High Current Linear Induction Accelerator Fault Diagnosis And Performance Assessment Techniques

The high current linear induction accelerator, as one of the most important diagnosis tools for hydro-dynamics in detonation experiments, must be operated stably and reliably. The opetation mode in single shot of LIA requires that the LIA should show the best performance at given moment. Therefore, to develop an initelligent fault diagnosis and performance assessment system is indispensable to LIA. Meanwhile, high current LIA also is a complicated and huge system. it is difficult to develop such a system. It is the reason that many foreign acceletator laboratories have been dedicating to develop such a system during the past decades, but still in searching and developing stage, whereas in starting phase in our country. In this paper, the applications of pulsed signal processing such as de-noising, singularity detection and time interval measurement by using wavelet transform are presented. Because of the complex physical mechanism of LIA and difficulties of systemic modeling, a new method of pattern recognition and fault diagnose for LIA based on wavelet packet transform -neural network has been put forward.Signal singularity detection and time interval measurement are important in the LIA's signal processing field. After analyzing the singularity exponent of LIA's signals, some selected wavelet function which can accurately detect the break point of diverse signals in the noised background have been given, time interval measurement automatic systems have been built based on the singularity detection principle. According to the distinct features towards the useful and noised signal in wavelet transform orders, a decomposition order algorithm by verifying the white noise in wavelet transform orders has been given. The experimental results demonstrate that the best de-noising effect can be obtained under the hard threshold and fixed threshold rule.In the paper, wavelet packet space is used as the characteristic space for pattern recognition. According to the corresponding relationship between the extremum of frequency point under Fourier transform and that of frequency band under wavelet packet transform, the methods of determining the decomposition order and extracting characteristic vector from energy on the coefficients of wavelet packet transform have been proposed. The experimental results demonstrate that characteristic vector represents feature of LIA's waveform well, and reduces the dimensions effectively.Based on the extracted characteristic vector, by using of improved nearest neighbor clustering algorithm, a RBF neural network prototype system towards accelerator cell with beam and a RBF neural network trend classification system towards beam current from the exit of LIA injector have been built. In order to satisfy the needs of LIA fault diagnosis, a distributed diagnosis strategy based on neural networks is also presented. As an example, a jointed inference prototype system which combining multi-networks and ruler classification has been built. Diagnosing results show these applications can detect faults effectively, give an automatic quick view of the state of the accelerator during the experiments, can detect LIA's performance drifting trend, and also can provide predictive information for precise maintenance of LIA.