Adcapsnet: A Novel Efficient and Robust Anomaly Detection Capsule Network Model for Deteriorated Iiot Sensor Images

Abstract

With the rapid development of the IIoT,  the anomalies generated in industrial production, especially the hidden anomalies, such as slight rotation of the sensor images and position reversed sample images, will cause great damage to the normal operation of industry. Detecting anomalies hidden in deteriorated sensor images has increasingly become a hot spot. Although there existed many related researches, there are still some problems that are difficult to deal with such hidden anomalies. This paper proposes an efficient and robust semi-supervised anomaly detection capsule network (ADCapsNet) by integrating two reconstructed components based on CapsNet and two novel operations. Reconstructed parts include a changed convolution structure to better extract the features of the data, and a new added SecondaryCaps layer to better extract spatial relationships for anomaly detection. New operations mainly embody the optimized vector selecting operation for dynamic anomaly detection routing and the scoring operation using modified probability mechanism. Specially, the modified probability mechanism is adopted to widen the score gap between positive and negative samples. This model can accurately identify and output the spatial relationships. Extensive experiments on four data sets show that the ADCapsNet has good performance of anomaly detection for deteriorated IIoT sensor images.

Keywords: IIoT, anomaly detection, capsule network, modified probability