<p>Accurate and explainable fault diagnosis in robotic systems is crucial to ensuring operational safety and efficiency in modern manufacturing environments. However, detecting anomalies arising from non-conservative dynamics—particularly those encountered in industrial settings—remains challenging due to their nonlinear and context-dependent nature. The proposed framework supports early-stage fault identification in industrial robot joints, such as abnormal friction, gear wear, and backlash in the reducer. This enables improved maintenance scheduling and reduced downtime in manufacturing lines. This paper introduces a physics-informed fault diagnosis framework that leverages explainable neural attribution techniques to uncover and classify these complex behaviors. We first employ a Deep Lagrangian Network (DeLaN) to model the robot’s conservative dynamics based on physical principles. The discrepancy between the DeLaN-predicted torque and the measured torque is treated as a residual signal that captures non-conservative forces. A separate multilayer perceptron is then trained to model this residual, representing non-conservative effects explicitly. To interpret the learned representations, we apply SHapley Additive exPlanations (SHAP) to estimate the attribution of each input feature to each hidden neuron. These SHAP values form a high-dimensional feature map that reflects how non-conservative behaviors influence internal network activations. By training a classifier on these SHAP profiles, we achieve accurate fault detection while maintaining transparency in decision-making. Experimental results on industrial robot datasets demonstrate high classification performance and provide explainable insights into the neural encoding of faults. This approach offers a scalable and physically grounded methodology for fault diagnosis, overcoming the separation between black-box neural models and explainable, physics-aware decision systems in robotic applications.</p>