Open Set Fault classification for Rotatory machine by DNN’s neuron activation similarity score

In real-world industrial settings, mechanical fault classification systems often encounter unknown fault classes beyond the K classes used for training. To prevent misdiagnoses and costly downtime, an effective system must possess open-set classification capabilities, distinguishing between known and unknown faults. Traditional deep neural network (DNN) approaches, relying on softmax probabilities as black-box outputs, frequently fail due to their inherent tendency to assign probabilities to all inputs, even those significantly deviating from the training distribution. In this paper, we propose a novel, generic, open-set classification technique for rotary machines. Our method is based on neuron activation similarity from convolutional layers in DNN architectures, which learn the different abstract level of features. This method incurs a very small amount of additional computational overhead during inference. Extensive empirical evaluations on eight diverse datasets, including six open-source benchmarks, demonstrate the state-of-the-art performance of our approach across different DNN architectures (ResNet18 and VGG11), highlighting its robustness and generalizability.