Elsevier

Neurocomputing

Volume 148, 19 January 2015, Pages 222-228
Neurocomputing

Material identification of loose particles in sealed electronic devices using PCA and SVM

https://doi.org/10.1016/j.neucom.2013.10.043Get rights and content

Abstract

The existence of loose particles left inside the sealed electronic devices is one of the main factors affecting the reliability of the whole system. It is important to identify the particle material for analyzing their source. The conventional material identification algorithms mainly rely on time, frequency and wavelet domain features. However, these features are usually overlapped and redundant, resulting in unsatisfactory material identification accuracy. The main objective of this paper is to improve the accuracy of material identification. First, the principal component analysis (PCA) is employed to reselect the nine features extracted from time and frequency domains, leading to six less correlated principal components. And then the reselected principal components are used for material identification using a support vector machine (SVM). Finally, the experimental results show that this new method can effectively distinguish the type of materials including wire, aluminum and tin particles.

Introduction

Sealed electronic devices are widely used in communication, remote control and scientific experiments in a satellite system. Their reliability is vital to the success of the mission and safety of personnel and devices [1], [2]. However, due to their complex structure and production process, some metal and nonmetal loose particles may be left inside, such as the wire pieces, aluminum scraps and tin dregs. When these loose particles are in the vibration or shock environment, they can be freed and collided randomly, which may lead to short circuits, component malfunction, system breakdown, or even aerospace catastrophes in system operations. Therefore, it is critical to investigate loose particle detection and identification technologies in order to increase the reliability of sealed electronic devices [3].

The Particle Impact Noise Detection (PIND) test is a traditional loose particle detection technique for sealed components, which is specified in MIL-STD-883E standard [4]. Fig. 1 represents the typical structure of PIND system. More specially, loose particles are first freed and then collided with component walls using a series of shocks and vibrations generated by a shaker. The resulting collision energy is transformed to acoustic signal by the sensitive acoustic emission (AE) sensor mounted on the test device. Finally, a speaker and oscilloscope present the captured acoustic signal, which can be used to estimate the presence of loose particles.

The PIND test mentioned above is an AE technology based non-destructive method for detecting loose particles inside the sealed components such as relays, transistors and switches. However, due to the many subjective factors of operators, wrong conclusions can be easily drawn, and the accuracy of particle detection is only around 44% [5]. Therefore, it is urgent to develop an automatic particle detection system. In order to eliminate the artificial factors, many efforts have been paid to automate the testing operation. Based on modern computer technology, the impact acoustic signal is collected by a data acquisition card, and the result is analyzed by the computer software instead of audio and visual determination. The structure of the novel particle detection system for sealed electronic devices is shown in Fig. 2.

Before detecting the loose particles, the experimental conditions for freeing the loose particles have to be determined. A dynamical mathematical model is established for particles collision within aerospace relays by considering the coefficient of restitution and quality of particles in [6] where the best vibration condition is derived and it indicates that output power of the particle is in proportional to the vibration acceleration and the amplitude of velocity of vibrator [3], [7]. For loose particle detection and material identification, in order to overcome the problems of the conventional manual methods and therefore improve their accuracies, Scaglione [5] and Ma [8] employed neural networks to produce more accurate results, leading to greater testing reproducibility and confidence. Another investigation carried out in [9] proposed to filter the negative effects of noise using wavelet and complex wavelet denoising to achieve a higher performance. Further work on identifying the types of loose particle material including tin dregs, glass scrap, wire and rubber particles reported in [10] extracted three features of pulse duration time, spectrum shape and the linear prediction coefficients of impact acoustic signals and then the artificial neural network are used to classify their material types. Recently, a new feature of the disturbance signal energy distribution vectors in vibration acceleration is used for loose particles classification based on back propagation (BP) neural networks [11]. In addition, the test objective is changed from relays to aerospace power, and a material identification method is presented by employing wavelet and neural networks [12].

As all the existing material identification methods are black-box methods. There is no qualitative or quantitative analysis between different material particles and acoustic signals. Thus, the features directly obtained from time, frequency and wavelet domains depend on experience. This can result in overlapping and redundancy. Though these methods can be used for small components, they are incapable to deal with large and complex sealed electronic devices as their performance is seriously affected by widespread impact position, complex path of wave propagation and serious background noise.

To solve this problem, typical particles impact acoustic signals are compared in time and power spectral distribution (PSD) domains, and nine features of impact acoustic signal are selected from time and frequency domains. PCA is then used for further feature extraction, dimension reduction and de-noising. The top six principal components are input to support vector machine as eigenvectors. Thus, a novel particle material identification method based on PCA and multi-SVM is developed for sealed electronic devices. Experimental results are finally presented, demonstrating the effectiveness of the proposed classification method.

Section snippets

Test system

The data used in this work were collected from PIND automatic detection system as shown in Fig. 3. With a high-power vibrator, vibrations make the particles to collide with the walls of the sealed electronic device. The collided energy is released in the form of elastic wave propagating along the walls. AE sensors mounted on the walls convert the collision signal into electronic signal. After amplification and data acquisition, the electronic signals are recorded by a computer. The sensor is

Principal component analysis

In traditional methods, the features directly obtained from different material particle collision signals could lead to overlapping patterns. In order to reduce the redundant information, PCA was performed for further feature extraction.

As a widely used statistical technique, PCA has been employed to reduce the dimensionality of problems and to transform interdependent coordinates into significant and independent ones [17]. This technique and other PCA-based methods have been successfully

Basis of SVM

In order to enhance the classification accuracy, the data mining method of SVM combined with PCA were used to identify the particle materials.

