APA-boosted segmentation neural network with flexible set expansion strategy for industrial fault detection

A crucial aspect of the industrial installation operation is proper maintenance. To ensure this, appropriate control and monitoring of the condition of the fitting are necessary. For this purpose, we propose new encoder–decoder architecture model for industrial fault segmentation. The neural network architecture is based on a three-block encoder/decoder module and a bottleneck. Information concatenation between the encoder and the decoder was also introduced to transfer information about low-level features from the encoder. Each decoder block was enriched with the Atrous Pixel Attention (APA) module, which allows for the enhancement of specific features, such as features of different scales, or spatial context and maintaining a global representation of the processed data. Additionally, we propose an algorithm for expanding the training set, which flexibly adapts to the current training progress. Technique extends the training set by adding modified samples with the highest Shannon entropy, which allows for reducing the impact of data imbalance in the original set. Moreover, the set can also be reduced in case the desired learning efficiency is achieved to prevent overfitting. We test the fault detection potential of this approach on a set of welding joint faults and a set of steel strips with patches, inclusions, and scratches using GDXray Welds and SD-saliency-900 datasets. The experiments performed showed that the proposed algorithm allows for high-accuracy detection and improves existing approaches reaching a Dice coefficient of 0.9347 and 0.8890 on GDXray Welds and SD-saliency-900, respectively.