1 January 2019 in Neurocomputing
1 May 2019 in Neural Networks
12 October 2018 in IEEE Transactions on Power Electronics
For the inverter fault diagnosis of PMSM drives, it is difficult to simply and conveniently use the small low frequency data without needing additional hardware and affecting the accuracy of the diagnosis results. This problem often leads to unnecessary computation and wastes a lot of corresponding system memory, increases the hardware requirements and realization difficulty, affects practicability. In order to improve this problem, to make the design, debugging and implementation of fault diagnosis more convenient and low cost, and further to improve the practicability of the algorithm, a novel multiswitches fault diagnosis algorithm is proposed. Firstly, a second low frequency processing method is used to obtain the small low frequency data simply from the feedback signals of the controller. These small low frequency data contain the main feature information of different switch states. Secondly, these small low frequency data are effectively normalized by the single extremum normalization method, which is based on the symmetry features of different state data. These processed data can keep the asymmetry feature of fault state of inverter very well. Thirdly, the main fault components and features are extracted from the distortion part and envelope change of the processed small low frequency data. Fourthly, echo state network is used by combining with the extracted features to implement the intelligent classification. Different from the traditional neural network, the design of structure of hidden layer network is simplified, and the network training speed is very fast. Therefore, the debugging of this network is very convenient. Compared with the existing fault diagnosis algorithms, the proposed algorithm is convenient and low-cost to realize the multi-switches fault diagnosis. The effectiveness of fault diagnosis algorithm is verified by the experiment.
1 January 2018 in Neurocomputing
19 October 2022 in Journal of the Franklin Institute
31 May 2020 in Neural Processing Letters