Generative adversarial nets in laser-induced fluorescence spectrum image recognition of mine water inrush

Abstract

Water inrush occurred in mines, threatens the safety of working miners which triggers severe accidents in China. To make full use of existing distinctive hydro chemical and physical characteristics of different aquifers and different water sources, this article proposes a new water source discrimination method using laser-induced fluorescence technology and generative adversarial nets. The fluorescence spectrum from the water sample is stimulated by 405-nm lasers and improved by recursive mean filtering method to alleviate interference and auto-correlation to enhance the feature difference. Based on generative adversarial nets framework and improved spectra features, the article proposes a novel water source discrimination-generative adversarial nets model in mines to solve the problem of data limitation and improve the discrimination ability. The results show that the proposed method is an effective method to distinguish water inrush types. It provides a new idea to discriminate the sources of water inrush in mines timely and accurately.

Keywords

Fluorescence generative adversarial nets mine inrush water source discrimination

Introduction

Water inrush in mines generally occurs when a large volume of water unexpectedly bursts into underground working areas in a short period of time, which jeopardizes the safe production of coal mine, causes casualties, and engenders severe economic losses. Many research works have verified that different aquifers and different water sources have distinctive hydro chemical and physical characteristics.^1,2–4 Based on these theories, the available methods to discriminate water source including trace elements,⁵ representative ions,⁶ water temperature level,⁷ and water flow conditions, in mines could be acknowledged in detail, that could be used to accurately forecast mine water inrush disasters.

However, still many unfavorable factors, such as lagging of monitoring responses, sample contamination, and heavy labor consumption, are caused by the method of water sample collection in mines in situ. It is impossible to measure water inrush type accurately in time. Recently, as some online sensors were proposed gradually, especially, the laser-induced fluorescence (LIF) technology is a twinkling innovation to measure the reflected fluorescence of the water sample online. The fluorescence is sensitive to hydro chemical and physical characteristics of different water and protects the tested water from being polluted.⁸ Therefore, the LIF severed as an effective online monitoring parameter used to discriminate water type for prediction of water inrush disaster.

However, discriminant model based on data driven is an important approach to solve the problem of discrimination of water type in mines nowadays. Related methods include the support vector machine (SVM),⁹ clustering analysis,¹⁰ extreme learning machine (ELM),³ back propagating neural network,¹¹ and so on. Gradually, these methods are applied to discriminate the type of water source in mine. Nevertheless, these methods adopt shallow frameworks, where useful features are necessary to be chosen in prior. It has to construct feature vectors in subjective and experiential ways. Randomly, initial weights and non-convex function of network also make the model be more likely to be trapped in local optimal solutions.

Recently, as deep neural network grows deeper and stronger, deep concept has been successfully applied to discriminate the type of water source in mines and predict the disaster of water inrush.¹ Particular, convolutional neural network (CNN) is specific to image data, which automatically learns from fluorescence spectra of LIF technology and avoids feature extraction and feature selection by human interference. Whereas, the large volume of data is vital to train an excellent discriminant model. However, available data are extremely limited in actual industrial production. This thorny phenomenon is the bottleneck problem to construct mine water source type discrimination model using CNN.

The generative adversarial nets (GAN) model was innovatively first proposed by Goodfellow who inspired by the game theory of two-person zero-sum game.^12–14 The modeling framework consists of a generated model and a discriminant model. The generated model is responsible for capturing the distribution of sample data and generating new data to the training model. Discriminant model is utilized to estimate the probability of the input form the real data or generated samples. The GAN provides a new idea to solve the shortcomings of validly data for discrimination of mine water inrush sources.

Methods

Recursive mean auto-correlation spectra feature construction

The recursive filer method can be used to reduce high-frequency fluctuation interference of signal. Equation (1) presents recursive average method, where N is utilized to control the flatness of the obtained instantaneous fluorescence spectra. Simple moving average filter could be used for this

y_{s} (n) = \frac{y (n - N) + y (n - N + 1) + \dots + y (n)}{N + 1}

(1)

where $y_{s} (n)$ is smoothed instantaneous fluorescence spectra, $y (n)$ is record of instantaneous fluorescence spectra, N is the length of the moving average filter. Then, it is further used to calculate the auto-correlation on collected data using LIF technology as following $R (n) = y_{s} (n) \times y_{s} (n)$ . Finally, the proposed image features of recursive mean auto-correlation spectra could be constructed for subsequent discriminant model.

