On-Line Acoustical Quality Control System for Automobile Stereo Components Using LabVIEW and Neural Networks

Author(s):
Peter Wischnewski - Data Analysis Department of MIT - Management of Intelligent Technologies GmbH

Industry:
Automotive

Products:
PXI/CompactPCI, LabVIEW

The Challenge:
Developing a high-accuracy, efficient acoustical quality control system for 100 percent on-line testing of automobile stereo components.

The Solution:
Building a LabVIEW-based automated test system using DAQ boards and Kohonen neural network analysis techniques. The Philips APM Company produces components for car stereo systems in several countries throughout the world. Most large automotive and consumer electronics manufacturers are Philips' customers.

Introduction
The Philips APM Company produces components for car stereo systems in several countries throughout the world. Most large automotive and consumer electronics manufacturers are Philips’ customers. These manufacturers demand that the components delivered meet their high quality standards regardless of where a component such as a tape deck or CD drive was assembled. End users expect to hear high fidelity stereo music in their car without disturbing noise caused by moving gearwheels inside the tape deck. To meet these customer expectations and to maintain their place in the competitive market of automotive subcontractors, Philips in Germany asked for assistance from the quality control experts at MIT. MIT responded by designing an automated test system using DAQ and LabVIEW.

Philips had found that the quality of final acoustic testing of car stereo components was dependent on the individual worker perfoming the tests during a particular shift. In addition, the accepted product quality varied with plant location. Their goal was to replace testing relying on human hearing with an automated on-line quality control system. The goal of the system was to standardize testing around the world, eliminate dependency on the hearing ability of workers, and avoid the uncertainty caused by variations in background noise level. The system task was to inspect 100 percent of the assembled products with very high accuracy. Moreover, because it is an on-line system, it would locate a defect on the production line, where it can be repaired immediately.

System Setup
MIT designed the system around a Pentium 100 MHz PC running Windows 3.11. A National Instruments AT-DSP2200 digital signal processing board (subject to change to the PCI-4452) samples the acoustical data from a piezoelectric sensing element that is pressed to the chassis of the stereo component during the test cycle. For the Normal and (Auto-) Reverse Play operating modes, data is collected for approximately four seconds at a sampling rate of 8 kHz. In addition to linearity test computations, the board processes the data in preparation for real-time classification. The heart of the LabVIEW-based analysis software is a Kohonen neural network, found in MIT’s DataEngine V.i toolbox.

Neural Networks Handle the Challenges
As usual in industrial product design, improvements or minor modifications to automobile stereo components are carried out periodically. Such changes in product design complicate the acoustical frequency analysis due to the varying frequency characteristic found in chassis vibration signals of slightly different tape decks. Quality control systems using neural net-works accommodate such frequency shifts because of their generalization and abstraction abilities. Neural networks also help find the relevant frequency com-ponents. If there is relevant information buried in the data, neural networks will discover it regardless of whether the signal of interest is acoustical, vibration, or imaging data. Neural networks are less affected by noise than other algorithms. They also save time in handling complex systems, because no mathematical model has to be derived. The threshold of human acoustical perceptibility and the subjective human feeling about acceptable noise emissions can be taken into account during network "training."Having been trained, neural networks are able to classify signal patterns that were not presented to the network before.

Even though MIT had installed several quality control and pattern recognition systems based on neural networks before, the special challenge in this project was to accomplish all of the complex system tasks within the given small budget and short time restrictions. In addition to the required high-level classification performance, the duty cycle of the pro-duction line cuts the computation time for classification to less than 1 second.

For the critical feature-extraction step, the data is analyzed in the time, frequency, and cepstral domains. The cepstrum was completely built of virtual instruments (VIs) included in the LabVIEW Advanced Analysis Library. The features have to describe the acoustical properties of the stereo components as completely as possible. The Kohonen network is trained with signal patterns of 250 preclassified components and is tested using another 250 components. The trained neural network, now called a classifier, is ready to be used in the on-line quality control system. The acoustical experts from Philips set up five classes, one class for good components and four for defective MSA-approved systems for automotive components.The class names relate directly to specific parts of a tape deck.With this "message" included in the classification results, Philips engineers have the ability to adjust the defective production item immediately and reduce the cost of a failure.

Conclusion
The system was first used in Germany and Hungary and will be installed in Mexico in 1998. MIT is proud of the fact that this system incorporating DataEngine V.i. reaches MSA-standard levels. Measurement Systems Analysis (MSA) provides a standard approach to analyzing measurement systems according to QS-9000. Neural networks from the DataEngine V.i toolbox used in LabVIEW applications can be the foundation of MSA-approved systems for automotive components.

Neural networks are a favorite solution for quality control tasks whenever cost requirements or system complexity prevent mathematical modeling of acoustical behavior of the components under test. Especially companies with world-wide production sites for the same product profit from applying neural networks in quality control systems because of their generalization property and ease of use. The introduction of these algorithms into quality control systems dramatically decreases system costs even in fields where quality standards are related to human hearing.

For more information, contact:

Peter Wischnewski

MIT GmbH

Promenade 9,52076 Aachen, Germany

Tel: +49-2408-94580

Fax: 49-2408-94582

E-mail: pw@mitgmbh.de

Browse All Case Studies »