Scholarlink

<p>Walnuts are susceptible to internal and external defects, such as shell damage and kernel shriveling, during processing and storage, which makes rapid and simultaneous quality-state recognition challenging. This study developed a free-fall impact acoustic method combined with deep learning to classify intact, damaged, and shriveled walnuts. Impact sounds were transformed into time–frequency features, and one-dimensional waveform features, Mel-frequency cepstral coefficients, and Log Mel filterbank (Logfbank) features were evaluated using several neural-network models. The ResNet18 model using Logfbank features achieved the best baseline performance. An optimized model was then constructed by combining frequency-domain attention, exponential moving average parameter smoothing, and SpecAugment with Noise-Fill. Under five-fold cross-validation, the optimized model achieved an Accuracy of 96.5%, a Macro-F1 of 96.5%, and a Recall of 96.4%. Grad-CAM and occlusion tests showed that the model attended to different spectral–temporal regions across the three walnut categories, providing model-level interpretability for the classification decisions. These results indicate that impact acoustic signals combined with deep learning can provide a rapid, low-cost, and non-destructive approach for walnut defect recognition and may serve as a methodological reference for acoustic quality detection of other nut products.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Non-destructive detection of internal and external defects in walnuts using impact acoustics

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Taoyi Liu,
  • Shenao Chen,
  • Xianghao Chen,
  • Junwei Wang,
  • Lixia Li

摘要

Walnuts are susceptible to internal and external defects, such as shell damage and kernel shriveling, during processing and storage, which makes rapid and simultaneous quality-state recognition challenging. This study developed a free-fall impact acoustic method combined with deep learning to classify intact, damaged, and shriveled walnuts. Impact sounds were transformed into time–frequency features, and one-dimensional waveform features, Mel-frequency cepstral coefficients, and Log Mel filterbank (Logfbank) features were evaluated using several neural-network models. The ResNet18 model using Logfbank features achieved the best baseline performance. An optimized model was then constructed by combining frequency-domain attention, exponential moving average parameter smoothing, and SpecAugment with Noise-Fill. Under five-fold cross-validation, the optimized model achieved an Accuracy of 96.5%, a Macro-F1 of 96.5%, and a Recall of 96.4%. Grad-CAM and occlusion tests showed that the model attended to different spectral–temporal regions across the three walnut categories, providing model-level interpretability for the classification decisions. These results indicate that impact acoustic signals combined with deep learning can provide a rapid, low-cost, and non-destructive approach for walnut defect recognition and may serve as a methodological reference for acoustic quality detection of other nut products.

Graphical Abstract