Effective scrap reduction: draw-in measurement in combination with advanced MLbased data analysis

One major aim in the field of production technology is the scrap reduction. This requires comprehensive monitoring of both the processes and the resulting product quality. For this purpose, process data, which is sensitive for the result of the manufacturing process, must be evaluated and monitored.

In order to do so, in the framework of FormPlanet project, Fraunhofer IWU assessed different in-process data acquisition systems in the context of sheet metal forming. Within the project Work Package “Part quality assessment and in-process measurements”, one of the approaches investigated in the context of deep drawing was the draw-in measurement using laser triangulation systems. This method seemed to be valid for predicting the material state and the resulting part quality. This blog post describes the underlying conditions, the methodology and the analysis of our study.

Conditions

The results of the drawing process, which is a representative process for forming technology, are mainly determined by the following parameters:

• The sheet metal material with its mechanical properties and blank size
• The forming tool
• The press, in particular the drawing device
• The tribological system influenced by the three parameters above

The sheet metal material itself is subject to fluctuations of the mechanical properties due to changes in the coil or variations from batch to batch. During production, the reciprocal effects of these fluctuations result in components with varying properties, or even lead to production of defective components.

Ultimately, all of the mentioned influencing variables manifest themselves in changing characteristics of the draw-in of the sheet metal during deep drawing. This draw-in can therefore be assumed to reflect the quality of the production process for an individual component or for entire series. Thus, in-line detection of this variable provides an opportunity to apply process monitoring or process control to deep drawing operations.

Methodology

For analysing the draw-in measurement, tests were carried out with material provided in OK and in not OK batches. An oil pan component was used as a demonstrator component. The draw-in measurement was realized by using laser triangulation sensors. The informative value of the draw-in measurement primarily depends on the sensor position. FE simulations were applied in order to identify the optimum sensor positions.

For this purpose, material cards for the OK and not OK materials were created and comparative simulations were performed. The sensors were mounted in the areas with the highest difference in draw-in. In the next step, the drawing tool was modified, and the sensors were installed. The experimental investigations were realised on a hydraulic press PYZ 250 with 2.500 k.

Figure 1: Determination of sensor positions – left) simulated draw-in for OK and not OK material, right) difference between simulated draw-in curves and resulting sensor positions

Figure 2: Integration of laser sensors in forming tool

Data analysis: Machine Learning and data interpretation

Due to the small deviations in draw-in behaviour, no correlation could be determined between the measured signal and the resulting part quality by using conventional methods of data analysis. At this point, Machine Learning (ML) techniques were involved in data interpreting.

The first step comprised pre-processing of the measured data. For this purpose, features were extracted from the draw-in curves. These features respectively the feature vector form the input for the ML models. The used dataset was split into a training dataset (80% of the data) and a validation dataset (20% of the data). Two different kinds of ML algorithms were investigated. At first the kNN (k-nearest neighbors algorithm) – a non-parametric method was used, followed by the Random Forest – an ensemble learning method. Using the kNN, a correct classification rate of 80% was realised on the validation dataset, and the Random Forest method resulted in a correct classification rate of 87%.

Having the small amount of training datasets (80) in mind, the results are promising. It is to be expected that the correct classification rate can be significantly increased by increasing the training data.

Results

A promising approach for the prediction of part quality lies in draw-in measurements in combination with advanced data analysis methods based on Machine Learning.

Due to its real-time capability, this system can be used as an in-process quality assessment method for efficiently reducing scrap. The findings of our study form the basis for the development of closed loop controls.

---

---