Fan Fault Diagnosis Using Acoustic Emission and Deep Learning Methods

Giuseppe Ciaburro, Sankar Padmanabhan, Yassine Maleh, Virginia Puyana-RomeroInformatics 2023 · DOI

TL;DR

This paper introduces a novel method for automatically detecting dust accumulation on axial fan blades using acoustic emission and deep learning. By analyzing sound spectrograms with a pre-trained Convolutional Neural Network (CNN) and transfer learning, the system achieves high accuracy in distinguishing between healthy and faulty fan conditions, paving the way for more efficient and proactive industrial maintenance.

What Problem Does This Paper Address?

Maintaining industrial equipment in optimal working condition is crucial for efficiency, safety, and energy conservation. Fans, commonly used in ventilation systems, are susceptible to dust accumulation on their blades. This accumulation not only affects their aerodynamic performance and energy efficiency but can also lead to the spread of allergens and microorganisms, posing health risks. Traditional maintenance often relies on scheduled checks or reactive repairs after a failure. The challenge lies in developing an automated, non-intrusive method to detect this specific type of fault (dust buildup) in real-time, enabling timely interventions and preventing production downtime.

Methods

The core of this research lies in a two-pronged approach: data acquisition and deep learning-based analysis.

  1. Acoustic Emission Measurement: The study utilizes acoustic emission signals generated by an axial fan. Measurements were taken in an anechoic chamber to minimize external noise interference. Two distinct operating conditions were simulated: "No-Fault" (clean blades) and "Fault" (artificially created dust deposits on blades).
  2. Spectrogram Generation: The raw acoustic signals were transformed into spectrograms. A spectrogram is a visual representation of the signal's frequency content over time, with color intensity indicating sound intensity. This conversion is achieved using the Short-Time Fourier Transform (STFT).
  3. Deep Learning with Transfer Learning: A Convolutional Neural Network (CNN) was employed for fault classification. To overcome the challenge of limited data, the researchers utilized transfer learning. They adapted a pre-trained CNN model (SqueezeNet), originally trained on the massive ImageNet dataset for general image recognition, to their specific task. This involved modifying the input layer to accept spectrogram images and retraining the final classification layers to distinguish between the "No-Fault" and "Fault" states.
  4. Data Augmentation: To further increase the dataset size and improve model robustness, various transformations (rotation, resizing, reflection) were applied to the spectrograms.

Datasets

The study did not rely on publicly available datasets for fan fault diagnosis. Instead, they created their own dataset by:

  • Recording acoustic emissions from an axial fan under two controlled conditions: "No-Fault" and "Fault."
  • The "Fault" condition was simulated by artificially applying fine-grained sand to the fan blades.
  • After initial data collection and segmentation, data augmentation techniques were applied, increasing the dataset from 120 samples to 840 samples.
  • The augmented dataset was split into a training set (70%) and a test set (30%).

Key Insights and Ideas

  • Acoustic Signatures of Dust: The research demonstrates that dust accumulation on fan blades significantly alters their acoustic emission profile, creating a distinct "signature" that can be detected. This is attributed to changes in aerodynamics and potential imbalances.
  • Spectrograms as Rich Features: Spectrograms effectively capture the complex interplay of time, frequency, and intensity in acoustic signals, making them highly informative for deep learning models. The visual patterns within spectrograms can reveal subtle differences indicative of faults.
  • Power of Transfer Learning: Leveraging a pre-trained model like SqueezeNet, originally designed for image recognition, proved highly effective. This approach significantly reduces the need for massive, task-specific datasets and accelerates the training process, making deep learning more accessible for specialized industrial applications.
  • Non-Intrusive Monitoring: The method is non-intrusive, relying solely on sound measurements. This is a significant advantage over vibration sensors or visual inspections, which might require system shutdown or physical access.

Key Results

The developed methodology achieved an impressive accuracy of 0.95 in classifying the fan's operating conditions (No-Fault vs. Fault) on the test set. This high accuracy confirms the effectiveness of using acoustic emission and deep learning for detecting dust deposits on fan blades. The study also noted that the 270° microphone position provided the most sensitive readings for detecting operational irregularities.

Limitations and Future Work

  • Artificial Fault Simulation: The "Fault" condition was simulated using artificial dust. While this provides a controlled environment, real-world dust accumulation can be more complex and varied.
  • Specific Fan Type: The study focused on an axial fan. Further research could explore the applicability of this method to other types of fans (e.g., centrifugal) or different industrial machinery.
  • Generalizability: While transfer learning helps, the model's performance might vary with different fan designs, operating speeds, or environmental conditions not covered in the study.

The authors suggest that this methodology can be integrated into automated maintenance systems for fans in industrial settings, particularly those dealing with dusty environments or where physical access is challenging. Future work could involve testing the system in real-world industrial scenarios and exploring its potential for diagnosing other types of fan faults.

Summary updated

BibTeX
@article{ciaburro2023fan,
  author = {Giuseppe Ciaburro and Sankar Padmanabhan and Yassine Maleh and Virginia Puyana-Romero},
  title = {{Fan Fault Diagnosis Using Acoustic Emission and Deep Learning Methods}},
  journal = {Informatics},
  year = {2023},
  doi = {10.3390/informatics10010024},
}
Figure