Operation faults monitoring of combine harvester based on SDAE-BP

In order to solve the problem of deep feature extraction of nonlinear feature signal of operation faults of combine harvester and improve the diagnosis accuracy of fault recognition, a method based on Stack Denoising Auto Encoder -Back Propagation neural network(SDAE-BP) model was proposed in this study. Lovol RG50 combine harvester was used as the test prototype, according to the analysis of the working procedure and failure mechanism of each component of combine harvester, NJK-5002C speed sensor was used to collect the rotation speed signal of feeding auger, impurity auger, grain beat auger, fan, threshing cylinder and conveyor chain harrow, and the frequency signal of sickles and straw walker, and the collected data sets were used as the input of the monitoring system. The monitoring system was consist of IPC-610L embedded industrial computer, USB-4711 data acquisition module, EYOYO interactive display screen and LTE-1101J sound and light alarm device. The SDAE model was used to extract the deep feature of the input signal, the extracted deep feature was sent to the BP neural network and then the operation status of combine harvester was classified. During the training process, the first step was to train the DAEs (Denoising Auto Encoder) under different gaussian noises distribution center respectively, after all the DAEs was trained, stacking the DAEs all together and fine-tuning the model's parameters through the error back propagation algorithm. The noise center of DAE was far away from 0, which meaned that the original data was seriously damaged, the model could learn global coarse grained features, the noise center of DAE was close to 0, which indicateds that the damage degree of original data was low, and the model could learn local coarse grained features, in other words, training each DAE with different gaussian noise centers, the SDAE model would learn both global and local coarse grained characteristics which was of great significant to improve the model's expressive ability of deep feature. The experiments were carried out in 2018 to verify the proposed method, and the results showed the diagnostic accuracy rate reached 99.00%, which improved by 1.50 and 4.50 percentage points respectively compared with SDAE and BP neural network. The DAE-BP model was updated with the test data of 2019, and tested with the data of 2018 and 2019. The results show that the fault identification accuracy rate of the updated model for the test data in 2018 was 99.25%, and that for the test data in 2019 was 98.74%,which increased by 6.52 percentage points than that of the unupdated model. The model established in this paper can accurately identify the fault type of combine harvester, and has good robustness, which has reference value for the fault monitoring and early warning of rotating machinery. © 2020, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.