INTERPRETABLE TRANSFORMER-BASED VIBRATION ANALYSIS FOR ANOMALY DETECTION IN INDUSTRIAL SYSTEMS

Unexpected equipment failures in industrial systems can lead to significant production downtime, increased operational costs, and reduced asset life cycles. Leveraging recent advancements in data science and intelligent decision-support systems, this paper proposes an explainable data-driven optimization and mathematical modeling framework for predictive maintenance scheduling in industrial environments. The framework integrates sensor-based condition monitoring data with machine learning models to forecast potential equipment failures before they occur. A supervised learning approach, implemented using gradient boosting and temporal feature engineering, predicts the remaining useful life (RUL) and failure probability of critical assets. An explainability layer, based on SHAP (SHapley Additive exPlanations) values, provides interpretable insights into the most influential factors contributing to predicted failures, enabling maintenance engineers to validate model outputs and trust automated recommendations. The predictive outputs are then embedded into a mixed-integer linear programming (MILP) model to generate optimal maintenance schedules that minimize total downtime and maintenance costs while satisfying operational constraints such as resource availability and production deadlines. The proposed framework is validated using a combination of publicly available predictive maintenance datasets and real-world industrial sensor data. Experimental results demonstrate a reduction of up to 22% in unplanned downtime and 15% in maintenance costs compared to reactive and preventive maintenance strategies, while maintaining high interpretability for domain experts. This integrated approach highlights the potential of combining explainable AI, predictive analytics, and mathematical optimization for sustainable and efficient industrial operations.

Background: Industry 4.0’s development requires digitalized manufacturing through Predictive Maintenance (PdM) because such practices decrease equipment failures and operational disruptions. However, its effectiveness is hindered by three key challenges: (1) data confidentiality, as traditional methods rely on centralized data sharing, raising concerns about security and regulatory compliance; (2) a lack of interpretability, where opaque AI models provide limited transparency, making it difficult for operators to trust and act on failure predictions; and (3) adaptability issues, as many existing solutions struggle to maintain a consistent performance across diverse industrial environments. Addressing these challenges requires a privacy-preserving, interpretable, and adaptive Artificial Intelligence (AI) model that ensures secure, reliable, and transparent PdM while meeting industry standards and regulatory requirements. Methods: Explainable AI (XAI) plays a crucial role in enhancing transparency and trust in PdM models by providing interpretable insights into failure predictions. Meanwhile, Federated Learning (FL) ensures privacy-preserving, decentralized model training, allowing multiple industrial sites to collaborate without sharing sensitive operational data. This proposed research developed a sustainable privacy-preserving Explainable FL (XFL) model that integrates XAI techniques like Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) into an FL structure to improve PdM’s security and interpretability capabilities. Results: The proposed XFL model enables industrial operators to interpret, validate, and refine AI-driven maintenance strategies while ensuring data privacy, accuracy, and regulatory compliance. Conclusions: This model significantly improves failure prediction, reduces unplanned downtime, and strengthens trust in AI-driven decision-making. The simulation results confirm its high reliability, achieving 98.15% accuracy with a minimal 1.85% miss rate, demonstrating its effectiveness as a scalable, secure, and interpretable solution for PdM in Industry 4.0.

Development of a river dissolved oxygen prediction model integrating spatial effects and multiple deep learning algorithm

Purpose: Predicting colorectal cancer liver metastasis (CRLM) is essential for prognostic assessment. This study aims to develop and validate an interpretable multimodal machine learning framework based on multiparametric MRI for predicting CRLM, and to enhance the clinical interpretability of the model through SHapley Additive exPlanations (SHAP) analysis and deep learning visualization. Methods: This multicenter retrospective study included 463 patients with pathologically confirmed colorectal cancer from two institutions, divided into training (n = 256), internal testing (n = 111), and external validation (n = 96) sets. Radiomics features were extracted from manually segmented regions on axial T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI). Deep learning features were obtained from a pretrained ResNet101 network using the same MRI inputs. A least absolute shrinkage and selection operator (LASSO) logistic regression classifier was developed for clinical, radiomics, deep learning, and combined models. Model performance was evaluated by AUC, sensitivity, specificity, and F1-score. SHAP was used to assess feature contributions, and Grad-CAM was applied to visualize deep feature attention. Results: The combined model integrating features across the three modalities achieved the highest performance across all datasets, with AUCs of 0.889 (training), 0.838 (internal test), and 0.822 (external validation), outperforming single-modality models. Decision curve analysis (DCA) revealed enhanced clinical net benefit from the integrated model, while calibration curves confirmed its good predictive consistency. SHAP analysis revealed that radiomic features related to T2WI texture (e.g., LargeDependenceLowGrayLevelEmphasis) and clinical biomarkers (e.g., CA19-9) were among the most predictive for CRLM. Grad-CAM visualizations confirmed that the deep learning model focused on tumor regions consistent with radiological interpretation. Conclusions: This study presents a robust and interpretable multiparametric MRI-based model for noninvasively predicting liver metastasis in colorectal cancer patients. By integrating handcrafted radiomics and deep learning features, and enhancing transparency through SHAP and Grad-CAM, the model provides both high predictive performance and clinically meaningful explanations. These findings highlight its potential value as a decision-support tool for individualized risk assessment and treatment planning in the management of colorectal cancer.

