This article presents an open-switch fault detection method for a hybrid active neutral-point clamped (HANPC) inverter based on deep learning technology. The HANPC inverter generates a three-level output voltage with four silicon switches and two silicon carbide switches per phase. The probability of open fault in switching devices increases because of the large number of switches of the entire power converter. The open-switch fault causes distortion of output currents. A convolution neural network (CNN) comprising several convolution layers and fully connected layers is used to extract features of distorted currents. A CNN network was trained using three-phase current information to determine the location of the open-switch fault. Our proposed CNN model can accurately detect approximately 99.6% of open-switch faults without requiring additional circuitry and regardless of the current level within an average time of 1.027ms. The feasibility and effectiveness of the proposed method are verified by experimental results. INDEX TERMS Open-switch fault detection, hybrid active neutral-point inverter, silicon carbide, deep learning, convolution neural network.

Smart factories should be able to respond to catastrophic situations proactively, such as recalls caused by production line disruptions and equipment failures. Therefore, the necessity for predictive maintenance technology, such as fault detection or diagnosis of equipment has increased in recent years. In particular, predicting the faults of collaborative robots is becoming increasingly crucial because smart factories pursue efficient collaboration between humans and devices. However, collaborative robots have the characteristic of executing programmable motions designed by an operator, rather than performing fixed tasks. If existing fault diagnosis methods are applied to non-fixed programmable motions, problems arise in terms of setting absolute criteria for fault analysis, interpreting the meanings of detected values, and fault tracking or fault cause analysis. Therefore, we propose a method of programmable motion-fault detection by analyzing motion residuals to solve the three problems mentioned above. The proposed method can expand the fault diagnostic range of collaborative robots.

No abstract

Abstract:The mechanical rubbing of a polyimide (PI) layer with a velvet cloth has been dominantly used to induce a uniform alignment of liquid crystal molecules in the manufacturing process of liquid crystal displays (LCDs). The rubbing process is affected by its process parameters and also by the nature and properties of the rubbing cloth used. We fabricated the rubbing cloths with various pile densities and different weaving structures and then investigated how the parameters were related to the effect of the cloth on the rubbing process and the properties of rubbed PI alignment layers (ALs). As the pile density increases at same rubbing process parameters, the degree of molecular orientation of the rubbed AL and its surface roughness increase with an increase of the contact area between the rubbing cloth and the AL surface. Furthermore, a more uniform rubbing is also achieved. The results in this work showed that higher pile density of the rubbing cloth induces a better rubbing effect, which is indeed more favorable to the LCD process. It was also observed that the weaving can exert some influence on the rubbing process.