top of page
Machine Learning based Quality Assurance and Feature Prediction
Keywords
Machine Learning, Image Processing, Time Series, PCA, k-means, SVM, Deep Learning, Neural Network, CNN, VAE, ResNet, LSTM
Tasks
The research institute PtU aims to investigate quality control using machine learning techniques to improve the quality of deep drawn paper cup. Sensors including cameras and force sensors are installed to collect measurements during the production process. Based on these data, machine learning based quality control methods should be developed and evaluated. Sensory data including side and top camera captures and 5 dimensional data: time, displacement, velocity, acceleration, force should be processed.
Implementations
Traditional machine learning methods like PCA, k-means clustering are used, but these methods are incapable of yielding good clustering results, which is crucial for quality control.
Therefore, deep learning based methods are implemented. To cluster the image data, ResNet50 is used to process the image data and get the latent representation. These compressed, low dimension representations are then clustered using k-means clustering. Finally, good clustering results are yielded.
Clustered sound paper cups
Clustered broken paper cups
Based on the clustering result, variational autoencoder model is trained solely using the force data of the sound paper cups. In this way, the VAE model only learns to construct the force curve of sound deep drawing process. If a cup is drawn to broken, the force curve is definitely different from the sound one, as shown in picture. Thus, the broken paper cup can be removed from the production line.
small reconstruction loss when processing the force curve of the sound drawing process
large reconstruction loss when processing the force curve of the broken drawing process
Furthermore, a LSTM-autoencoder is trained with the force curve. In this way, a more precise and detailed anormaly detection procedure can be carried out. The LSTM-autoencoder model is also trained to reconstruct the force curves of the sound drawing process. When it is implemented to process the continuous force measurement, it can detect anormaly in realtime.
overview of the LSTM-Autoencoder model
continuous anormaly detection using the LSTM-Autoencoder
Presentation Powerpoint
Thesis
Final grade
1.0/1 (equivelant to 100/100)
bottom of page