Ammunition Supply System Fault Prediction Based On PCA-KLD And Deep Learning

The ammunition supply system is the core of artillery weapon system.Its stuck fault seriously restricts the timing of combat.So it is of great significance to carry out health monitoring and forecast the potential failure.In this paper,the working principle and common faults of the large caliber Gatling gun were analyzed.On the basis of this,measuring points were placed and multi-field signals were collected.In order to solve the problem of high noise in multi-field signal,an adaptive block threshold denoising method based on biorthogonal spline wavelet was proposed.The core of this method is to regard the Stein unbiased risk estimation as the constraint condition,and to optimize the threshold and the neighborhood length adaptively.The proposed method is better than the block threshold wavelet denoising and the adjacent coefficient wavelet denoising,because this method has the best effect on weak impact signal,and it can make the noise reduction results have the biggest correlation with the simulation impact signal.The results showed that this method highlighted the energy from 1 to 10 times working frequency.In order to solve the problem of complexity of the ammunition system signal,the method of principal component analysis(PCA)and profile likelihood was used to reduce the dimension.This method can eliminate redundant components and ensure the sensitivity of fault identification.The effectiveness of this method was validated via a simple uniform experiment.Then,the idea of potential fault prediction was explained.Firstly,constructing PCA model by the reference data.Secondly,the Kullback-Leibler Divergence(KLD)was used to measure the dissimilarity between the probability density function of reference and unknown working condition latent scores.Then the potential fault identification ability of KLD was verified.The results showed that KLD had a better performance over Wilcoxon Rank Sum Test(WRST)when a priori information on the distribution was known.Then wavelet time-frequency graph of the ammunition supply system was set as the input data for PCA-KLD algorithm.And the wavelet selection criteria was set within the biorthogonal spline wavelet.Because the spectrum energy was mainly concentrated in the working frequency sequence,the working frequency sequence was transformed to scale sequence,which was used to compute the wavelet time-frequency graph.There PCA models were constructed by three healthy data respectively.Development trend of the KLD values that computed from measured data was analysed,and the symptom of fault was found out.Finally,three kinds of data(healthy,faulted and deteriorating),were classified by the improved depth belief network(DBN).The results showed that the improved DBN was better than BP neural network and traditional DBN.