Neural Network Evaluates Auto Hi-Fi Components

A version of this article ran in the April-May 1999 edition of Test & Measurement Europe.
Download the article as a PDF.

By substituting a neural network for human evaluation in production-line tests, you can speed up production and eliminate discrepancies in quality that result when two human operators’ judgement of acceptable quality level differs. In contrast with other methods of automating such tests, neural networks let you eliminate the need to develop complex mathematical models of characteristics such as acoustic noise.

Philips Components has put neural networks to work replacing the human ear in a production-line automated test of acoustic noise generated by a car tape-deck chassis. According to Ralf Hofmeier, an engineer at the Automotive Playback Modules division of Philips Components (Wetzlar, Germany), Philips’ quality-control staff member traditionally have manually determined whether moving gearwheels inside an automotive tape-deck chassis generated unacceptable levels of acoustic noise. Quality results thus depended on the worker performing the test as well as on the unavoidably varying levels of background noise in the production-test environment.

In conjunction with Management Intelligenter Technologien GmbH (MIT, Aachen, Germany), Philips substituted a Kohonen (see Footnote 1) neural network classifier for the human ear in production-line tests of its automotive tape decks. According to Manfred Fochem, an engineer at MIT, such networks are robust in the presence of input signals infested with electrical noise and in the presence of high background acoustic-noise levels. In addition, he notes, the networks exhibit a capacity for abstraction and generalization. Consequently, the networks can classify patterns that do not correspond precisely to learned patterns.

Data Acquisition
The Philips approach employs a piezoelectric acoustic sensor applied to the bottom of the tape-deck chassis. As the tape-deck motor drives the gearwheels, the sensor output is preamplified and digitized at 32 kHz using a National Instruments AT-DSP2200 data-acquisition board. Next, National Instruments’ LabView software extracts the fast Fourier transform (FFT) and cepstrum from the acquired time-domain signal. The cepstrum (whose first four letters are the reverse of the first four letters of “spectrum”) is the inverse Fourier transform of logarithmic FFT:

C(q)=|F^–1{log[|F(t)|²]}|²

Here, F(t) is the FFT of the original time-domain signal, and q represents quefrency (the cepstrum domain’s inverse of frequency, measured in units of time).       

Figure 1 shows typical time-domain and cepstrum signals (the FFT is not shown) derived from sample tape decks. Parts (a) through (d) represent decks that human experts judge to be unacceptable, while part (e) represents a satisfactory deck.

Fochem points out that the time-domain signals of some faulty decks exhibit obvious fault-specific properties, such as in Figure 1d. Some fault properties are more subtle in the time domain but are more noticeable in the cepstra (Fig. 1b).

Figure 1. Time-domain and cepstrum signals such as the ones shown here representing four noisy tape decks (a through d) and one good one (e) serve as a basis for training a neural network. In some cases (d), the noise characteristic is readily apparent in the time domain; in others (b), it’s less obvious but can be extracted from the cepstrum. (Courtesy MIT GmbH.)

The cepstrum function’s benefit, in contrast with the FFT, is the cepstrum’s ability to report total power content in a series of harmonics. Faulty bearings or gears might manifest themselves in the frequency spectrum as a series of unobtrusive harmonics, each of which barely rises above the noise floor. In the cepstrum domain, however, the power content of the entire harmonic series is encapsulated in one very noticeable rahmonic—the cepstrum domain’s analog of the frequency domain’s harmonics (see Footnote 2). Figure 1b, representing a faulty tape deck, shows at least eight rahmonics, compared with five in Figure 1e, which represents a good deck.

The Philips and MIT GmbH team employ the time-domain signal and the FFT as well as the cepstrum as input to its neural-network classification system ( Fig. 2). MIT GmbH engineers employed their firm’s DataEngine V.i software, an add-on program to LabView that provides neural-network tools within the Labview environment, to implement the Kohonen neural-network itself.  

Figure 2. A data-acquisition board gets data from an acoustic sensor mounted on a tape-deck chassis, and LabView software derives FFT and cepstrum information from the acquired data. The Feature blocks in the neural network extract information from time, frequency, and cepstrum data. The go/no-go block yields a pass/fail decision and writes results to a database. The block labeled T+L represents neural-network training and labeling. (Courtesy MIT GmbH.)

Training the Network
To train the network, the team took acoustic signals from 500 sample tape decks that had been judged by mechanical-noise-analysis experts. The time-domain, FFT, and cepstrum data files plus the experts’ corresponding judgements served as input for the Kohonen neural network’s unsupervised learning process, during which the network develops two-dimensional self-organizing neuron configurations, or “feature maps” to use Kohonen’s terminology.

As the learning process progresses, vectors associated with each neuron adapt, effectively changing the neuron’s location on the feature map. These vectors represent weights corresponding to the degree to which the associated neuron is related to the input signals. When training is complete, similar input signals excite localized groups of neurons, leading to consistent outputs.

In the Philips production test system, the neural network presents a good/bad classification result as green/red semaphore. Detailed fault-class information (for instance, which data representation in Figure 1 most closely matches that of a failed production unit) gets written to a database. Philips engineers needn’t make hardware changes to adapt the test to new tape-deck designs; they need only retrain the network.

Fochem and Hofmeier conclude that neural networks can serve in quality control whenever cost requirements or system complexity forbid the mathematical modeling of acoustical behavior of components under test. Hofmeier notes that the system reaches Measurement System Analysis (MSA) quality levels defined by Chrysler, Ford, and General Motors and suggests that neural networks can be at the core of MSA-approved quality-control systems for automotive components. T&MW

FOOTNOTES
1. Kohonen, T., Self-Organizing Maps, Second Extended Edition, Springer Series in Information Sciences, Vol. 30, Springer, New York, 1997, 1-800-SPRINGER, www.springer-ny.com/ordernew.html.

2. Harris, Cyril M. (editor), Shock and Vibration Handbook, Third Edition, McGraw-Hill Book Co., New York, NY, 1988, pp. 13-43 through 13-45, and 16-16.

FOR FURTHER READING
Fochem, Manfred, Peter Wischnewski, and Ralf Hofmeier, “Quality Control Systems on the Production Line of Tape Deck Chassis Using Self Organizing Feature Maps,” ESIT 97, the First European Symposium on Applications of Intelligent Technologies, Aachen, Germany, 1997, www.mitgmbh.de.