​​Machine learning (ML) is present everywhere in today's fleet management, driving all aspects from predictive maintenance to driver risk profiling. A paramount issue surfaces: the "black box" aspect of such models. This lack of transparency presents a major hurdle for regulators, fleet owners, and other non-technical stakeholders who must trust and verify the fairness and safety of such automated decisions. This paper serves the dual purpose of advancing predictive maintenance performance and addressing the critical gap of interpretability in high-performing but opaque ML models. We resolve this by proposing a framework that integrates state-of-the-art Explainable AI (XAI) techniques for transparency and compliance. We demonstrate the framework with a use case example: vehicle maintenance needs prediction from a real-world telematics dataset. Our contribution is threefold: first, we show a system architecture that integrates an explainability layer directly into the fleet ML pipeline. Second, we experimentally compare three widely used models: XGBoost, Random Forest, and TabNet, using SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) for generating both global and local explanations. Third, we provide interpretations of explanations from a regulatory perspective, accompanied by step-by-step examples of how these explanations can be used to answer fundamental questions about model behavior, feature importance, and potential biases. Our key findings are that SHAP delivers powerful global and local explanations and that LIME generates insightful, human-understandable explanations of individual predictions. In our experiment, the XGBoost model achieved an F1-score of 0.77 with a recall of 0.72, reducing projected unplanned maintenance cost by over 55% compared to less interpretable models. By implementing XAI, we provide a clear path toward rendering opaque fleet management systems transparent and compliant operations that satisfy regulatory mandates for safety, fairness, and accountability.

This study seeks to improve the accuracy of air conditioning load forecasting to address the challenges of load management in power systems during high-temperature periods in the summer. Given the limitations of traditional forecasting models in capturing different frequency components and noise within complex load sequences, this paper proposes a multi-level decomposition forecasting model using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), sample entropy (SE), variational mode decomposition (VMD), and long short-term memory (LSTM). First, CEEMDAN is used for the preliminary decomposition of the raw air-conditioning load series, with modal components aggregated by sample entropy to generate high-, medium-, and low-frequency subsequences. VMD then performs a secondary decomposition on the high-frequency subsequence to reduce its complexity, while LSTM is applied to each subsequence for prediction. The final prediction result of the air-conditioning load is obtained through reconstruction. To validate model performance, this paper uses air-conditioning load data from Nanchong City and Sichuan Province, for experimental analysis. Results show that the proposed method significantly outperforms the LSTM model without decomposition and other benchmark models in prediction accuracy, with the Root Mean Square Error (RMSE) reductions ranging from 40.26% to 74.18% and the Modified Mean Absolute Percentage Error (MMAPE) reductions from 37.75% to 73.41%. By employing the SHAP (Shapley additive explanations) method for both global and local interpretability, the model reveals the influence of key factors, such as historical load and temperature, on load forecasting. The decomposition and aggregation approach introduced in this paper substantially enhances forecasting accuracy, providing a scientific foundation for power system load management and dispatch.

Wind speed forecasting approach using conformal prediction and feature importance selection

Artificial Intelligence (AI) has emerged as a transformative force in the realm of industrial predictive maintenance (PdM), enabling businesses to anticipate equipment failures and optimize maintenance schedules. This paper examines the integration of AI technologies such as machine learning, deep learning, and IoT systems into predictive maintenance frameworks. Drawing on case studies and current implementations across various sectors, including manufacturing, steel production, and ICT infrastructure, the research examines key processes such as data acquisition, preprocessing, model development, and decision-making. It also evaluates the use of advanced sensors and wireless networks, the architecture of predictive models, and practical implementation strategies. Benefits, including reduced operational costs, minimized downtime, and enhanced safety, are discussed alongside challenges such as data quality, model accuracy, and organizational maturity. Ultimately, this paper underscores AI’s pivotal role in advancing Industry 4.0 by facilitating data-driven maintenance that is both intelligent and proactive. Keywords: Predictive Maintenance (PdM), Artificial Intelligence (AI), Industry 4.0, Machine Learning (ML), Internet of Things (IoT), Deep Learning, Remaining Useful Life (RUL), Condition Monitoring.

