Boosting Trustworthiness in 6G-Enabled Smart Industries 5.0 Through Explainable Artificial Intelligence on Industrial Machine Anomaly Data

New industrial paradigms, like Industry 5.0 and the Industrial Internet of Things (IIoT), are developing in industrial contexts with the objective of foster quality. The growth of 6G-enabled smart industries requires interpretable, secure, and adaptive anomaly detection structures able to operate under adversarial, dynamic, and resource-constrained environments. Inappropriately, the usage of these technologies, successively, strengthens the attacking surface and poses a significant attack to the novel goals of Industry 5.0, like enhancing the user experience, resilience, and sustainability. In this setting, human-machine collaboration and cyber-physical integration request that anomaly detection systems surpass only prediction and include resilience, interpretability, and scalability. Traditional anomaly detection systems in Industry 4.0, like machine learning (ML)-based detection show remarkable challenges. These contain vulnerability to antagonistic attacks owing to vulnerable updates of the model, rigidity in adjusting to concept and temporal drift, and unclear decision processes that hamper trust and accountability. Therefore, this study develops an explainable artificial intelligence with a feature selection strategy enabled anomaly detection (XAIFSS-AD) technique for improving trustworthiness in 6G driven smart industry 5.0. The primary objective is to unravel the security issues in the smart industry environment via the anomaly identification process with the XAI approach to elaborate the proposed model. The presented XAIFSS-AD technique involves data scaling using a robust scalar which uses median and interquartile range to standardize the network data. In addition, the selection of features can be performed using the recursive feature elimination (RFE) process to eliminate the least significant features and enhance the classification process. A variational autoencoder with attention bidirectional long short-term memory (VAE-ABiLSTM) is deployed for anomaly detection. The performance of the classification model can be fine-tuned by Tree-structured Parzen Estimator (TPE) which optimally chooses the hyperparameter values. Finally, the Local Interpretable Model-Agnostic Explanations (LIME) methodology is applied to offer descriptions of the presented methods. The performance of the XAIFSS-AD approach is evaluated utilizing the benchmark datasets. The outcomes specify that the proposed model surpasses the other inspected systems and advanced models.