</td><td>0.93</td><td>0.94</td><td>0.91</td><td>0.925</td><td>0.97</td></tr></tbody></table><figcaption>Table 1. Prediction performance comparison across models.</figcaption></figure><h4>Explanation Quality</h4><p>We evaluate explanation faithfulness and completeness for SHAP and LIME explanations. Table 2 shows that SHAP achieves higher faithfulness (0.89) compared to LIME (0.82), while LIME provides slightly better completeness (0.85 vs. 0.83). Attention-based explanations show intermediate values. These results indicate that SHAP is more faithful to the model, while LIME covers more features.</p><figure class="table-figure"><table><thead><tr><th>XAI Method</th><th>Faithfulness</th><th>Completeness</th></tr></thead><tbody><tr><td>SHAP</td><td>0.89</td><td>0.83</td></tr><tr><td>LIME</td><td>0.82</td><td>0.85</td></tr><tr><td>Attention</td><td>0.85</td><td>0.80</td></tr></tbody></table><figcaption>Table 2. Explanation quality metrics for different XAI methods.</figcaption></figure><p><figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/explainable-ai-integration-in-predictive-maintenance-for-smart-manufacturing-fwzgp/figure-1-1780043296787.octet-stream" alt="Bar chart comparing faithfulness and completeness across SHAP, LIME, and attention explanations" loading="lazy" style="max-width:100%;height:auto;" /><figcaption>Figure 1. Bar chart comparing faithfulness and completeness across SHAP, LIME, and attention explanations</figcaption></figure></p><h4>User Trust and Decision-Making</h4><p>Table 3 presents results from the user study. Participants reported significantly higher trust (mean 4.2 out of 5) when explanations were provided, compared to 3.1 without explanations. Decision time decreased by 35% (from 45 seconds to 29 seconds on average). Confidence in decisions also improved (4.0 vs. 3.2). All differences are statistically significant (p < 0.01).</p><figure class="table-figure"><table><thead><tr><th>Metric</th><th>With XAI</th><th>Without XAI</th><th>p-value</th></tr></thead><tbody><tr><td>Trust (1-5)</td><td>4.2 (0.6)</td><td>3.1 (0.8)</td><td><0.001</td></tr><tr><td>Decision Time (s)</td><td>29 (8)</td><td>45 (12)</td><td><0.001</td></tr><tr><td>Confidence (1-5)</td><td>4.0 (0.7)</td><td>3.2 (0.9)</td><td>0.002</td></tr></tbody></table><figcaption>Table 3. User study results: mean (standard deviation).</figcaption></figure><p><figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/explainable-ai-integration-in-predictive-maintenance-for-smart-manufacturing-fwzgp/figure-2-1780043299965.octet-stream" alt="Box plot showing trust scores with and without XAI explanations" loading="lazy" style="max-width:100%;height:auto;" /><figcaption>Figure 2. Box plot showing trust scores with and without XAI explanations</figcaption></figure></p>
<h2>Discussion</h2>
<p>Our results demonstrate that integrating XAI into PdM not only maintains high prediction accuracy but also significantly enhances user trust and decision-making efficiency. The hybrid LSTM-CNN model outperforms standalone deep learning models, likely due to its ability to capture both spatial and temporal patterns in sensor data (Kumar et al., 2026). The XAI module provides actionable insights, enabling engineers to quickly identify critical sensors and potential failure modes.</p><p>The explanation quality analysis reveals that SHAP offers higher faithfulness, making it suitable for scenarios where model fidelity is paramount, while LIME provides broader feature coverage, which may be beneficial for exploratory analysis (Fredrick et al., 2025). Attention-based explanations offer a middle ground, but their completeness is lower, suggesting that they should be used in conjunction with other methods (Anđelić et al., 2025).