EXPLAINABLE AI BASED PREDICTIVE MAINTENANCE AND ANOMALY DETECTION SYSTEM FOR INDUSTRY 4.0 USING MACHINE LEARNING AND DEEP LEARNING

Abstract

The advent of Industry 4.0 has revolutionized the way manufacturing companies operate by monitoring, maintaining and optimising industrial equipment. One of the most effective uses of this transformation is Predictive Maintenance (PdM), which moves maintenance from a reactive and time-based approach to a data-driven, condition-based and anticipatory approach. There is a major lack in the literature and currently in systems: most of the predictive maintenance models focus either on predictive accuracy or explainability, or on explainability without considering robustness under realistic data scenarios, like sensor noise and class imbalance. This big project proposes an Explainable AI (XAI) framework for Predictive Maintenance and Anomaly Detection System, designed and tested with the widely used UCI Machine Learning Repository AI4I 2020 Predictive Maintenance Dataset. To systematically compare the three representative machine learning models, Logistic Regression (LR), Random Forest Classifier (RF) and a Feed-Forward Neural Network (Multi-Layer Perceptron, MLP), three carefully designed experimental conditions are provided: (1) a clean baseline dataset, representing ideal data quality; (2) a synthetically noised dataset, with Gaussian noise added to the data, representing real world sensor drift, calibration errors and measurement uncertainty; and (3) a synthetically over-sampled dataset (via Synthetic Minority Over-sampling Technique, SMOTE), to address the class imbalance encountered in industrial failure prediction tasks. The project puts in place an Autoencoder-based unsupervised module for anomaly detection in addition to supervised classification. The Autoencoder is trained with operational data without any failure labels and considers a failure as being flagged when the reconstruction error exceeds the 95th percentile set using a statistically motivated threshold, which is determined by the learned data, setting the limit to be the values of the normal machine behaviour. Two complementary approaches are used to obtain model explainability: using SHAP (SHapley Additive exPlanations) with the Random Forest through TreeExplainer, and LIME (Local Interpretable Model-Agnostic Explanations) with the Neural Network. SHAP offers explanations at the individual instance level using waterfall plots and explanations at a higher level of abstraction by ranking the most important global features via a beeswarm plot, the mean absolute SHAP bar plot, and a feature interaction dependence plot. Local linear surrogate explanations are offered for individual predictions by LIME, allowing maintenance engineers to grasp the model behaviour at the instance level. Experimental results indicate that the Random Forest model gives the best F1 score (1.00) for the clean dataset, followed by SMOTE augmented data (0.99), and lowest under Gaussian noise (0.95), which is the highest noise robustness among the three models. Even on well-structured tabular PdM data, Logistic Regression shows a consistent F1-score of 0.99 for all three data conditions, making it a very well performing simple and interpretable model that remains competitive over a variety of data conditions. The highest variance is in the case of the Neural Network, which achieves F1=0.93 for clean data, F1=0.98 for noisy data (with implicit regularisation effects of the added noise), and F1=0.93 for SMOTE data. The binary failure mode flags — PWF (Power Failure), HDF (Heat Dissipation Failure), OSF (Overstrain Failure) — clearly stand out as the most important failure predictors, while Rotational Speed and Torque are the most impactful continuous feature drivers. LIME explanations from the Neural Network align with SHAP results, thereby offering cross-method validation of explainability results. The distribution of the reconstruction errors of the Autoencoder reveals the separation between the normal and anomalous instances, and the Autoencoder is able to isolate operational records that are anomalous. The results above clearly proof the synergy of the classical and deep learning models with the explainable AI, unsupervised anomaly detection, resulting in a strong, explainable, practically viable PdM solution that can be well adapted to the Industry 4.0 manufacturing environments. It is executed with open source Python libraries and is accessible on standard hardware, so it can be utilized by small and medium scale enterprises without any special computational hardware.