電気通信システムと管理

A neural network approach for detecting surface defects in hot rolling process

Abstract

Mr. Sam Geogi

In  hot  rolled  steel  production  surface  defects  occur  due  to  material  problems  or  process  problems.  The production  quality  should  be  monitored  real  time  to  identify  the  surface  defects  occurring.  This  will  help  to easily identify the cause of the defect and  solve  it.  Thus,  by correctly identifying the defects  in real time  we could rectify the problem avoiding defective production and saving in material and process cost. The production speeds for hot rolled steel sheets will reach up to 18m/s. This makes the real time monitoring extremely difficult as  the  detection  system  must  have  a  very  high  detection  speed.  The  detection  system  also  must  have  a  high prediction  accuracy  to  conform  to  the  industrial  quality  management  standards.  The  surface  defect  detection problem is to identify defect class in a steel surface. Automatic visual inspection systems were in the industry for a while, but they were so sensitive to the environment and could only use for a particular system. Here  we  are  using  convolutional  neural  network  approach  for  identifying  the  type  of  defect.  With  the advancement  of  Deep  Learning  especially  Convolution  Neural  Network  (CNN)  the  image  classification  has become more sophisticated and accurate. CNN has the potential for high detection speed and at the same time high accuracy predictions. For training the neural network a dataset of 1800 images belonging to 6 defect classes were selected from North Eastern University-USA steel defect dataset. From experimentation we have learned that for this dataset sequential model architecture could be used and 8 layered CNN model is used. Loss function used is categorical cross-entropy. The optimizer function used is Nadam. The  images  have  been  pre-processed  using  Keras  pre-processing  to  improve  the  dataset  variability  the various parameters that have been changed during pre-processing include width shift, height shift, shear zoom, horizontal flip and rotation. After the pre-processing process using the Keras a dataset with more variability is obtained. Max pooling is done to reduce the amount of data without affecting the quality at each level and also used a fully connected layer at the end to enable classification. The model been built to 8 layers with alternate convolutional and maxpooling layers and fully connected layer at the end. The built model then has been trained with image dataset the training is done on COLAB.  The convolutional layers  used  3*3  filters.  The activation function used is Rectified Linear Unit. The input shape of the images is 200*200 gray scale. The classification activation is done by softmax function. There is 68,16,198 total trainable parameters in the model. The model is trained  up  to 500  epochs. To improve the accuracy tuning  of hyper  parameters have been carried  out various parameters  that  have  been  fine-tuned  were,  loss  function  which  is  used  to  find  the  error  from  model  output against the desired output. Then the optimiser functions were fine tuned optimizers are functions used to change the attributes of neural network such as weights to reduce the losses. The performance evaluation of the CNN model and tuning of hyper parameters were carried out to obtain maximum accuracy. The model output accuracy of 99.36% was obtained through the fine tuning. The detection speed has been reduced to microseconds. The images collected from the local industries can be used to test the validity of the model. The implementation with YOLO V3 can be done to obtain for faster image detection on real time manufacturing scenarios.