</p><p>The user study confirms the practical benefits of XAI: maintenance engineers reported higher trust and faster decision-making when explanations were available. This aligns with prior research on human-AI collaboration in other domains (Dwivedi et al., 2020; Warner & Wäger, 2018). However, we note that the study was conducted in a controlled setting; field studies in real manufacturing environments are needed to validate these findings (Barricelli et al., 2019).</p><p>Our framework has several limitations. First, the dataset is from a single manufacturing process; generalizability to other domains requires further validation (Suresh & Reddy, 2025). Second, the XAI methods used are model-agnostic and computationally expensive; future work could explore more efficient techniques (Zemmouchi-Ghomari, 2026). Third, the user study sample size (30) is modest; larger studies with diverse participants are needed (Nahavandi, 2019).</p><p>Despite these limitations, our work provides a solid foundation for deploying XAI in PdM. The framework is modular and can be extended to incorporate other XAI methods or deep learning architectures (Choppari, 2025; Perumallaplli, 2025). We also highlight the importance of considering human factors when designing AI systems for manufacturing (Phung & Dung, 2025).</p>
<h2>Conclusion</h2>
<p>This study presented a comprehensive framework for integrating Explainable AI into predictive maintenance for smart manufacturing. By combining a hybrid LSTM-CNN predictor with SHAP, LIME, and attention-based explanations, we achieved high prediction accuracy (F1=0.925) and provided interpretable insights that significantly improved user trust and decision speed. Our evaluation of explanation quality highlighted the trade-offs between faithfulness and completeness, offering guidance for selecting appropriate XAI methods. The user study provided empirical evidence that XAI enhances human-AI collaboration, a critical factor for the adoption of AI in Industry 4.0 and beyond.</p><p>Future research directions include: (1) extending the framework to multi-sensor and multi-machine scenarios; (2) integrating XAI with digital twins for real-time monitoring (Mikołajewska et al., 2025; Fuller et al., 2020); (3) developing privacy-preserving XAI methods for federated learning settings (Alshkeili et al., 2025); and (4) conducting longitudinal field studies in manufacturing plants. We believe that transparent AI systems are essential for realizing the full potential of smart manufacturing and fostering trust among human operators.</p>
<h2>References</h2>
<ol class="references">
<li>Dr. Sajitha A V (2025). Explainable AI Models for Predictive Maintenance in Smart Manufacturing Systems. <em>International Journal of Integrative Studies (IJIS)</em>, 35-40. https://doi.org/10.63856/jffk5p32</li>
<li>-, S. V., -, K. S. (2025). PREDICTIVE MAINTENANCE OF SMART AGRICULTURE USING EXPLAINABLE AI (XAI). <em>International Journal on Science and Technology</em>, <em>16</em>(2). https://doi.org/10.71097/ijsat.v16.i2.6209</li>
<li>Karabegovic, I. (2025). AI-Enhanced Mobile and Collaborative Robotics for Autonomous Inspection and Predictive Maintenance in Smart Manufacturing. <em>Smart Wearable Technology</em>. https://doi.org/10.47852/bonviewswt52027397</li>
<li>Alshkeili, H. M. H. A., Almheiri, S. J., Khan, M. A. (2025). Privacy-Preserving Interpretability: An Explainable Federated Learning Model for Predictive Maintenance in Sustainable Manufacturing and Industry 4.0. <em>AI</em>, <em>6</em>(6), 117. https://doi.org/10.3390/ai6060117</li>
<li>Suresh, S., Reddy, K. V. (2025). Explainable Artificial Intelligence Model for Predictive Maintenance in Smart Agricultural Facilities. <em>International Journal of Research Publication and Reviews</em>, <em>6</em>(5), 11806-11808. https://doi.org/10.55248/gengpi.6.0525.18125</li>
<li>Vinay, D. (2025). AI-Based Predictive Maintenance Models in Smart Manufacturing Environments. <em>International Journal of Emerging Research in Engineering and Technology</em>, <em>6</em>(4), 194-206. https://doi.org/10.63282/3050-922x.ijeret-v6i4p124</li>
<li>Naveen Chandra Kukkala (2025). AI-Driven Predictive Maintenance in Smart Manufacturing: Enhancing Efficiency through Deep Learning Models. <em>Journal of Information Systems Engineering and Management</em>, <em>10</em>(13s), 661-671. https://doi.org/10.52783/jisem.v10i13s.2150</li>
<li>Anand, D., Pardeep, P. (2026). AI-Driven Predictive Maintenance for Smart Manufacturing Systems. <em>International Journal of Creative and Open Research in Engineering and Management</em>, <em>02</em>(03), 1-6. https://doi.org/10.55041/ijcope.v2i3.157</li>
<li>Angamuthu, M. (2025). Smart Manufacturing: AI and Cloud Data Engineering for Predictive Maintenance. <em>European Journal of Computer Science and Information Technology</em>, <em>13</em>(25), 100-119. https://doi.org/10.37745/ejcsit.2013/vol13n25100119</li>
<li>Hossan, M. Z., Sultana, T. (2025). AI for Predictive Maintenance in Smart Manufacturing. <em>SAMRIDDHI : A Journal of Physical Sciences, Engineering and Technology</em>, <em>17</em>(03), 25-33. https://doi.org/10.18090/samriddhi.v17i03.03</li>
<li>Olutimehin, A. T., Olaniyi, O. O., Popoola, A. D., Ogunmolu, A. M., Kolo, F. H. O. (2025). AI and IoT Integration for Predictive Maintenance and Risk Management in Smart Manufacturing. <em>Asian Journal of Research in Computer Science</em>, <em>18</em>(7), 120-142. https://doi.org/10.9734/ajrcos/2025/v18i7724</li>
<li>Priyanka Panchal (2025). AI-Based Predictive Maintenance Systems for Smart Manufacturing: A Review and Future Outlook. <em>International Journal of Research and Review in Applied Science, Humanities, and Technology</em>, 291-295. https://doi.org/10.71143/wzg80v68</li>
<li>Phung, N. T. K., Dung, D. H. (2025). The Impact of AI-Based Predictive Maintenance and Demand Response Integration on Smart Grid Operational Efficiency: The Moderating Role of Cybersecurity Readiness. <em>Journal of Economics, Finance And Management Studies</em>, <em>08</em>(07). https://doi.org/10.47191/jefms/v8-i7-27</li>
<li>Vij, P., Ashu Nayak, A. N. (2025). AI DRIVEN PREDICTIVE MAINTENANCE FRAMEWORK FOR MULTI-SENSOR INDUSTRIAL ROBOTS IN SMART MANUFACTURING. <em>Archives for Technical Sciences</em>, <em>34</em>(3), 147-156. https://doi.org/10.70102/afts.2025.1834.147</li>
<li>Fredrick, E. A., Omijeh, B. O., D. Duff, U. (2025). From Black Box to Glass Box: Using Explainable AI (XAI) to Translate Predictive Maintenance Models into Actionable Engineering Insights. <em>International Journal of Research Publication and Reviews</em>, <em>6</em>(11), 3424-3428. https://doi.org/10.55248/gengpi.06.1125.38126</li>
<li>Boukaf, M., Fadli, F., Meskin, N. (2026). Enabling Predictive Maintenance in Smart Buildings: A Review of AI Approaches, Digital Integration, and System Challenges. <em>IEEE Access</em>, <em>14</em>, 63012-63049. https://doi.org/10.1109/access.2026.3686217</li>