Discrimination model of mine water inrush source based on GAN

The straightforward GAN framework of mine water inrush source is shown in Figure 1, which includes a generating model labeled by G and a discriminant model labeled by D. It defines a prior random noise variable $z$ and maps into generating model. Then, the generating model maps variables into data space as $G (z; θ_{g})$ to learn the distribution of inputs $z$ . Besides, the discriminant model $D (x; θ_{d})$ is utilized to detect the input data from the generated model or real data, where outputs of a single scalar $D (x; θ_{d})$ is the probability x from the real data rather than $z$ . $θ_{g}$ and $θ_{d}$ , respectively, represent parameters set of the generating model and discriminant model. Futhermore, to train D maximize the probability for judging the data source as accurately as possible; and minimize $\log (1 - D (G (z; θ_{g})))$ for G to estimate the distribution of generated data and training data as far as possible.

Figure 1.

GAN model structure diagram.

Based on the given generator G and the optimization discriminator D, the loss function is

obj (θ_{d}, θ_{g}) = E_{x ~ p_{data} (x)} [logD (x; θ_{d})] + E_{z ~ p_{z} (z)} [\log (1 - D (G (z; θ_{g})))]

(2)

where $p_{data} (x)$ is the distribution of real data x, $p_{z} (z)$ is the generating distribution of noise variable $z$ , and represents the expected value. Fixing the generating model $G (z; θ_{g})$ , the optimal discriminant model $D (x; θ_{d})$ is

D_{G}^{*} (x) = \frac{p_{data} (x)}{p_{data} (x) + p_{z} (x)}

(3)

The GAN model is a semi-supervised learning process. Suppose there are _m input train samples $(X = [x^{(1)}, x^{(2)}, \dots, x^{(m)}], x^{(i)} \in ℜ^{n + 1}, i = 1, 2, \dots, m)$ , where _m is the number of input samples, and $x^{(i)}$ is defined in $n + 1$ dimensional Euclidean space. The corresponding labels are $Y = [y^{(1)}, y^{(2)}, \dots, y^{(m)}]$ . Suppose there are also _m noise samples $Z = [z^{(1)}, z^{(2)}, \dots, z^{(m)}]$ from noise prior $p_{g} (z)$ .

Figure 2 shows training process of water source discrimination-generative adversarial nets (WSD-GAN) model. First, the auto-correlation fluorescence spectra severed as important features that are calculated by recursive mean filtering and auto-correlation method. It is further used to build a deep convolution generation network and initialize parameters of discriminant network. After trained by auto-correlation fluorescence data, the network model, finally, could be saved.

Figure 2.

Training process of WSD-GAN.

WSD-GAN model experiment and result analysis

Materials and spectral data acquisition

The water samples were gathered from the No. 2 Xinji mine of Huainan Mining Group from June 2016 to June 2017 at four different hydrogeology environments. These samples were named as Goaf hydrops, Sandstone water, Limestone water, and Ordovician Limestone water. All collected samples in situ have been placed in airtight polypropylene sample bottles, which had been cleaned by 10% HCl and distilled water before. All the water samples were measured in laboratory. The water samples included 46 items from the Goaf hydrops, 59 items from Sandstone water, 42 items from Limestone water, and 14 items from Ordovician Limestone water.

The fluorescence spectrum has been captured in totally dark environment avoiding influence from background and personnel activities. The wavelength of fluorescence spectrum ranges from 340 to 1340 nm with 0.5-nm resolutions. Each spectrum data has 2048 dimensions. The spectrum images of four different water sources are shown in Figure 3. Then, spectrum energy is focused in the range of 420–720 nm. These four curves are mainly composed of two peaks. The first peak emerges at 517 nm. The second peak arises at 560–580 nm.

Figure 3.

The spectrum of four different water sources.

