Detection of gas leakage sound using modular neural networks for unknown environments
Introduction
The detection of gas leakage sound from pipes is important in petroleum refining plants and chemical plants, as often the gas used in these plants are flammable or poisonous. Gas sensors can also detect the gas leakage, but it is not always suitable for the early detection of the gas leakage. The detection ability of the gas sensor is strongly influenced by the direction of the wind, the force of the wind, the density of the gas, and by other factors. On the other hand, if the leakage sound of the gas was used, we could immediately detect the leakage when the gas leaked from pipes. Furthermore, this type of diagnosis uses the sound detected by microphone, and therefore has the excellent features of non-contact detection and easy handling.
We have already evaluated the detection performance with the neural network for relatively short-term experiment (a few weeks) [5], that is considered as a stationary state. The applied network is the multi-layered feed-forward model whose learning algorithm is the error back-propagation (BPN) [12]. Experimental results have shown that the neural approach was effective to detect the gas leakage. On the other hand, the stable diagnosis technique in the long term is necessary to apply for practical use. It is supposed that the regular inspection and the deterioration of many machines around a microphone affect the background noise. Changes in background noise influence the reliability of diagnosis results. Such changes in the background noise mainly depend on the operating situation of machines around a microphone. Since there is a possibility that the operating situation is back to the former state after a while, the background noise could return to the former sound, too. We consider that the property of the background noise is basically stationary. However, the temporal changes in the acoustic properties of sounds emitted from a machine impose a dynamic environment on a diagnosis system, which can be modeled as a transition of successive stationary environments. This means that there is a non-stationary state during the change of stationary states.
It is necessary to examine the diagnostic method that is able to adapt to the above-mentioned property of environment. Here, the environment is defined as factors to affect the background noise. The simple method to adapt to the environment is to re-train the network when environment changes. However, the network after the re-training might change the diagnosis result even for the same data as the network diagnosed before the re-training. It seems that the re-training strategy to adapt to the new environment results in modifications of the network. Therefore, the examination of a novel strategy of neural network model is necessary to overcome this difficulty. A model of modular neural network is one of the promising solutions. Decomposing a task into modular components leads to the improving performance [13]. Several modular networks have been studied such as divide and conquer approaches [4] and class decomposition approaches [1], [7]. Furthermore, modular neural networks have been applied to the non-stationary signal processing [3], fault diagnosis [11], and so on. Since most of these studies relate how to decompose the given task automatically, these studies are not suitable for our requirement regarding the following points: the trained modules are not modified and keep their weights if a new module were added to the network to adapt to the new environment. That is to say, we need a constructing strategy for a network model to adapt to new tasks without changing results for already trained tasks under dynamic environment.
The purpose of this study is to propose a novel strategy of neural networks to adapt to dynamic environment. We show an example of the modular neural networks to realize the strategy. Furthermore, we apply the application of the modular networks to the sound collected over about 18 months' experiments in the petroleum refining plants to evaluate the effectiveness of the proposed strategy.
This paper is organized as follows: in Section 2, we propose a novel model of modular neural networks. In Section 3, we describe experiments for gas leakage and acoustic characteristic of the sound. We show results of experiments and discuss the effectiveness of the proposed model in Section 4. Finally, we give a brief summary and concluding remarks in Section 5.
Section snippets
Architecture and learning algorithm
There seem to be various models of modular neural networks to realize our strategy. In this section, we propose an example of modular neural networks to adapt to the dynamic environments and keep the diagnostic result for the same data as the network diagnosed before the adaptation to the new environment. The modular neural network consists of an initial module and some supplementary modules as shown in Fig. 1. The initial module is for the known environment and the supplementary module is for
Experimental setup and acoustic characteristics of gas leakage
The leakage sound from the defect of pipes is influenced by the configuration of the defect, the gas pressure inside the pipe, the kind of gas, and so on. Since the “pure” leakage sound of the gas is added to the background noise, the sound detected by a microphone suffers greatly from the influence of the background noise. The “pure” leakage sound means the leakage sound without the background noise. The leakage sound in this paper is the sum of the “pure” leakage sound and the background
Results of experiments
Here, we evaluated the effectiveness of the modular neural network using the above-mentioned acoustic data. The input information was 200 dimensional logarithmic power spectra and the output of the network was the diagnostic result. Each module had 200 input units and an output unit. The target pattern was for the background noise and for the leakage sound. The input data was judged “normal” for the output over , “abnormal” for the output below , and “unknown” for the other
Discussion
Here, we compared the proposed network with the simple network being non-modular networks with the re-training strategy. The network is the Gaussian Potential Function Network (GPFN) proposed by Lee [6]. In the GPFN, the hidden units are automatically allocated to maintain good classification performance. The GPFN is able to perform classification based on a set of potential fields synthesized over the domain of input space by a number of Gaussian potential function units. The GPFN is composed
Conclusions
We have described the detection method of gas leakage from pipes for practical use. We have proposed a novel strategy of neural networks to adapt to the dynamic environments and have shown an example of the modular neural networks to realize the strategy. Experiments were performed for an artificial gas leakage device under various experimental conditions over about 18 months in a petroleum refining plant. Experimental results showed that the proposed network could adapt the structure to
Manabu Kotani has received the B.E. and M.E. degrees in Instrumentation Engineering from Kobe University, Kobe, Japan in 1981 and 1983, respectively. He received the D. Eng. from Kobe University in 1994. He had been with Kobe Steel, Ltd., had researched for the non-destructive testing with the eddy current method. He is now an associate professor in the Department of Computer and Systems Engineering, Kobe University. His current research interests are the acoustic diagnosis and the neural
References (14)
- et al.
Efficient classification for multiclass problem using modular neural networks
IEEE Trans. Neural Networks
(1995) - M.H. Hassoun, Fundamentals of Artificial Neural Networks, MIT Press, Cambridge, MA, 1995, pp....
- et al.
Modular learning strategy for signal detection in a nonstationary environment
IEEE Trans. Signal Process.
(1997) - et al.
Adaptive mixtures of local experts
Neural Comput.
(1991) - et al.
Acoustic diagnosis for gas leakage in pipes under noisy environments
- et al.
Multilayer feedforward potential function network
- et al.
Task decomposition and module combination based on class relationsa modular neural network for pattern classification
IEEE Trans. Neural Networks
(1999)
Cited by (0)
Manabu Kotani has received the B.E. and M.E. degrees in Instrumentation Engineering from Kobe University, Kobe, Japan in 1981 and 1983, respectively. He received the D. Eng. from Kobe University in 1994. He had been with Kobe Steel, Ltd., had researched for the non-destructive testing with the eddy current method. He is now an associate professor in the Department of Computer and Systems Engineering, Kobe University. His current research interests are the acoustic diagnosis and the neural networks for pattern recognitions.
Masanori Katsura has received the B.E. and M.E. degrees in Computer and Systems Engineering from Kobe University, Kobe, Japan, in 2001 and 2003, respectively. He had studied the application of neural networks to pattern recognitions. He is now with Hitachi, Ltd.
Seiichi Ozawa received his B.E. and M.E. degrees in Instrumentation Engineering from Kobe University, Kobe, Japan, in 1987 and 1989, respectively. In 1998, he received his Ph.D. degree in computer science from Kobe University, Kobe, Japan. His is currently an Associative Professor in Graduate School of Science and Technology, Kobe University. His primary research interests include incremental learning for artificial neural networks, reinforcement learning, life-long learning, and pattern recognition.