Health condition monitoring of a complex hydraulic system using Deep Neural Network and DeepSHAP explainable XAI. (January 2023)

• Record Type: Journal Article
• Authors: Keleko, Aurelien Teguede; Kamsu-Foguem, Bernard; Ngouna, Raymond Houe; Tongne, Amèvi
• Is Part Of: Advances in engineering software. Volume 175(2023)
• Publication Date: 2023-01
• Subjects: Hydraulic system -- Deep Neural Networks (DNN) -- Faults classification -- Sensors -- Condition Monitoring (CM) -- Anomaly detection -- Diagnosis -- EXplainable Artificial Intelligence (XAI) -- Deep SHapley Additive exPlanations (DeepSHAP) -- Ethics; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2022.103339
• Languages: English
• ISSNs: 0965-9978

Abstract

Abstract: This paper presents a detailed framework for Condition Monitoring (CM) based on hydraulic systems and multi-sensor data. Nowadays, the CM technique is increasingly deployed for the optimization of quality and manufacturing processes. It is used as a decision-making support tool in maintenance operations or activities. In this environment, the diagnosis, prognosis, or monitoring of interconnected machines has become a crucial issue for improving the cost-effectiveness of manufacturing industries. Some models are available to monitor or predict the degradation of elements within a hydraulic system such as coolers, valves, internal pump leakage, or the condition of the hydraulic accumulator. In this case, we have focused on a data-driven approach, concentrating on the Deep Neural Networks (DNN) multi-class classification for imbalanced data that are adapted to predict the real operating states of the system. Despite their performance, questions remain concerning the reliability of the DNN as "black box" models when used in complex applications, notably regarding the decision-making processes and the possible ethical, socioeconomic, and transparency impacts upon stakeholders. Regarding the explanation approach, we have exploited the Deep SHapley Additive exPlanations (DeepSHAP) methodologies to provide trustworthy results, and to explain the importance (weight) or role that each sensor plays and its contribution to the DNN algorithm's decision-making. The obtained framework based on two principal modules illustrates that the DNN classifier model when evaluated by Accuracy, F1-Score, Recall, and Precision metrics are robust and perform efficiently. Finally, using the DeepSHAP technique provides an explanation of the results of the developed model, and helps humans to understand, interpret and trust the model, with an associated increase in the support or the stimulation of Artificial Intelligence (AI) models applications on large-scale problems including industrial sectors. Highlights: A Detailed framework for condition monitoring (CM) based on hydraulic systems. Using the DNN for Fault classification based on multi-sensor data. The DNN module is robust and efficient for multi-class and unbalanced data. Overview of eXplainable Artificial Intelligence (XAI) approaches. Explains the decision-making rules by the DNN known as a "black box" model. Using the DeepSHAP approach to highlight local and global contributions. Outline the most important sensors to be monitored. Module helps to understand and support practitioners in maintenance operations.