Figure 4 presents auto-correlation fluorescence spectrum image. Consider distribution characteristics of the fluorescence spectrum, the last 1024 dimensions of fluorescence are almost overlapping together while the first 1024 dimensions of fluorescence are chosen to construct auto-correlation fluorescence spectrum image. Autocorrelation and normalization of one-dimensional (1D) spectral data after smoothing processing are performed to obtain the two-dimensional (2D) spectral diagram as shown in Figure 4, which has 1024×1024 frame size. Color is applied to distinguish the received energy among four different kinds of water emitted by the fixed laser. The brighter the color is the more spectral energy the band receives, where larger the fluorescence component of the sample is stimulated. Compared with the fluorescence spectra, it can better describe the changes of spectral curves in the whole band.

Figure 4.

Auto-correlation fluorescence spectrum: (a) Goaf hydrops, (b) Sandstone water, (c) Limestone water, and (d) Ordovician Limestone water.

These four fluorescence spectra, including Goaf hydrops, Sandstone water, Limestone water and Ordovician limestone water, present obvious differences in Figure 4. Considering, Goaf hydrops gradually coming with increased organic compounds, the auto-correlation fluorescence spectra shows seven levels of energy stages and a tendency to diverge gradually. Sandstone water gradually developing with reduced organic compounds in water, the auto-correlation fluorescence spectra shows six levels of energy stages and a tendency of the lowest three levels to diverge gradually. Limestone water growing with further reduced organic compounds in the water, where the lowest two levels energy of the auto-correlation fluorescence spectra shows a tendency to diverge gradually. Ordovician limestone water accumulating with fewer organic compounds in clear water, the auto-correlation fluorescence spectra shows an obvious tendency toward convergence. These results conclude that the proposed auto-correlation fluorescence spectra achieve a clear discrimination between Sandstone water and Limestone water and displays obvious convergence in the Ordovician Limestone water.

WSD-GAN model result analysis

Figure 5 presents the WSD-GAN model diagram, including the generating network and discriminant network. The generating network consists of three layers of transposed convolution. The first layer uses the convolution kernel 128@5 × 5 and the activation function is ReLU. The second layer uses the convolution kernel 64@4 × 4 and the activation function is ReLU. The third layer uses convolution kernel 1@2 × 2, and the activation function is Tanh. The discriminant network D consists of the convolution layer, the pooling layer, full connection layer, and the output layer. Convolution layers and the pooling layers connect alternately. The convolution layer adopts various specific convolution kernels and an activation function to compute output features. It adopts 5-fold cross-validation; 20% random samples used for the test and the remaining 80% used to train the model.

Figure 5.

WSD-GAN model of mine water inrush source.

Among three layers of convolutions, the first layer uses convolution kernels 32@2 × 2 and the activation function is LeakyReLU. The second layer uses convolution kernel 64@2 × 2 and the activation function is LeakyReLU. The third layer uses convolution kernel 96@2 × 2 and the activation function is ReLU. The pooling layer adopts the mean pool to select output features. The classifier is constituted by fully connected layer and the output layer using logistic regression, softmax regression, and SVM. The hyper parameters of the model includes: batch = 16, epoch = 150, learning rate λ = 0.0001, stochastic gradient descent (SGD) optimization algorithm, cross-entropy loss function, the penalty factor of SVM C = 1.5 and radial basis function (RBF).

The experimental results of WSD-GAN model with improved fluctuation feature are shown in Figure 6. By contrast, Figure 7 demonstrates the results of discrimination accuracy using those unprocessed fluorescence data and the WSD-GAN model. It could be observed that whether the improved feature set or the original data set, the proposed WSD-GAN model performs a gradually increasing tendency of correct recognition as iterations increase. These results are further used to verify the proposed WSD-GAN model performance on data robustness. Specially, based on the proposed WSD-GAN model, the improved features achieve more excellent recognition rate at the end of the iteration.

Figure 6.

Discrimination accuracy of WSD-GAN model with improved features.

Figure 7.

Discrimination accuracy of WSD-GAN model with unprocessed auto-correlation feature.

