| The hydraulic dam project is of paramount importance in solving our country’s water source problem.Due to the complex construction technology of the dam,the uncertainty of geological conditions,and long-term erosion by flowing water,high-risk dam surface diseases such as leakage and cracks are easily generated on the surface of the dam.If the diseases cannot be detected and warned in time,they will bring serious safety hazards to downstream towns.Currently,the dam inspection mainly relies on manual on-site inspection,which has a long inspection cycle and low efficiency.At the same time,inspection personnel also face high risks.Therefore,in view of the above problems,this paper proposes a deep learning-based algorithm for detecting surface diseases of hydraulic dams.The algorithm can maintain high detection accuracy and speed in complex environments,and has low computational requirements,making it suitable for deployment on mobile and embedded devices,and can effectively complete remote inspections.The specific research contents are as follows:(1)The thesis describes an improved detection algorithm based on YOLOv5.Firstly,in order to address the leakage and crack targets on dam surfaces that have nonrigid characteristics and features that are difficult to extract,an improved multi-scale visual Transformer structure is adopted to optimize the backbone network.This structure uses the multi-scale global information correlated with the multi-scale Transformer structure and the local information extracted by the convolutional neural network to construct an aggregate feature,fully utilizing the multi-scale semantic information and position information to enable the network to more effectively extract and learn target features.Secondly,a co-attention mechanism is introduced before each detection head in the feature detection layer to construct long-distance associations between pixels on the feature map,enhancing the network’s target localization ability in complex environments.Finally,the sampling algorithm for optimizing the network’s positive and negative training samples is optimized.By building the average fit and difference between prior and actual bounding boxes,the algorithm screens samples to assist candidate positive samples to respond to prior bounding boxes that are similar in shape,thereby helping the network converge faster and better,and ultimately improving the overall performance and generalization of the network.(2)The model has been optimized for lightweight.Firstly,redundant structures have been optimized by removing or replacing certain network layers.Then,the model has been pruned using a structured pruning method based on the BN layer’s γ coefficient,followed by fine-tuning to restore the pruned model’s accuracy,resulting in a significant reduction in the model’s parameter and computation complexity at the expense of minimal precision loss.Lastly,through structure reparameterization,the multiplebranch structure during the training phase has been transformed equivalently to a single-path structure during the inference phase,which further compresses the model while almost maintaining the same level of accuracy.Finally,on the dataset of dam surface diseases constructed by several small and medium-sized dams in Guizhou,the model improved in this paper has improved by10.7% and 17.3% in mAP(0.5)and mAP(0.5:0.95),respectively,compared to the original algorithm.After lightweight modification,the model becomes more compact,with a 77% reduction in parameter quantity,43% reduction in computing power,and79% reduction in model size,while achieving a 170% FPS improvement.Additionally,the development and application testing of the surface disease detection system for hydraulic dams have been accomplished.In conclusion,the algorithm developed in this paper effectively improves the efficiency of dam surface inspection,reduces manual risks,and holds engineering significance for the intelligence of water conservancy and hydropower. |
PDF Full Text Request