An accurate and rapid identification is required for the Clivia miniata pests and diseases in the greenhouse and garden in recent years. In this study, the YOLO V4-TLite algorithm was proposed to detect the Clivia miniata pest using a mobile terminal. The real-time performance and high accuracy were also achieved to reduce the over-reliance on the high-computing and high-power hardware. Firstly, the dataset of Clivia diseases and insect pests was collected from the real planting in the greenhouse. The images were captured at the early and middle stages of Clivia diseases and insect pests in winter and spring. Secondly, a low-cost improved partial convolution was used to replace the traditional one in the backbone network using the YOLO V4-Tiny model. The improved model was then obtained with the high speed of operation and the low consumption of memory. Thirdly, an improved structure of the inverse residual network was used to form a lightweight backbone network. The hardware compatibility was also enhanced to reduce the large consumption of random storage in the depth of the backbone network in the YOLO V4-Tiny model. The high operation speed of the model was obtained with the compatibility of the mobile terminal with the limited resources. Fourthly, the weight-sharing convolution was combined with the conventional convolution for channel fusion. The high robustness and accuracy of the network were obtained to reduce the redundant feature maps and their attention distraction in the traditional convolution layer of the YOLO V4-Tiny model. Finally, the improved model was deployed on the ROCK 5B mobile. Three types of Clivia miniata pests were then tested: leaf blight, maculopathy, and coccid. The experimental results showed that the better performance of the improved model was achieved with the mean average precision (mAP) of 78.5% at an intersection over union (IoU) ratio of 0.5, memory usage of only 4.8MB, and the floating point operations (FLOPs) of 1.3 G. The desktop single detection speed was 0.005 s with 70 W power consumption. On the mobile side, the CPU single detection speed was 0.239 s with 10 W power consumption. The NPU single detection speed was 0.018 s with 7 W power consumption. Compared with the original YOLO V4-Tiny model, the mAP50 of the YOLO V4-TLite model increased by 12.6 percentage point, whereas, the model size decreased by 78.6%. The computational efficiencies of the YOLO V4-TLite model were improved by 37.5 and 85.9 percentage point on the desktop and mobile side, respectively. While the power consumption demands were reduced by 26 and 2 W, respectively. The mAP50 values were 3.9, 2.3, 1.6, and 1.3 percentage point higher, respectively, compared with the target detection models of YOLOV11-N, YOLO V10-N, YOLO V7-Tiny, and YOLO V5-S. The YOLO V4-TLite model can be expected to detect the Clivia Miniata pest and disease on the low resource and power mobile. Better performance was also achieved, compared with the existing mainstream YOLO series models.

A new lightweight convolutional neural network (CNN) model called the deep separable convolutional neural network (DSCNN) based on structural reparameterization is proposed, aiming at the single-branch deep convolutional approach used in the majority of the current CNN models, which not only affects the expressive ability of the model but also occupies a large number of parameters and Flops. Firstly, the feature extraction module (reparameterization MiXer, RepMiX) of the model can fuse information between different channels and spatial locations, realizing multi-scale feature fusion. Secondly, the reparameterized asymmetric spatial operator (RepASO) in the DSCNN learns different channel feature information through a multi-branch structure with different functions, which improves the model feature learning ability; meanwhile, both RepMiX and RepASO combine the structural reparameterization technique and the idea of depth wise separable convolution (DS-Conv) to realize structural decoupling in the training and inference phases, which accelerates the model inference while reducing the model parameters and Flops. Finally, comparative experiments are carried out on the Tiny-imagenet-200, CIFAR-10, and CIFAR-100 datasets in addition to a self-constructed dataset for the classification of aluminum ingot surface defects. The experimental results demonstrate that the DSCNN maintains competitive accuracy while achieving higher floating-point speeds.

In order to improve the prediction accuracy of ultra-short-term wind power, a combined prediction model based on variational modal decomposition (VMD), hierarchical principal component analysis (HPCA), and gated recurrent unit (GRU) neural network optimized by COOT algorithm was proposed. Firstly, the submode number of VMD was determined by the energy difference method so that the original power sequence with strong nonlinearity was decomposed into a set of relatively stationary submodes. Secondly, the correlation degree value between high-dimensional meteorological features and power sequence was calculated by gray relation analysis, and the ranking and stratification were carried out. The first principal component of feature variables in each layer was extracted by principal component analysis (PCA) to realize the dimensionality reduction of high-dimensional meteorological features. Finally, the COOT algorithm was introduced to optimize the hyperparameters of the GRU prediction model, accelerate the model convergence speed, and improve the model prediction accuracy. Simulation analysis was carried out on the measured data of a wind farm in Guizhou Province, and the results show that compared with the prediction results of the traditional GRU model, the root mean square error, mean absolute error, and mean absolute percentage error of the proposed method are reduced by 67.41%, 72.25%, and 45.69%, respectively, and the prediction accuracy of the proposed method is higher than that of the other four combined prediction models, which effectively improves the prediction accuracy of ultra-short-term wind power.