Subscribe Button 1

Predictive Quality – What It Is and Who Can Benefit

Quality controls of components are essential for manufacturing companies. Depending on the type of component, the checks lead to high non-productive times . 100% controls are therefore often not economical. In addition, quality controls can usually only be carried out after the manufacturing process. This means that reacting to errors is only possible at the end of the entire production process – when the defective components have already been produced.

One possible solution: Predictive Quality. With this approach, the quality of components is predicted in parallel with the process using AI models – based on production data acquired at high frequency. This enables 100% checks, reduces non-productive times and detects process errors at an early stage.

Predictive Quality – What Is It?

When we talk about ‘predictive quality’ we mean the approach of predicting component quality in parallel with the process. The resulting component quality is derived from current production data, for example using an artificial neural network (ANN) .

The network learns connections between the production data and the required quality characteristics of manufactured components. The network can be used to predict quality characteristics for new production data . In the area of ​​machining, this can be form deviations, for example.

Which Data is Suitable For Predictive Quality?

For process-parallel quality control, two types of data is needed: production data and data for assessing the component quality. The drive currents of the machine axes are particularly suitable as production data in the area of ​​cutting machine tools. These are already available as machine-internal signals in the machine control. Various methods can be used to access this data.

Most control manufacturers offer their own interfaces for acquiring machine signals. Real-time capable fieldbus systems such as Profibus or Profinet offer the possibility of recording machine-internal signals at high frequencies. In addition, open standards such as OPC UA or MTConnect are also becoming increasingly popular. The latter can usually be implemented with less effort, but have a lower sampling frequency.

Before choosing the interface, you should answer these questions:

– What interfaces does the control offer?
– What data do I need? Can this data be accessed via the respective interface?
– What sampling rate do I need to evaluate the data?
– How many different data streams do I want to transfer?
– How quickly do I want to be able to react to the transferred data?

In order for an artificial neural network (ANN) to be trained for quality prediction, quality control data must also be recorded in addition to the production data. CNC-controlled measuring machines in particular offer great potential: They record measurement data automatically and with high repeatability and make them available for further use. However, data from analog measuring equipment can also be used for predictive quality. The quality of the measuring equipment used determines the maximum achievable accuracy of the quality prediction.

How Is The Artificial Neural Network Trained?

In the next step, you train your AI with the recorded production and measurement data. The goal is to predict the quality before the measurement and thus be able to react more quickly to quality changes or to make quality control unnecessary in the long term. Artificial neural networks can also learn complex relationships based on a variety of data. Production data represents the input variables and the quality control data represents the output variables of the neural network.

Spindle torque of a milling process to predict shape deviations and the surface quality of the milled component

The neural network is trained, for example, using ‘supervised learning. Provide the ANN with the input and output variables to teach it the correlations between the inputs and the corresponding outputs. These correlations are stored in the parameters of the network. The trained AI model is then able to predict the associated quality characteristics of the manufactured components based on new production data.

Predictive Quality – Which Companies Can Benefit From It?

Quality controls of manufactured components are an indispensable step for all manufacturing companies. In particular, productions in which manufacturing processes are frequently recurring (for example in series production) benefit from predictive quality. However, the principle can basically be implemented in all manufacturing processes, for example to predict shape and position errors in the manufactured components. The approach is less suitable for small batch sizes of frequently changing products. However, Predictive Quality can also be used across products for similar or identical form elements.

Predictions made by neural networks do not achieve 100 percent accuracy. They cannot therefore completely replace quality controls. Nevertheless, the tolerance of the predicted quality variables can be quantified using the ascertainable accuracy of the network. This tolerance helps in assessing whether the ANN meets the quality control requirements or not.

Author: Aleks Arzer Research Assistant – Institute of Manufacturing Technology and Machine Tools

For more information: