Seal defect detection in food packaging integrity inspection presents substantial challenges, primarily owing to minimal textural disparities between defective and intact regions under industrial illumination. These challenges intensify in production environments where moderate class imbalance intersects with pronounced inter-class similarity. This study introduces an enhanced MobileNetV4 architecture incorporating adaptive feature learning mechanisms to overcome these limitations. A novel LocalAttention module, employing dynamically weighted rectangular sliding windows, augments the model's texture discrimination capabilities by effectively isolating elongated structural anomalies characteristic of seal defects. To address suboptimal illumination conditions, we integrate Pinwheel-shaped Convolution (PConv), which leverages Gaussian-inspired distribution patterns to amplify weak signal detection in poorly lit environments. Additionally, an Adaptive Similarity-Modulated Loss function concurrently mitigates class imbalance while refining feature space organization for morphologically close categories. Evaluated on a grain packaging dataset from Zhangzhou, Fujian, the proposed framework achieves 99.54% accuracy and 99.75% F1-score with only 3.7M parameters and 25.27 FPS on Raspberry Pi 4B. To validate industrial reliability, robustness experiments under simulated LED degradation and sensor noise confirm consistent performance advantages over baseline architectures, and an 8-hour continuous stress test demonstrates stable edge deployment without thermal throttling. Cross-domain validation on NEU-CLS further achieves 99.72% accuracy with superior parameter efficiency, suggesting transferability to related industrial defect classification tasks.

Steel surface defect classification constitutes a critical component of industrial quality control. Current computer vision-based classification approaches, however, exhibit substantial dependence on dataset characteristics, where inconsistent defect dimensions and low-contrast ambiguities significantly compromise detection accuracy. While emerging State Space Models (SSMs) such as VMamba and Vision Mamba offer efficient long-range modeling, they often lack the multi-granular feature extraction capabilities necessary for identifying such subtle industrial anomalies. To bridge this gap, we propose HiAM-Mamba, a hierarchical architecture that synergizes the linear complexity of SSMs with adaptive multi-scale processing. Distinct from existing Mamba variants, our approach integrates a Multi-Scale Fusion (MSF) module to aggregate features across varying granularities and a Gated CNN to capture fine-grained textural patterns. Furthermore, we introduce an Adaptive Feature Recalibration mechanism to dynamically suppress environmental noise and a hierarchical head for holistic decision fusion. Extensive experiments on the NEU-CLS and X-SDD benchmarks demonstrate that HiAM-Mamba achieves state-of-the-art accuracy (99.90% and 99.15%), significantly outperforming current CNN and Transformer-based methods while maintaining superior computational efficiency.

Effective and intelligent fault diagnosis is essential for ensuring the operational safety and reliability of gearbox systems. In practical engineering environments, however, weak fault-related features are often obscured by strong background noise, pronounced nonstationarity, and time-varying operating conditions, which significantly degrade the performance of conventional feature extraction techniques. To address these challenges, this paper proposes an adaptive feature extraction approach that integrates the complementary advantages of variational mode decomposition (VMD), Teager energy operator (TEO), and multi-scale permutation entropy (MPE) to enhance the characterization of weak and transient fault signatures. Vibration signals associated with different fault conditions are first adaptively decomposed into a series of intrinsic mode functions (IMFs) using VMD, enabling the effective separation of fault-sensitive components and enrichment of fault-related information. Subsequently, an enhanced multi-scale permutation entropy (EMPE) method is developed to emphasize transient impulsive characteristics and capture fault-induced complexity variations across multiple temporal scales. By jointly exploiting instantaneous energy modulation and multi-scale dynamical complexity analysis, the proposed approach exhibits improved sensitivity to weak fault signatures and enhanced robustness against variable operating conditions. The effectiveness and generalization capabilities of the proposed framework are validated using three experimental datasets involving gearboxes and rolling bearings under diverse operating conditions. Comparative results demonstrate that the proposed method outperforms conventional entropy-based approaches in terms of fault feature separability and diagnostic performance, highlighting its potential for practical condition monitoring and fault diagnosis of rotating machinery.