International audience ; One of the primary challenges associated with bolts, frequentlyused in assembly structures and systems, is the issue of torqueloosening. This problem can arise due to shock and vibration,potentially leading to significant damage and structural failure.The difficulty in identifying and monitoring torque looseningarises from the variability and nonlinear effects present in boltedjoints. In this paper, we proposed a machine-learning algorithmarchitecture designed for pattern recognition, detection, andquantification of torque loosening in bolted joints. This approachcombines unsupervised and supervised machine learning algorithms toaddress the challenges of assessing the bolt torque looseningissue. Our algorithm utilises a damage index, calculated from thefrequency response of the jointed system using the FrequencyResponse Assurance Criterion, as input data for theunsupervised–supervised classification algorithm. Thisclassification ML algorithm effectively identifies and categorisesinstances of torque loss by analysing indirect vibrationmeasurements, even in situations where the bolted system’s state isunknown. Additionally, we introduce a regression algorithm toquantify torque loosening levels. The results obtained from ourproposed machine learning algorithm to overcome torque loosening inbolted joints show that the inherent uncertainties of a data-drivenapproach intrinsically influence torque-related issues. Thisassessment is based on experimental raw data collected underdiverse test conditions for the bolted structure. We employ a rangeof validation and cross-validation metrics to evaluate theeffectiveness and accuracy of these ML algorithms in detecting anddiagnosing torque-related issues. These metrics play a crucial rolein assessing the algorithms’ efficiency and precision indetermining the state of the bolted connection and thecorresponding torque levels with associated uncertaintyquantification.