ABSTRACT
Figure 8 shows a general layout o f a typical set up for automated acoustic inspec tion. A microphone or any kind o f vibration sensor is used for acquisition o f sig nals, generated by systems under inspection. The acquired signals have first to be smoothed or filtered to reduce the noise components present in the signals, and then to be digitized and stored for further processing needed for pattern recogni tion and pattern classification. Based on this, the inspection set up should identify whether the sample under test is O K or N O T O K . In case o f N O T O K statement, it should further specify the quality class degradation to which the sample be longs. In this way the set up w i l l finally:
• Increase the inspection and classification reliability • Enable a higher inspection rate
When designing a computer-based setup for automated acoustic inspection, care should be taken that it:
• Needs a relatively simple training form (that is, attributes o f simple training with an acceptable learning efficiency based on a relatively small number o f training samples)
• Guarantees a high degree o f adaptability to the new test experiment • Requires only a simple procedure when operating in the field • Has a relatively low prime cost
Currently, various setups that meet the majority o f the criteria listed above are commercially available. Furthermore, the professional setups offer some addi tional features important for:
• Establishment and management o f inspection data files • Automated specific access to the databases
• Statistical analysis o f inspection data formulation and on-line adaptation o f quality criteria to meet the current product test and current status o f the production line
• On-line training o f automated device on new products • Possibility o f direct intervention in product quality control via process
interfaces and related control software
Methods o f signal analysis used in acoustic inspection predominantly rely on analysis o f frequency spectrum, containing the signals' characteristic features to be extracted. The extracted features, usually the amplitude values o f around 250 frequencies, selected as components o f the signal features vector, are then for warded to a pattern classifier, implementing a conventional pattern recognition algorithm, or to a neural network as a cognitive pattern recognizer [43]. In this way the automatic diagnosis o f system status is possible. For instance, using neu ral networks for implementation o f a feature extractor and pattern classifier i t was possible to identify the class o f internal errors o f a combustion engine based on the explosion sound records o f the engine under operation [99,100].