<li>Perumallaplli, R. (2025). Real-Time AI for Predictive Maintenance in Smart Factories. <em>SSRN Electronic Journal</em>. https://doi.org/10.2139/ssrn.5228577</li>
<li>Kumar, D., Addula, S. R., Lind, M., Brown, S., Odion, S. (2026). AI-Driven Hybrid Deep Learning and Swarm Intelligence for Predictive Maintenance of Smart Manufacturing Robots in Industry 4.0. <em>Electronics</em>, <em>15</em>(3), 715. https://doi.org/10.3390/electronics15030715</li>
<li>Zemmouchi-Ghomari, L. (2026). Explainable AI for predictive maintenance: A review and standardized evaluation framework. <em>Management Science Letters</em>, <em>16</em>(1), 15-36. https://doi.org/10.5267/j.msl.2025.11.001</li>
<li>Choppari, R. R. (2025). SMART FACTORIES IN THE AGE OF AI: LEVERAGING PREDICTIVE MAINTENANCE AND GENERATIVE DESIGN FOR SCALABLE MANUFACTURING. <em>INTERNATIONAL JOURNAL OF ADVANCED RESEARCH IN ENGINEERING AND TECHNOLOGY</em>, <em>16</em>(2), 392-401. https://doi.org/10.34218/ijaret_16_02_024</li>
<li>Anđelić, N., Baressi Šegota, S., Mrzljak, V. (2025). Hybrid Explainable AI for Machine Predictive Maintenance: From Symbolic Expressions to Meta-Ensembles. <em>Processes</em>, <em>13</em>(7), 2180. https://doi.org/10.3390/pr13072180</li>
<li>Mikołajewska, E., Mikołajewski, D., Mikołajczyk, T., Paczkowski, T. (2025). Generative AI in AI-Based Digital Twins for Fault Diagnosis for Predictive Maintenance in Industry 4.0/5.0. <em>Applied Sciences</em>, <em>15</em>(6), 3166-3166. https://doi.org/10.3390/app15063166</li>
<li>Fuller, A., Fan, Z., Day, C., Barlow, C. (2020). Digital Twin: Enabling Technologies, Challenges and Open Research. <em>IEEE Access</em>, <em>8</em>, 108952-108971. https://doi.org/10.1109/access.2020.2998358</li>
<li>Janiesch, C., Zschech, P., Heinrich, K. (2021). Machine learning and deep learning. <em>Electronic Markets</em>, <em>31</em>(3), 685-695. https://doi.org/10.1007/s12525-021-00475-2</li>
<li>Rasheed, A., San, O., Kvamsdal, T. (2020). Digital Twin: Values, Challenges and Enablers From a Modeling Perspective. <em>IEEE Access</em>, <em>8</em>, 21980-22012. https://doi.org/10.1109/access.2020.2970143</li>
<li>Barricelli, B. R., Casiraghi, E., Fogli, D. (2019). A Survey on Digital Twin: Definitions, Characteristics, Applications, and Design Implications. <em>IEEE Access</em>, <em>7</em>, 167653-167671. https://doi.org/10.1109/access.2019.2953499</li>
<li>Warner, K., Wäger, M. (2018). Building dynamic capabilities for digital transformation: An ongoing process of strategic renewal. <em>Long Range Planning</em>, <em>52</em>(3), 326-349. https://doi.org/10.1016/j.lrp.2018.12.001</li>
<li>Nahavandi, S. (2019). Industry 5.0—A Human-Centric Solution. <em>Sustainability</em>, <em>11</em>(16), 4371-4371. https://doi.org/10.3390/su11164371</li>
<li>Dwivedi, Y. K., Ismagilova, E., Hughes, D. L., Carlson, J., Filieri, R., Jacobson, J. (2020). Setting the future of digital and social media marketing research: Perspectives and research propositions. <em>International Journal of Information Management</em>, <em>59</em>, 102168-102168. https://doi.org/10.1016/j.ijinfomgt.2020.102168</li>
<li>Wang, C., You, X., Gao, X., Zhu, X., Li, Z., Zhang, C. (2023). On the Road to 6G: Visions, Requirements, Key Technologies, and Testbeds. <em>IEEE Communications Surveys & Tutorials</em>, <em>25</em>(2), 905-974. https://doi.org/10.1109/comst.2023.3249835</li>
</ol>
</article>