Wavelet Neural Network For Ultrasonic Detection Of Materials With Coarse-grained

In ultrasonic detecting of coarse-grained material specimen or workpiece withlimited size, there are many kinds of noise such as boundary reflection waves andacoustic-electric crosstalk signals in received signals besides the scattered waves, flawecho and bottom echo. In this paper, first, the origin of the noise signals andde-noising method have been researched experimentally. Then, the primarycharacteristics of the scattered wave, such as the earliest time it arrives to receivingtransducer, the position of its maximum value and its frequency spectrumcharacteristic in the propagation process have been verified experimentally. Amongtheoretical analysis, we calculate the sound pressure distribution along axis usingpulse frequency spectrum composition method for transmitting and receiving complexfield. According to above calculation we give a physical explanation to the position ofthe maximum pressure value, We also use short time Fourier transform to analyzethe spectrum of the scattered wave during the propagation process, and obtain someuseful results.Because of the scattering characteristic of the coarse-grained material, theuseful signals are covered up by the scattered waves which come from themicro-structures of the material. In order to minimize the influence of the scatteredwaves on the useful signals and improve the signal-to-noise ratio (SNR), waveletneural network is applied. Wavelet neural network is a non-linear filtering methodthat can be used to reduce the noise of ultrasonic signals adaptively. Improvedgradient descent algorithm with gauss mother wavelet has been used to train theneural network. In the process of training, the network has a dynamic adaptivelearning rate, which can be updated automatically according to gradient descent. The experimental results show that wavelet neural network is an effective method inde-noising of ultrasonic signals. With improved gradient descent algorithm and gaussmother wavelet, the network has completed training in 21 times with the error0.001177. Compared with the normal gradient descent algorithm, the learning speedof the procedure described here is evident increased. It effectively avoids convergingat a local minimum. These advantages show that wavelet neural network inde-noising of ultrasonic signals is a quite useful tool in pre-processing stage in flawdetection. It has a great significance in identifying useful signals and accomplishingreal time processing in ultrasonic detection.