Neural Networks : WS23/24 Project

Introduction : The Acoustic Weld Inspection Project (AKoS)

• Submitted by Shashwat Pandey (spande2s), Talha Riyaz Shaikh (tshaik2s)

• Source : https://www.akos-projekt.de/

The Acoustic Weld Inspection Project (AKoS) enhances weld seam quality assurance for safety-critical components using innovative machine learning, particularly autoencoders. It addresses limitations in traditional methods like visual examination and ultrasonic testing. AKoS leverages acoustic data to develop a robust system for accurate anomaly detection in weld seams. It aims to develop a non-destructive testing (NDT) method for welding seams in safety-critical components.

Solution Overview

The proposed solution leverages autoencoders, a type of neural network designed for unsupervised learning. Autoencoders consist of an encoder and a decoder, with the primary objective of learning a compressed representation of input data. In the context of AKoS, this compressed representation captures essential features of acoustic signals associated with normal weld seams The proposed solution utilizes autoencoders, a neural network architecture tailored for unsupervised learning, to encapsulate crucial characteristics within acoustic signals associated with normal weld seams in AKoS. The workflow involves:

1. Data Understanding:

   • Analyzing the frequency representation of acoustic signals through Power and Mel spectrograms for only the files associated with "Microphone Gefell 1 = MK301, Amplification "32"

2. Feature Extraction:

   • Mel Spectrogram: A frequency-domain representation of the signal's spectrum, emphasizing the distribution of energy across different frequency bands.
   • Pore Amplitude Feature: Captures amplitude information within specified frequency ranges associated with pores or defects in weld seams.
   • Spectral Contrast Feature: Measures the difference in amplitude between peaks and valleys in the spectrum, with statistical descriptors (mean and standard deviation) computed for added context.
   • Statistical Descriptors: Computes mean and standard deviation for spectral contrast
   • Feature Vector: Combines mel spectrogram, pore amplitude, and statistical descriptors
   • Length Adjustment: Zero-pads or truncates feature vectors for consistent length
   • Output: Returns a matrix with the extracted features

3. Scaling and Feature Agglomeration Application:

   • Applying scaling and feature agglomeration techniques to enhance data representation.

4. Autoencoder Training (Normal Data - Non-Pore):

   • Training the autoencoder on only the normal data to capture inherent patterns.

5. Testing and Evaluation:

   • Assessing the model's performance based on reconstruction errors
   • First take a test signal, find reconstruction error compare it with the reconstruction error of known signal (non-pore).
   • Check if the reconstruction error is high, it's understandable that this signal is not known to the model so its an anomaly or in this case ( a pore)

6. Classification Comparison:

   • Comparing reconstruction errors for Pore and Non-Pore classification.

7. Further Investigations

• Despite achieving a 95.83% accuracy, further investigations are suggested.
• These include exploring different microphones and frequencies.
• Extended analysis of pore distribution and audible cues from pores is recommended.
• Applying existing techniques for pore recognition can enhance the model's capabilities.
• These investigations aim to improve weld inspection comprehensively.

Conclusion

• The Acoustic Weld Inspection Project uses autoencoders and advanced machine learning to improve weld seam inspection.
• Ongoing research aims to establish it as a reliable tool for quality assurance in welding applications.
• The project shows promising results for defect detection in weld seams, offering a reliable method for safety-critical components.
• Further refinement and enhancement of the approach are underway for more robust weld inspection.

References:

• Source : https://www.akos-projekt.de/

1. https://github.com/aldente0630/sound-anomaly-detection-with-autoencoders/tree/main -

   • We got the basic idea behind this anomaly detection from here.
   • We also tried to understand their feature extraction also which was very helpful for the purpose of this project
   • The autoencoder architecture used by them was useful to figure out the right bottleneck layer

2. https://chat.openai.com/ -

   • How to effectively improve accuracy when the data was overfitting and what strategies are possible to tune the hyperparameters
   • For drafting some parts of this report on Markdown and using proper scientific terms

3. 0_read_i32.py -

   • Original file used to read the audio signals, we referenced this to read the files needed to make our Normal data

4. Anomalous Sound Detection using unsupervised and semi-supervised autoencoders and gammatone audio representation - https://arxiv.org/abs/2006.15321 -

   • Understanding the architecture and how to build autoencoders for unsupervised learning

5. Deep Autoencoders for Acoustic Anomaly Detection: Experiments with Working Machine and In-Vehicle Audio - https://repositorium.sdum.uminho.pt/bitstream/1822/81433/1/aeaad2.pdf

   • Formulating the algorithm for using autoencoders in anomaly detection