Convolutional Neural Network For Predicting Welding Quality

click to learn more

• Chinese

PREREQUISITES

• numpy
• tensorflow
• imblearn

OPTIONAL

• tensorflow-gpu
• sklearn
• matplotlib

Introduction

According to the curve waveform of voltage, current and electrode position, judge the quality of flash welding. The data is multi-dimensional time series. We have 2000 of good quality and 50 of bad quality in ./data/. In this network, I use data augmentation to increase the number of bad. Experiments show that CNN is more effective than BP-network and Dropout is effective. I think convolution can identify the relative positional relationship between multi-dimensional time series, which reduces the over-fitting of the model to some extent. As shown below, the origin data is multi-dimensional time series.

START

• git clone git@github.com:wzx140/welding_prediction.git
• conda install --yes --file requirements.txt. Import dependency to anaconda
• change some variables in resource/config.py
• cd welding_prediction
• python main.py train to train the model and save the model in resource/model. There is a trained model in this folder
• tensorboard --logdir resource/tsb, to see the visualization of the data after training
• python main.py predict path-to-mode path-to-sample to predict the quality of the welding

Features

Regularization

Implemented Dropout

• set keep_prob in resource/config.py range from 0~1. 1 means Dropout is disabled

DATA AUGMENTATION

Since we only have 50 bad samples, we use ADASYN to expand the bad samples.

For more information, you can read my blog about ADASYN

TENSORBOARD

After training, the data for tensorboard will store in resource/tsb. Just run tensorboard --logdir resource/tsb

DEBUG

If you want to use tfdbg, you should,

• install pyreadline by pip
• set enable_debug True in resource/config.py
• run python main.py train --debug in project dir

For more information, you can read official document

DEMO

• see the demo
• we implement more model like DTW and KNN to classify our data and compared with our neural network model

MORE

• CNN model's test accuracy is more than 0.96 and train accuracy is more than 0.99
• the best network architecture is (18 - 36 - 72 - 144)