Fracture energy plays a pivotal role in ensuring the safe design of concrete structures. Currently, experimental testing remains the predominant methodology for exploring fracture energy in concrete. Nevertheless, this approach is hindered by protracted sample production cycles and test loading conditions that contribute to elevated expenses. Moreover, owing to the complex nonlinear behavior exhibited by concrete during the fracturing process, existing empirical formulas exhibit restricted precision when forecasting fracture energy. Therefore, in order to swiftly and accurately predict the fracture energy of concrete and investigate the impact of various factors on it, this study employs a deep learning algorithm to establish the correlation between parameters and fracture energy. Additionally, an interpretable deep learning prediction model for fracture energy is proposed, which is then compared with existing empirical formulas. Finally, the SHapley Additive exPlanations (SHAP) interpretability method is utilized to interpret and analyze the prediction results. The SHAP method can identify and visualize the contribution direction (positive/negative) and magnitude of the input features and reveal the relative importance of parameters at both local and global levels simultaneously. This analysis effectively explains the decision-making mechanism of the “black box” model and significantly improves the problem of insufficient interpretability that is common in traditional machine learning methods. The findings demonstrate that over 87% of the prediction results from the deep learning model in this study exhibit a relative error of less than 10% on the test set. The model effectively captures the intricate nonlinear relationship among characteristic parameters, exhibiting superior accuracy and generalization capabilities compared to empirical formulas. The SHAP values of the input parameters are visualized to assess their influence on fracture energy: initially, fracture energy increases and then decreases with increasing compressive strength, age, and coarse aggregate proportion; fracture energy increases with increasing maximum particle size of aggregate until it reaches 20 mm, after which it stabilizes; a high water–binder ratio reduces fracture energy; within the range of 400 mm, fracture energy increases with height, exhibiting a noticeable size effect; fracture energy increases with specimen width, but the size effect diminishes beyond 150 mm width; fracture energy decreases as span–height ratio increases; seam height ratio exhibits an initial increase followed by a decrease in fracture energy, with larger ratios showing a more pronounced size effect; an increase in ligament height enhances fracture energy while maintaining a significant size effect.

To improve the accuracy and interpretability of neural network enabled spectrum prediction, an explainable spectrum prediction framework based on Variational Mode Decomposition (VMD) and Long Short-Term Memory (LSTM) networks, integrated with the Shapley Additive Explanations (SHAP) method (VMD-LSTM), is proposed in this work. Firstly, the raw spectrum data is decomposed into multiple Intrinsic Mode Functions (IMFs) via VMD to reduce sequence complexity. These IMFs are then fed into the LSTM network in parallel to improve prediction accuracy. Secondly, the SHAP method is incorporated to evaluate the impact weights of individual IMF components on the prediction outcomes, revealing the model's decision-making logic. Finally, we weight the input data by multiplying each IMF by its SHAP value to optimize prediction performance. Simulation results based on real spectrum data demonstrate that the proposed VMD-LSTM significantly outperforms baseline models on the metrics of Weighted Quality Evaluation Index (WQE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and mean absolute error (MAE). By incorporating SHAP weights to refine the model input features, the framework not only provides transparent explanations for the black-box model but also reduces the average WQE, RMSE, and MAPE by 3.99%, 3.23%, and 3.67%, respectively.

A novel prediction model construction and result interpretation method for slope deformation of deep excavated expansive soil canals

Predictive maintenance (PdM) is a critical component of modern industrial systems, aiming to reduce downtime, enhance efficiency, and lower maintenance costs. With the increasing complexity of industrial equipment and the vast amounts of sensor data generated, traditional maintenance methods are no longer sufficient. Deep learning approaches have emerged as a powerful tool to address these challenges, offering enhanced capabilities for anomaly detection, fault prediction, and condition monitoring. This paper reviews the latest deep learning techniques applied to predictive maintenance in industrial systems, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), autoencoders, and reinforcement learning models. We explore their application to time-series data, sensor data analysis, and fault diagnosis in various industrial contexts. Additionally, the integration of deep learning with Internet of Things (IoT) and Industrial Internet of Things (IIoT) frameworks is examined, highlighting the synergies between these technologies. We also discuss the challenges and limitations, such as data quality, model interpretability, and computational cost, as well as the potential future directions for research. This work underscores the transformative potential of deep learning in optimizing predictive maintenance practices and advancing the efficiency of industrial operations.

