通过计算机视觉和机器学习改进腐蚀管理并确保资产完整性

The oil and gas sector is challenged with coating breakdown and corrosion detectionacross the entire asset management lifecycle. Energy companies lose billions of dol-lars annually due to damage caused by ineffective asset integrity management. Oiland gas industry leaders can mitigate the problem by moving from reactive to proactiveasset maintenance to prevent asset failure and downtime, lower maintenance costs, andensure employee and environmental safety. Discover how to apply advanced technologies to proactively monitor for corrosion, screenfor anomalies, and detect additional defects. THE PINNACLESUCCESS STORY Our client,Pinnacle, is a leader indeveloping data-driven solutions forthe oil and gas, chemical, mining,and water and wastewater industries. Thecompany helps industrial facilities betterleverage their data to ensure increased pro-duction, optimized maintenance spending,and improvements in process safety andenvironmental impact. Data collection and organization areamong Pinnacle's strong points, and theirclients can confidently make strategicdecisions. However, they wanted tofurther help their customers createrobust data reliability programs. Pinnacle wanted to embrace cutting-edgetechnology and directly link data collectionand intelligence segments in one solution.Recognizing they needed edge technologyexpertise, Pinnacle reached out toSoftServe. SoftServe was asked to provethe feasibility and validate the businessvalue of a computer-vision based approachfor detecting and classifying coatinganomalies and corrosion using visual data. Pinnacle is focused on helping its customersmanage risks, optimize costs, and ensurecompliance through its Data-DrivenReliability framework. This frameworkconsists of four categories: data collection,organization, intelligence, and strategicdecisions. Pinnacle’s Research andDevelopment team continues to invest inthe intelligence segment of this framework,specifically through QuantitativeReliability Optimization (QRO), to deliverdata-driven insights for its clients. OUR APPROACH SoftServe’s experts helped Pinnacledevelop effective ML models forcorrosion segmentation and textureclassification. The final model can generatecoating breakdown and corrosion maps,evaluate damaged areas, and mark corro-sion damage according to a few severityclasses. We split the dataset of images into threeparts: training, validation, and testing forboth corrosion segmentation and textureclassification. The texture classification solution uses thesliding window approach and classifies thecrops of an image with the size of 128 px x128px. This size showed the best result inour previous experiments compared to oth-er tested sizes. During experiments with thesupervised and unsupervised classificationof the textures, we tested the most pop-ular machine learning algorithms, such asXGBoost, K-NN, Centroid-based, SVC, plusothers. We decided to use ResNet18 as a fi-nal model for texture classification becauseit gave the best data results. We conducted several experiments to ex-amine the problem from a different angleto help determine the best solution for theclient. We tried using open-source data anda classifier based on embedding vectors,but the metrics were worse. Since therearen’t many open-source textural datasets,we decided to gather our own big datasetswith specific textures. For corrosion segmentation, we tested sev-eral fully convolutional networks with theUNet-like architectures. During the hyperpa-rameter tuning process, we tested severalparameters of the segmentation pipelinesuch as model architecture, model encoder,loss function, number of epochs, batch size,input size, and image input method. Thisallowed us to find the optimal configurationof the pipeline. In the final implementation,we used the PSPNet architecture and anEfficientNet encoder. Pinnacle provided panoramic images con-verted to a 2D view suitable for labelingpurposes. SoftServe’s experts, in collabora-tion with the Pinnacle team, conducted thelabeling workshop and developed imageanalytics to evaluate distribution, calculatestatistics of coating damage and corrosion,etc. The labels for the datasets were: •Corrosion segmentation:corrosion and stains. •Texture classification:severe corrosion, light corrosion, paintdefects, rust staining, and normal paint. CORROSIONINFERENCE PIPELINE Close-up image inference SOLUTION OUTCOMES Together with Pinnacle, the SoftServe team successfullycompleted the following project deliverables: •CORROSION SEGMENTATION MLMODEL.Detects corrosion and ruststains on images of the facilities takenby a panoramic handheld camera. •USER INTERFACE.A simple browserapplication hosted on the securedVM allowing a user to tune processingparameters, run ML models in real timeon the selected images, and see the bigpicture on the map. The UI allows theuser to process images one by one witha preview or a batch processing of theentire directory passed as a parameter. •TEXTURE CLASSIFICATION ML MODE