Adaptive Neural Network for Real-World Dynamic Learning Environment

The management of real world dynamics environments, characterized by evolving data distribution, poses a challenge for the reliable deployment of machine learning models in engineering applications. Machine learning models, particularly convolutional neural networks (CNNs), have shown remarkable effectiveness in tasks involving visual inspection and fault detection which has great potential to facilitate quality monitoring and control in a wide range of real world applications. However, the training pipelines of popular CNNs are typically based on the static data datasets, a condition that is hardly representation of real-world engineering systems where data often evolves over time. Therefore, CNNs can struggle to learn new tasks progressively and efficiently, leading to catastrophic forgetting over previous learned ones. In this paper, continuous learning strategies are compared to support the development of adaptive machine learning models capable of consistently updating model knowledge while preserving balanced performance under real world dynamic environments. A case study on solder joint quality control is provided is provided to illustrate their practice usage in production settings. Using a multiphase training strategy, we evaluate: (i) orthogonal gradient descent (OGD); (ii) learning without forgetting (LwF), augmented with entropy based confidence weighting; (iii) structured vertical parameter freezing, where a subset of filters is kept fixed across layers; and (iv) a frozen backbone with trainable head strategy that freezes early convolutional layers and adapts only the higher level classifier. The experimental simulations show that the naïve transfer learning technique quickly forgets previous patterns, while pure freezing strategies offer stability but limited adaptation facing distribution shift. In contract, OGD improves stability, but can still be adapted in new tasks. Our study demonstrates how different continuous learning strategies impact the use of features and plasticity of the model. It indicates that combined configuration of OGD, entropy weighted LwF, and structured parameter freezing yields the best stability–plasticity trade off and is recommended for long life and self-updated inspection systems.