Standard Profil makes sealing profiles for automotive manufacturers all around the world and it is one of STREAM-0D end-users. In this article Cetri, that in the project leads data fusion, data sparsification and data-driven modelling, provides a glimpse of the activities carried out on Standard Profil’s production line.
The development of the data-driven models (DDMs) for Standard Profil (SP) requires the completion of the following two tasks:
- Extraction of the critical dimensions from rubber profile images and
- building machine learning models for prediction of potential product defects where machine parameters are used as inputs.
The extraction of critical dimensions from rubber profile images is required to determine whether a product is defected or not, based on the specifications of the critical dimensions of the given product. This enables Cetri to label the desired outputs as 1 for ‘defected’ and 0 for ‘not defected’ items. The labelling process is a necessary step for the development of the DDMs: following the supervised learning approach, the objective is to infer an I/O mapping (this work is in progress).
To do so, Cetri developed a multi-step process using image processing techniques to analyse the provided images. The images to be analysed are pictures of the profiles of the rubber taken every hour using the existing profile scanner system of SP. Specifically, Cetri follows two approaches to address the problem:
Approach 1: Detect interest points and estimate the distances between them. The main advantage of this method is that it is invariant to rotation and translation. This is important because the pictures show rubber profiles with different orientations and in different positions. As a result, technicians should rotate and center, in principle, all profiles images to make them aligned.
A multi-step automated process has been developed to remove noise and detect interest points in the images, ensuring dimensionality accuracy and robust preservation of details. The process is repeatable for different samples and conditions. Cetri uses blurring and thresholding techniques that provide solid removal of the background and shadows/distortions from the scanning equipment (see top image). This technique enables measurements via region of interest (ROI) generation. ROI selection is performed using the Canny edge detection method. The interest points are detected using the SURF algorithm.
Cetri is facing some issues in finding a common set of interest points by matching them between images. A lot of the interest points capture features that are repetitive (i.e., not unique) in the images due to the fact that the profiles are not information rich.
Approach 2: Align profile images via rotation and translation. The idea here is that once all of the images are aligned, dimensions can be extracted following the lines specified by the product specification.
Initial processing steps including blurring, thresholding and edge detection are performed (same as in Approach 1). Implementation is carried out to extract global properties of the profile images, such as histograms of edge orientations that enable Cetri to align the profiles between them.
The actual dimensions of the surface in the background of the profile images are a requirement for this approach. This is needed as a reference point for measuring distances from images.
Once the process of automatically transforming images to critical dimensions is finalised, Cetri categorises the samples to ‘defected’ and ‘non-defected’. This allows to associate machine parameters to ‘desired’ outputs; defected or not, using binary classification methods. A list of input features and outputs are presented in image 2:
Cetri developed a multi-step data cleaning process to ensure the high quality of the data before building the models.
Firstly, critical lengths of the cross-section from images are extracted. variables of interest used as input features or target variables are then selected. Subsequently, entries with missing values in any of the parameters of interest are removed, making sure that no duplicates are present in the data. Outliers are detected and then removed. Lastly, the data is normalised (standardization, 0 mean, 1 standard deviation), so it is compatible with the machine learning algorithms used to train the models.
Regression and classification models are to be developed; one model for simultaneous outputs and separate models for singular outputs. Cetri is planning to use random forests and deep neural networks as model architectures. The overall design of the development of data-driven models can be summarised in the following image.
This article was provided by CETRI, STREAM-0D consortium partner
Follow STREAM-0D on: