Revolutionizing Boiler Maintenance: Image Processing Techniques for Boiler Scale Detection

Abstract

Scale formation is a problem that affects the performance of industrial process systems, reducing thermal efficiency, consuming extra energy, and causing equipment breakdown. But so far, although identification and classification of scale type are also of vital importance in industry, there is still no automatic scale identification system available to scale types from a visual perspective. However, the automatic identification of organic scale on boiler has not been studied in the prior research. In this paper, the first such system is proposed based on machine learning and deep learning to classify boiler scale deposits from the real industrial images that, belonging to the most frequent types (CaCO3, Fe3O4, Miscellaneous Scales), discriminated so far. A real world dataset was created with the images obtained directly from the operation of the boiler systems of the industries in Bangladesh and were annotated by the experts of the domain. To improve the model accuracy and robustness, we tested different pre-processing pipelines and feature extraction methods including classical descriptors (HOG, LBP, GLCM), CNN embeddings (VGG16, 2DSCN), and hybrid methods. Between the classical machine learning classifiers the optimized XGBoost attained the highest accuracy of 90.12%. In the realm of deep learning, we introduce a ScaleNet V1 with custom-designed blocks inspired by both deep residual networks and squeeze-and-excitation modules, which also includes a trainable attention module and is designed in Keras Tuner. The results showed that ourScaleNet V1outperformed conventional networks such as EfficientNet V2L,, MobileNet V2 and ResNet18, by achieving a test accuracy of 93.42%, lower prediction latency, and higher class-wise results. We validated the contribution of each architectural component with an ablation study and performed LIME-based explainability during the training phase of the model, gaining interpretability of model decisions, and improving the industrial applicability and trust. The proposed framework provides for the first time a scalable, interpretable, efficient and real-time solution for the boiler scale monitoring, which can constitute another benchmark in intelligent maintenance systems in industrial process optimization.