| Mechanical equipment are used in almost all industries,including manufacturing,construction,power generation,transportation,and mining.Their degradation as a result of aging and wear decreases their performance reliability and increases the potential for faults and failures.Ferrography analysis is a powerful tool used in predictive maintenance to detect and analyze wear particles in lubricating oil samples.However,despite several research conducted in ferrography analysis,there are still many drawbacks restricting the automation of ferrography analysis.To address the drawbacks of previous research and establish a well-thought-out intelligent system for ferrography analysis,this research proposed an object detection-based approach to segment wear particles from ferrographic images,identify them,extract their parameters,and determine the wear severity.First,a custom ferrographic images dataset was created by ferrographic experiment to pave the way for the training and evaluation experiments of wear particle segmentation models.The custom dataset construction is achieved by executing an image acquisition experiment,and employing different augmentation methods such as geometric transformation,color space conversion,and spatial filtering.In addition,the identification rules and characteristics required to predefine the wear particles in the dataset are demonstrated.At the end,the custom dataset is split into training,validation,and test set with a proportion of 70%,20%,and 10% respectively.The achieved custom dataset included 2750 images with 6400 particle instances available in these images.Furthermore,the varying wear particle segmentation models are established,alongside a comprehensive understanding of their working principles,mathematical operation,and involvement in particle feature extraction from ferrographic images.Mask RCNN,Cascade Mask RCNN,and fully convolutional neural networks(FCN)were proposed for segmentation of wear particle.Furthermore,these models are modified based on specific criteria such as anchor size,backbone selection,and other factors to be appropriate for wear particle segmentation.Subsequently,the established wear particle segmentation algorithms are trained and evaluated by carrying out multiple experiments.Firstly,the several training and optimization experiments are carried out to determine the optimal hyperparameters and achieve the best outcome of Mask RCNN.Then,followed by training and evaluating the Cascade Mask RCNN which demonstrated an increased segmentation accuracy of4% over Mask RCNN.At the end,FCN is trained and evaluated,this framework also exhibited a satisfactory segmentation result and provided a basic framework to further extract the particles parameters such as size and coverage area.Finally,the wear severity judgment based on ferrographic image and convolutional neural networks(CNNs)was studied.This work first extracted wear particles parameters such as size,quantity,and coverage area were using the FCN and Open CV algorithms.Then,the extracted parameters were statistically analyzed to categorize ferrographic images into three levels of wear severity,namely are Level-1 which represent the normal wear state of the machine;Level-II which denotes a progressive wear state;Level-III which indicates a catastrophic wear condition.At the end,the proposed models(Res Net-34,Mobile Net V2,Google Net)were trained and tested to evaluate their performance and select the model with highest accuracy to realize the classification task.This research includes 60 figures and 14 tables. |