SVM is a machine learning method using small samples. Based on structural risk minimization principle, SVM minimizes the empirical risk and Vapnik–Chervonenkis (VC) dimension simultaneously [24]. SVMs can efficiently perform non-linear classification problems by implicitly mapping their inputs into high dimensional feature spaces using the Kernel tricks.

Conclusions

In this paper, the particle material identification method based on principal component analysis is investigated for the sealed electronic devices. In the literature, features are extracted directly in the time, frequency and wavelet domains. They are less reliable due to their redundancy and they have not been applied to the large and complex devices. In order to address this problem, nine eigenvectors are extracted in the time and frequency domains. PCA is then performed for further feature

Acknowledgments

This research is supported by the National Nature Science Foundation of China under grants 61271347 and 51077022.

Guofu Zhai was born in 1964. He received the Ph. D. degree from Harbin Institute of Technology, Harbin, China, in 1998. He is currently a professor of the School of Electrical Engineering and Automation, Harbin Institute of Technology. His main research interests include reliability technique of electrical apparatus and electronic device, and nondestructive testing technique of electromagnetic ultrasound.

References (29)

  • X.H. Li et al.

    Estimation of crowd density based on wavelet and support vector machine

    Transaction of the Institute of Measurement and Control

    (2006)
  • F. Tong et al.

    Evaluation of tile–wall bonding integrity based on impact acoustics and support vector machine

    Sensors and Actuators A: Physical

    (2008)
  • D.R. Salgado et al.

    An approach based on current and sound signals for in-process tool wear monitoring

    International Journal of Machine Tools & Manufacture

    (2007)
  • S. Iplikci

    Support vector machines based neuro-fuzzy control of nonlinear systems

    Neurocomputing

    (2010)
  • Cited by (24)

    • A weight recognition method for movable objects in sealed cavity based on supervised learning

      2022, Measurement: Journal of the International Measurement Confederation
      Citation Excerpt :

      Finally, the obtained test results were verified by discriminant rules. A typical case - weight detection of movable particles in sealed electronic components, is selected as the verification of the method proposed in this paper[40]. Electronic component with movable particles has a single logic function which creates electronic circuits for specific functions, such as amplifiers, wireless receivers, and oscillators.

    • Data-driven simultaneous fault diagnosis for solid oxide fuel cell system using multi-label pattern identification

      2018, Journal of Power Sources
      Citation Excerpt :

      On one hand, it can avoid the large values dominating the small ones. On the other hand, scaling attribute values into a small range will avoid dimension disaster and decrease calculation load [43–46]. As a result, in this research the PCA is utilized to extract the features in redundant and correlative data from both the sensors and the controllers.

    • An intelligent system for quality measurement of Golden Bleached raisins using two comparative machine learning algorithms

      2017, Measurement: Journal of the International Measurement Confederation
      Citation Excerpt :

      This reduction is accomplished by identifying directions, called principal components (PCs), along which a set of correlated variables are transformed into a set of values of linearly uncorrelated maximal variables in the data. This conversion process is defined in such a way that the first principal component has the largest possible variance, and each following component in turn has the highest variance possible on the condition that it be orthogonal to the previous components [16,21]. Fig. 6 illustrates two principle components PC1 and PC2, whereby PC1 equals the maximum variance and PC2 is perpendicular to the first PC1 and it accounts as much as possible for the remaining variance.

    View all citing articles on Scopus

    Guofu Zhai was born in 1964. He received the Ph. D. degree from Harbin Institute of Technology, Harbin, China, in 1998. He is currently a professor of the School of Electrical Engineering and Automation, Harbin Institute of Technology. His main research interests include reliability technique of electrical apparatus and electronic device, and nondestructive testing technique of electromagnetic ultrasound.

    Jinbao Chen received the M.S. degree from the Harbin Institute of Technology, Harbin, China, in 2010. He is currently pursuing the Ph.D. degree with the Department of Electrical Engineering, Harbin Institute of Technology, Harbin, China. His current research interests include automatic detection and material identification of loose particles in sealed electronic devices.

    Shujuan Wang was born in 1967. She received the Ph. D. degree from Harbin Institute of Technology, Harbin, China, in 1998. She is currently a professor of the School of Electrical Engineering and Automation, Harbin Institute of Technology. Her research interests are reliability research and fault diagnosis of power electronic devices, the detection about electronic reliability of passenger train, and the application of power electronic technology in military hybrid electrical apparatus.

    Kang Li was educated in Xiangtan University, Harbin Institute of Technology, and Shanghai Jiaotong University, all in China. Dr. Li is a Reader in intelligent systems and control at Queen׳s University Belfast. His research interests include nonlinear system modeling and identification, advanced algorithms for training and construction of neural networks, fuzzy neural nets and support vector machines, as well as evolutionary algorithms, with application to power plants and systems, polymer extrusion, environmental monitoring, food security and biomedical systems. Dr.Li serves in the editorial board of Neurocomputing, and Transactions of the Institute of Measurement and Control. He is the current Secretary of the IEEE UKRI Section, a Chartered Engineer and a member of IEEE and InstMC

    Long Zhang was received the M.S. degree from the Harbin Institute of Technology, Harbin, China, in 2009. He is currently pursuing with the Ph.D. degree from Queen׳s University Belfast. His currently pursuing interests include non-linear system modelling and identification using grey-box method, neural networks and genetic algorithms.

    View full text