Research On Ultrasonic Flaw Signal Recognition By BP And NNRS Models

As one of the five common non-destructive detection method, the Ultrasonic nondestructive testing technology has the object to be measured a wide range of great depth detection; defect location accurate, high detection sensitivity; Low cost, easy to use; Speed is quick, harmless to the human body and convenient for field use and other characteristics, these make it to become the highest frequency of use and rapid development of a non-destructive testing techniques, it is the most widely used at home and abroad. The ultrasonic defect signal instabilities and nonlinear characteristics make the defect types of discrimination requires a great deal of artificial technology, how to make use of advanced science and technology to automatically identify and the defect information defect qualitative evaluation, also need more scientific research workers to contribute.Artificial neural network is adaptive, self-learning and parallel distributed information processing network structure. Since the 1980 s, the study on artificial neural network has made great progress in theory. There are hundreds of artificial neural network models, and widely used in signal processing, pattern recognition, image processing, medicine, meteorology, automatic control, financial forecasting and other fields with idiographic network structure and the performance of different algorithms and applications. It has excellent performance in the system of fault diagnosis to identify.Based on the experiment of laser ultrasonic surface defects of reflected wave and transmission wave signal data, this paper extracts can represent the characteristics of the ultrasonic defect signal. Using artificial intelligence technology, by using BP Neural Network model, the improved BP Neural Network model and Neural Network Regime Switching model(NNRS) to build a defect diagnosis system, gradually improve defect diagnosis system accuracy, stability and generalization ability, shorten the time complexity of the algorithm. In this paper, the main research contents include:1. The third chapter introduces the basic principle of the Mel Frequency Cepstral Coefficient method(MFCC), using the method of MFCC extraction of laser ultrasonic surface defects of reflected wave and transmission wave signal Frequency domain features. This chapter turns the high-dimensional signals of small sample data into variety characteristics of low-dimensional data, using normalization method to eliminate the high Frequency characteristic dimension, provide support for the neural network model input data.2. The Chapter 4 describes the development and basic principle of the neural network technology, build defect diagnosis system by using the BP neural network. This paper discusses the network structure and parameters selection, and starting from the shortcomings of BP and limitations, using additional momentum method to improve BP model, avoiding network in fixed weights may be trapped in local minima problem. With the actual measured data training test, the performance of additional momentum BP network model is better than BP. The defect signal classification accuracy up to 80%, the diagnostic rate of 100%.3. In the Chapter 5, we discuss the feedforward neural network architecture and its improved method, and put forward to introduce NNRS neural network model to construct ultrasonic surface defect signal diagnosis system. NNRS model increases layers in hidden layer, and set up linear connected between the input layer and output layer to enhance model nonlinear mapping ability and stable performance. There are model topology and parameters elaborated. With five groups of laser ultrasonic defect detection experimental data validation repeatedly, NNRS neural network model correct classification rate of 90%, the diagnosis rate of 100%. Compared with the BP network or with additional momentum, NNRS model nonlinear mapping ability and stability are stronger. And again through five groups of experimental data, it has better generalization ability.4. In Chapter 6, we analyze the complexity of algorithms in time and space of NNRS model. Using principal component analysis(PCA) for reducing dimension to optimize the time complexity of model in the aspect of data. By the PCA method processing reflection wave defect signal five groups of experimental data, and apply NNRS network model in classification. The diagnosis system classification accuracy is as high as 97%, the algorithm running time is reduced to the original 50%.