Scholarlink

<p>This work presents the development of an intelligent diagnostic system based on sound analysis and artificial neural networks, aimed at fault identification and severity level classification in worm gear reducers. The following faults were reproduced in a dedicated experimental test bench: angular misalignment, overload, and inefficient lubrication. Sound signals were captured using a smartphone, providing a practical and low-cost solution. Signal analysis was conducted in both time and time–frequency domains, with the extraction of statistical metrics (kurtosis, RMS, and crest factor) and the energy of Discrete Wavelet Transform (DWT) coefficients, using 11 levels of decomposition. Input variables for the neural networks were selected based on ANOVA and Random Forest tests, ensuring model robustness and interpretability. Two independent Multilayer Perceptron (MLP) networks were implemented: one to classify the operating condition (normal, misalignment, overload, or inefficient lubrication), and another to classify severity levels (mild, moderate, or severe). The final architecture used 15 neurons in the hidden layer, with ReLU activation and a Softmax function in the output layer. Accuracy reached 98.889% for operating condition classification and 98.045% for severity classification. Six-fold cross-validation yielded average accuracies of 97.738% and 97.403%, respectively. Analysis of cross-entropy loss curves confirmed a good model fit, with no overfitting. The main contribution of this research lies in the use of a smartphone to capture acoustic signals, providing an accessible and practical solution, and in delivering a complete diagnosis with automatic identification of both the fault and its severity level. The results demonstrate the effectiveness of the proposed approach, combining low cost, high performance, and strong potential for application in predictive maintenance systems for industrial rotating equipment.</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

Intelligent diagnosis of operating condition and severity level in worm gear reducers based on sound analysis and neural networks

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Raul Bernardo de Pontes Pires,
  • João Manoel de Oliveira Neto,
  • Andersson Guimarães Oliveira,
  • Marcelo Cavalcanti Rodrigues

摘要

This work presents the development of an intelligent diagnostic system based on sound analysis and artificial neural networks, aimed at fault identification and severity level classification in worm gear reducers. The following faults were reproduced in a dedicated experimental test bench: angular misalignment, overload, and inefficient lubrication. Sound signals were captured using a smartphone, providing a practical and low-cost solution. Signal analysis was conducted in both time and time–frequency domains, with the extraction of statistical metrics (kurtosis, RMS, and crest factor) and the energy of Discrete Wavelet Transform (DWT) coefficients, using 11 levels of decomposition. Input variables for the neural networks were selected based on ANOVA and Random Forest tests, ensuring model robustness and interpretability. Two independent Multilayer Perceptron (MLP) networks were implemented: one to classify the operating condition (normal, misalignment, overload, or inefficient lubrication), and another to classify severity levels (mild, moderate, or severe). The final architecture used 15 neurons in the hidden layer, with ReLU activation and a Softmax function in the output layer. Accuracy reached 98.889% for operating condition classification and 98.045% for severity classification. Six-fold cross-validation yielded average accuracies of 97.738% and 97.403%, respectively. Analysis of cross-entropy loss curves confirmed a good model fit, with no overfitting. The main contribution of this research lies in the use of a smartphone to capture acoustic signals, providing an accessible and practical solution, and in delivering a complete diagnosis with automatic identification of both the fault and its severity level. The results demonstrate the effectiveness of the proposed approach, combining low cost, high performance, and strong potential for application in predictive maintenance systems for industrial rotating equipment.