A Novel Industrial Defect Recognition Method Under Label Noise Based on Parallel Hybrid Penalty Network

Defect recognition optimizes industrial manufacturing processes by monitoring the health of equipment and structures. The success of current deep learning (DL)-based defect recognition methods relies on the assumption that the samples are label-noise free, yet the low-quality of industrial data inevitably introduces label noise, resulting in a substantial drop in recognition performance. To tackle this problem, this paper develops a parallel hybrid penalty network (PHL-Net). It solves two essential issues in defect recognition under label noise: low sample utilization and difficulty in recognizing the ambiguous sample in the decision boundary region, so that the industrial defect recognition performance can be greatly enhanced under label noise. PHL-Net consists of three stages: data selection (DS) for explicitly selecting clean and noisy data; parallel feature mining (PFM) for improving the sample utilization by label-driven mining and distance-driven mining; hybrid penalty learning (HPL) to improve decision reliability to label noise, thereby the ambiguous samples in the decision boundary region can be accurately recognized. In PHL-Net, PFM and HPL are co-optimized to facilitate data selection. In turn, reliable data selection prevents the model from overfitting label noise. Three groups of experimental results on real-world industrial datasets MPV-W and NEU-CLS have demonstrated the effectiveness of our PHL-Net over state-of-the-art alternatives. Note to Practitioners—Defect recognition plays a vital role in industrial production, ensuring product quality and efficiency. When training deep learning-based defect recognition models, having datasets with precise annotations is essential. Unfortunately, label errors, known as label noise, are common when annotating industrial datasets due to the need for domain expertise, subjective judgments, and confusing samples. This makes it challenging to create a large-scale and accurate industrial dataset. To address this issue, our goal is to develop a robust model that can achieve optimal recognition performance even on datasets with noisy labels. On this basis, this paper proposes a parallel hybrid penalty network (PHL-Net). It improves recognition performance under label noise by improving sample utilization and recognizing ambiguous samples near the decision boundary. In experiments, we utilize the noise transition matrix to introduce label noise for MPV-W and NEU-CLS. The results of the experiments demonstrate that our PHL-Net outperforms the comparison methods on industrial datasets, highlighting the practical significance of our approach. This provides inspiration and reference for industrial defect recognition under label noise.