Multivariate Time Series (MTS) data analysis is an important technique for detecting and localizing anomalies in industrial predictive maintenance. A real-world industrial mechanical system, such as a Cyber-Physical System (CPS), is a highly coupled system composed of multiple subsystems and devices, each equipped with sensors or actuators for control purposes. While abnormal system states can be detected by analyzing streaming data, accurately identifying their precise location remains challenging due to compensatory mechanisms inherent in CPS. Graph Convolutional Networks (GCNs) have recently proven effective for capturing spatial correlations among measured variables in CPSs. This chapter introduces a new method for uncovering hidden temporal and variable correlations in multivariate time series data from CPS. Specifically, we introduce an anomaly detection and localization framework that integrates multi-scale Graph Structure Learning (MAD-GSL) within a multi-stream sequence reconstruction architecture. This approach combines LSTM for temporal information extraction with GCNs to capture variable correlations across multiple scales, including long-term static and short-term dynamic relationships. A self-learning mechanism is used to identify long-term static graph structures, while feature similarities are utilized to infer short-term dynamic structures.

BackgroundAdmission to the neonatal intensive care unit (NICU) is a critical event for preterm infants, with significant implications for resource allocation and parental counseling. However, existing prediction tools are often limited by low accuracy or lack of interpretability. This study aimed to develop an interpretable machine learning (ML) model for predicting NICU admission in preterm infants using readily available prenatal and intrapartum features, with a focus on both the overall cohort and the clinically challenging subgroup of late preterm infants (34–37 weeks).MethodsA retrospective cohort of 2,610 preterm infants was analyzed. Features were selected using Boruta and least absolute shrinkage and selection operator (LASSO). Multiple models were trained and optimized via 5-fold cross-validation. The optimal model was evaluated using area under the curve (AUC), calibration, and decision curve analysis. Subgroup analysis was performed in late preterm infants (34–37 weeks) to assess model performance in this population. Interpretability was assessed with Shapley Additive exPlanations (SHAP).ResultsThe random forest (RF) model demonstrated superior and robust performance, achieving an AUC of 0.861 [95% confidence interval (CI): 0.830–0.891] in the validation set and 0.869 (0.841–0.897) in the testing set. SHAP analysis identified birth weight (mean |SHAP| value =0.17), prenatal checkup status (0.13), and gestational age (0.09) as the three most influential predictors. Low birth weight, lack of prenatal care, and gestational age below 32 weeks were associated with a significantly elevated risk of NICU admission. In the late preterm subgroup (34–37 weeks), the RF model maintained robust performance with an AUC of 0.842 (validation) and 0.838 (test), demonstrating good calibration and positive net benefit on decision curve analysis.ConclusionsThe interpretable ML model developed in this study accurately identifies preterm infants at high risk of NICU admission, with consistent performance in the late preterm subgroup. By providing individualized risk quantification and visual explanation via SHAP, it facilitates timely clinical decision-making and enhances clinician-parent communication. This tool holds significant potential for optimizing resource allocation and improving perinatal care pathways.

BackgroundRestless legs syndrome (RLS) is a common and debilitating complication in end-stage renal disease (ESRD) patients undergoing dialysis, significantly impairing sleep quality and quality of life. Screening of prevalent cases remains challenging. This study aimed to develop and validate an interpretable machine learning-based classification model for identifying RLS status in ESRD patients.MethodsA total of 396 ESRD patients (173 hemodialysis, 223 peritoneal dialysis) were enrolled from April to October 2024. Patients were randomly divided into training (70%, n = 287) and testing (30%, n = 109) sets. Feature selection was performed using LASSO regression with five-fold cross-validation, followed by Akaike Information Criterion (AIC) refinement. Nine machine learning algorithms were developed: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), K-Nearest Neighbors (KNN), Decision Tree (DT), Artificial Neural Network (ANN), Multivariate Adaptive Regression Splines (MARS), and Quadratic Discriminant Analysis (QDA). Model performance was evaluated using discrimination (AUC-ROC), calibration (Brier score, calibration curves), and clinical utility (Decision Curve Analysis, DCA). SHapley Additive exPlanations (SHAP) was employed to enhance model interpretability.ResultsFive variables were selected: β2-microglobulin, hemoglobin, diabetes mellitus, coronary heart disease, and alcohol consumption. SVM demonstrated optimal performance with AUC of 0.791 (95% CI: 0.702–0.879) in the testing set, outperforming other models. SVM achieved accuracy of 0.761, sensitivity of 0.711, specificity of 0.797, F1-score of 0.711, and Brier score of 0.183. Calibration curves showed good agreement between estimated and observed probabilities. DCA confirmed favorable net clinical benefit across threshold probabilities. SHAP analysis identified β2-microglobulin (mean |SHAP| = 0.131) and anemia as the most influential variables with diabetes, coronary heart disease, and alcohol consumption contributing moderately. SHAP dependence plots revealed interactions between β2-microglobulin and hemoglobin, as well as diabetes modifying the protective effect of higher hemoglobin.ConclusionWe developed and validated an interpretable SVM-based classification model for identifying RLS in ESRD patients using readily available clinical variables. This model demonstrates promising performance and requires prospective external validation in multi-center cohorts before clinical implementation. This tool may facilitate screening of prevalent RLS cases and inform clinical decision-making.