MVDR-LSTM Distance Estimation Model Based on Diagonal Double Rectangular Array

Xiong Zhang, Wenbo Wu, Jialu Li, Fan Dong, Shuting Wan2023 · DOI

TL;DR

This paper introduces a novel method for estimating the distance of roller faults in belt conveyors using a combination of a specialized microphone array and deep learning. By employing a diagonal double rectangular microphone array with MVDR beamforming and an LSTM network, the model achieves high accuracy in fault distance estimation, even in noisy industrial environments, outperforming traditional methods.

What Problem Does This Paper Address?

Belt conveyors are critical components in many industrial settings, including straw power plants. The smooth operation of these systems heavily relies on the health of their idlers. Idler failures can trigger a chain reaction, posing safety risks. Traditional inspection and maintenance methods for idlers are time-consuming, costly, and can lead to overdue replacements of degraded parts. There's a need for non-contact acoustic methods to accurately detect and locate idler faults, enabling proactive maintenance and improving operational safety and efficiency. Existing methods often struggle with the complex acoustic environments and long distances involved in industrial settings.

Methods

The core of this research lies in a two-pronged approach:

  • Diagonal Double Rectangular Microphone Array: This novel array configuration is designed to improve sound source localization and reduce noise and echo interference. It consists of two staggered rectangles, allowing for better capture of sound direction information and enhanced spatial filtering. The authors argue this configuration offers superior robustness compared to simpler linear, cross-shaped, circular, or rectangular arrays for this specific application.
  • MVDR-LSTM Model:
    • MVDR (Minimum Variance Distortionless Response): This adaptive beamforming algorithm is used to process the acoustic signals captured by the microphone array. MVDR excels at suppressing noise and interference while preserving the desired signal, leading to improved signal-to-noise ratio and spatial resolution. It adaptively adjusts to changing environmental conditions.
  • LSTM (Long Short-Term Memory): This deep learning architecture is employed to analyze the processed acoustic features. LSTMs are well-suited for time-series data and can learn long-term dependencies, making them effective for identifying patterns in the acoustic signals that indicate fault distances. The model uses a two-hidden-layer LSTM architecture.

The process involves acquiring audio signals with the microphone array, applying MVDR beamforming to extract relevant features, and then feeding these features into the LSTM model for fault distance estimation.

Datasets

The paper describes experimental setups and data collection rather than using pre-existing public datasets. The authors collected their own audio data from a simulated belt conveyor system with controlled fault sound sources and background noise. They generated datasets for training and testing the MVDR-LSTM model, splitting them into 90% for training and 10% for testing.

Key Insights and Ideas

  • The Power of Specialized Arrays: The diagonal double rectangular array isn't just a minor tweak; it's presented as a crucial element for achieving better spatial resolution and noise suppression than other common array configurations, directly impacting the accuracy of fault localization.
  • MVDR's Noise Resilience: The MVDR algorithm's adaptive nature and its ability to minimize noise and interference are key to extracting meaningful fault signatures from the noisy industrial environment.
  • LSTM's Temporal Memory: The LSTM's capacity to "remember" past acoustic patterns is vital for distinguishing subtle fault signatures from background noise and accurately estimating the distance of the fault.
  • Synergy is Key: The combination of MVDR for signal enhancement and LSTM for pattern recognition proves more effective than either technique alone or simpler beamforming methods paired with LSTM.

Key Results

  • High Accuracy: The MVDR-LSTM model achieved very high accuracy in fault distance estimation, reaching 100% in some experiments under background noise interference.
  • Superior Performance: Compared to CBF-LSTM and FBF-LSTM models, the MVDR-LSTM model demonstrated significantly higher average accuracy (e.g., 99.74% vs. 94.60% and 74.69% for the diagonal double rectangular array under background noise).
  • Robustness to Noise: The model showed strong performance even under challenging conditions like white noise and irregular impact noise, with accuracy converging to high levels after a reasonable number of iterations.
  • Improved Localization: The diagonal double rectangular array, when used with the MVDR-LSTM model, provided better localization accuracy and side-lobe suppression compared to linear, cross-shaped, circular, and rectangular arrays.

Limitations and Future Work

  • Computational Complexity: The authors acknowledge that LSTM models, while powerful, have higher computational complexity than CNNs due to their gating mechanisms. This can lead to longer training times and higher resource requirements, which they suggest optimizing in future research.
  • Model Structure Optimization: While the paper settled on a two-hidden-layer LSTM, they noted that increasing to three layers increased computation time, implying that further fine-tuning of the model architecture could be beneficial.
  • Generalizability: While the paper suggests potential applications in other industrial testing fields, further validation across diverse scenarios would be valuable.
  • Real-World Deployment: The study was conducted with simulated fault conditions and controlled noise. Testing and validation in actual, dynamic industrial environments would be a crucial next step.

Summary updated

BibTeX
@misc{zhang2023mvdr,
  author = {Xiong Zhang and Wenbo Wu and Jialu Li and Fan Dong and Shuting Wan},
  title = {{MVDR-LSTM Distance Estimation Model Based on Diagonal Double Rectangular Array}},
  year = {2023},
  doi = {10.3390/s23115094},
}
Figure