Air and structure-borne sound analysis with AI in production
A test system uses machine learning to combine mathematical calculation functions with classification data sets. The result: acoustic verification of the manufacturing process of mechatronic components by body as well as airborne sound in the presence of loud ambient noise!
» Read more!
Noise analysis in the manufacturing process of products is a proven quality assurance measure. With the aid of structure-borne sound, conclusions can be drawn about the synchronization and assembly quality of moving systems. An assessment of sound quality, i.e. whether a sound is perceived as pleasant and typical or as unusual and disturbing, can only be determined on the basis of airborne sound. Test systems based on structure-borne sound are already in use in production. However, in order to detect noise patterns in airborne sound, new methods need to be established in the manufacturing process.
A new generation of acoustic test systems from GÖPEL electronic is to meet the requirements of customers who want to analyze their production acoustically based on airborne sound in addition to structure-borne sound. In the new generation, the objective analysis of products by predefined mathematical algorithms is extended by a subjective evaluation by means of psychoacoustic measurement methods and artificial intelligence. Noise in production no longer has a significant influence on the acoustic analysis due to the use of AI.
Sound analysis in the manufacturing process
Until now, it has hardly been possible to carry out suitable test procedures using airborne sound due to the loud ambient noise in production. Therefore, most process monitoring is realized via structure-borne sound. Suitable technological measures, such as decoupling the production line and coupling structure-borne sound sensors, ensure that only the vibrations generated in the test specimen are recorded and analyzed.
As a rule, this measurement method can only be used to monitor undamped structures with motor drives that allow conclusions to be drawn about their assembly or mechatronic subassemblies. Of course, it is also possible to evaluate the test object via the current consumption or the voltage curve.
This view is mainly of interest to manufacturers for the optimization of their own manufacturing processes. Often, significant errors in production can be found through this monitoring, which also influence customer requirements. However, proving the acoustic properties of uncoupled or sprung subsystems is difficult.
For example, analyzing the acoustic properties of an upholstered vehicle seat or a soft-sprung running rail of a window shutter on the basis of structure-borne sound can hardly succeed. Here, the end customer decides on the quality based on comfort and also on the noise pattern generated by the test specimen. Both determine the incentive to buy the product. A check of the test specimens with regard to these subjective characteristics must ultimately be carried out by airborne sound analysis. Until now, this test was only possible at great expense.
Test system for use in noisy environment
To meet the requirements of manufacturers and end customers, GÖPEL electronic has developed the test system CARoLINE, which is now available in the fourth generation. The extension of the signal conditioning cassettes to 8 input channels offers the possibility of a combination of body and airborne sound analyses. The front panel of the CARoLINE 4 cassette features indicators for system status and data recording. Via a headphone connection with volume control it is possible to listen to the acoustic signal of the test object during the measurement. On the rear panel are the 8 input channels for body or airborne sound sensors as well as further connections for analog input signals. A shaker can be connected for comparative measurements. Data recording is done via a PCI-DAQ card in the PC or via an autonomous, internal data acquisition with Ethernet coupling. Within the PC environment, a software driver takes over the signals and transfers this data stream to the analysis program.
The analyses of the recorded signals can, as already with the previous CARoLINE versions, be performed via standard mathematical functions. In addition, airborne sound recordings are now also possible. By analyzing the airborne sound data, the disadvantages of the structure-borne sound analysis can be compensated. The difficulty is to realize an airborne sound analysis that works reliably independent of the disturbing sound in the factory halls.
Here, GÖPEL electronic offers an alternative approach to the acoustic shielding currently required for airborne sound analysis via machine learning. The method combines the previous mathematical calculation functions with the collection of data sets for the classification of specific features and noise patterns. During production cycles, the acoustic records obtained now form a basis for feature identification, with the existing noise also being recorded. These data sets are the prerequisite of an accurate prediction of subjectively determined classifiers. Clear and distinct noise is no longer dependent on frequency response or variations in speed, but only on the selection, i.e. an annotation, of the disturbances that have occurred in the product. Detailed mathematical knowledge about the determination of the noise is therefore no longer necessary. They only have to be picked out in the recorded audio signal.
In connection with CARoLINE 4, tools are offered to support the process of machine learning. With AnEKa (AI Annotation Tool) the data can be annotated to prepare an analysis of the acoustic signals for the development of the classifiers. The tool is used to listen to the recorded DUT data and assign the fault sounds. This is done by selecting and naming areas directly in the audio signal. Professional acousticians are no longer necessary for this work.
Subsequently, the annotated data can be used with MaLen (AI Model Handling Tool) for the generation, further development and evaluation of classification models. The schematic illustration shows the steps necessary to generate a model using machine learning and then apply it in production.
- Recording of audio data from the DUT.
- Annotation and range selection of the noise errors in the data.
- Training a model using the selected data.
- Use of the classifier in CARoLINE.
Independently of this, models that have already been created can also be integrated directly into the CARoLINE environment via a dialog.
CARoLINE 4.0 combines proven mathematical methods as well as the application of AI. This enables effective acoustic verification of the manufacturing process with structure-borne and airborne sound that meets the requirements of both manufacturers and end customers.