T2Net: An Attention-Enhanced Network for Trustworthy Thermal Imaging-Based Lubrication System Condition Assessment

The lubrication system is essential for the normal operation of rotating machinery, as its malfunction often results in premature transmission failures. Therefore, accurately assessing the condition of the lubrication system is vital for implementing effective preventive maintenance. Traditional condition assessment methods often depend on vibration sensors, which present challenges such as difficult setup in harsh operational environments and limited sensitivity to non-structural damage in lubrication system. To address this issue, an attention-enhanced network is proposed for trustworthy thermal imaging-based assessment of lubrication system conditions. First, a multi-scale attention mechanism is integrated into the network to improve its ability to identify fault-related temperature features, thereby enhancing the accuracy of condition assessment. Next, Bayesian batch normalization is employed to train the network by introducing randomness into batch normalization, enabling the estimation of uncertainty in the model's predictions and facilitating effective out-of-distribution detection. Finally, experimental results from thermal imaging data of gearbox demonstrate that the proposed method not only improves assessment performance but also achieves high accuracy in out-of-distribution detection.