The Research Of Blade Attachment Recognition Method Based On Semantic Segmentation Networks For The Tidal Stream Turbine

With the rise of global population and industrialization acceleration,the energy demand for all countries is increasingly promoted.However,the reserves of traditional energies such as oil,coal and natural gas are limited.Therefore,developing and utilizing renewable energies has become the consensus of researchers.Although the utilization ratio of wind and solar energies is relatively large,in recent years,the tidal stream energy stands out with its high energy density and predictability which gets the wide attention of the power generation industry.The tidal stream turbine(TST)is a mechanical device that absorbs the tidal stream energy to generate electric energy.Different from the onshore wind turbine,the TST operates in marine environments for a long time.Hence,land resources are not required and the life of surrounding residents is not affected.Nevertheless,the seawater contains corrosive elements and microorganisms that could gradually corrode the TST blades and breed visible attachments,which desperately influences the efficiency of power generation.Therefore,studying the attachment recognition method of TST blades is critical for prompt device maintenance.At present,the attachment fault detection method based on electrical signal analysis can judge whether faults exist or not.The diagnosis method based on image-level classification network can further determine the fault types.In spite of the high precision,the two methods cannot provide the results of fault localization and recognition.The semantic segmentation network(SSN)that realizes pixel-level classification can be a feasible solution.However,with the increasing operation speed of TST,the collected images will become blurry and difficult to recognize.Accordingly,this paper takes the TST blade attachment as the research object and proposes three SSNs to recognize the attachment under different operation conditions.The specific work includes:(1)sampling TST image data with five uniform and one non-uniform attachment distributions under the static to high speed operation conditions;then,conducting the semantic annotation on the image data.(2)for the static to medium speed operation conditions,employing the image generation algorithm based on random rotation augmentation to greatly reduce the burden of manual annotation;proposing a coarse-fine SSN where the coarse segmentation branch is responsible for the global recognition and the fine one accomplishes the local contour refinement;the recognition maps with precise contour localization can be outputted through the adaptive integration of the two branches;to obtain better overall recognition performance,C-Seg Net based on the techniques of max-pooling index preservation and feature concatenation is proposed.(3)for the medium-high and high speed operation conditions,utilizing the image generation algorithm based on continuous rotation augmentation to create more realistic labeled data;then proposing a semi-supervised video segmentation network(SVSN)based on the conditional adversarial generation strategy to complete accurate attachment recognition on the TST video frames with large motion blur.The generator and discriminator of SVSN are C-Seg Net and conditional fully convolutional network,respectively.(4)adopting the Monte Carlo sampling technique or the discriminator outputs of SVSN to estimate the recognition uncertainty.(5)validating the effectiveness of the three proposed SSNs on the public dataset of Cam Vid and TST dataset.