Finally, as a comparison, we have compared our method with the other four different state-of-the-art classification methods, namely, feedforward chaotic neural network (FCNN),¹⁵ Bp network, light gradient boosting machine (LightGBM),¹⁶and convolution neural network (CNN). In the experiment, the LightGBM parameters include learning rate $η = 0.1$ , the num_leaves $α = 100$ , the max_depth $h = 12$ , and the number of trees $n_estimators = 300$ , negative binomial log-likelihood as the objective function for classification. The network parameter settings are the learning rate $η = 0.1$ , the correcting-weight impulse coefficient $α = 0.05$ , the training derivation 0.0001, and the max training time J = 1000. The FCNN model is a single-hidden-layer including output 14 nodes, input 4 nodes, and hidden layer 8 nodes. The structure of CNN model is the same as the discriminant network of WSN-GAN model. The CNN parameters include dropout ratio 0.2, moment 0.9, learning rate 0.0025, regularization type L₂, and minibatch 64.

Table 1 reports the discrimination accuracy of different models with an improved feature spectra. It shows that our proposed method could perform optimal ability of water source discrimination than the other four methods. Specifically, our method is 0.62%, 0.91%, 16.33%, and 12.40% over CNN, LightGBM, Bp, and FCNN. More interestingly, based on the proposed improved recursive mean auto-correlation spectra, WSD-GAN algorithm performs similar discrimination accuracy of CNN and LightGBM.

Table 1.

Comparison of different models.

Model	Recognition rate (%)
FCNN	82.34
Bp	78.41
LightGBM	93.83
CNN	94.12
WSD-GAN algorithm	94.74

FCNN: feedforward chaotic neural network; GBM: gradient boosting machine; CNN: convolution neural network; WSD-GAN: water source discrimination-generative adversarial nets.

Conclusion

This article developed a new method to discriminate the source type of water inrush based on improved fluorescence and GAN for water inrush forewarning of mines. In the first stage, we proposed a novel method to construct features made by recursive mean filtering method and auto-correlation calculation, so that we can benefit from the alleviated fluctuation of interference and the magnified area around the resonance peaks. Some individual differences among different water sources are obviously presented. The characteristic of organic compounds in water can be effectively distinguished. In the second stage, the proposed WSN-GAN model is used as the classifier to solve the problem of the exciting small-sample set. The theoretical analysis and experimental results show that the proposed method is an effective assessment method to define inrush water source types in a mine. It provides a new method to solve discrimination of inrush water sources in mines.

Footnotes

Acknowledgements

The authors would like to thank anonymous reviewers and the associate editor, whose constructive comments helped to improve the presentation of this work.

Handling Editor: Konstantinos Kalpakis

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work is supported by National Natural Science Foundation of China (Nos 51875100, 61673108, and 41674133), University Science Research Project of Jiangsu Province (No. 19KJB470024), Science Application Basic Research Program of Xuzhou City (No. KC18015).

ORCID iD

Jing Li

References

Yong

Jianhua

Jing

, et al. Mine water inrush sources online discrimination model using fluorescence spectrum and CNN. IEEE Access 2018; 6: 47828–47835.

Zeng

Liu

, et al. Vulnerability assessment of water bursting from Ordovician limestone into coal mines of China. Environ Earth Sci 2016; 75(22): 1431.

Zhao

. Forecasting model of coal mine water inrush based on extreme learning machine. Appl Math Inform Sci 2013; 7(3): 1243–1250.

Zhu

Tuia

Mou

, et al. Deep learning in remote sensing: a comprehensive review and list of resources. IEEE Geosci Remote S 2018; 5(4): 8–36.

Cui

Zhao

, et al. Type classification and main characteristics of mine water disasters. J Chin Coal Soc 2013; 38(4): 561–565.

Wang

Meng

, et al. Seawater inrush assessment based on hydrochemical analysis enhanced by hierarchy clustering in an undersea goldmine pit, China. Environ Earth Sci 2014; 71(12): 4977–4987.

Meng

Xie

. A geological assessment method of floor water inrush risk and its application. Eng Geol 2012; 143: 51–60.

Wang

Hou

. Geothermal water at a coal mine: from risk to resource. Mine Water Environ 2016; 35: 294–301.

Sun

Gui

. Hydro-chemical evolution of groundwater and mixing between aquifers: a statistical approach based on major ions. Appl Wat Sci 2015; 5(1): 97–104.

10.

Qiang

Yulong

. Vulnerability assessment of groundwater inrush from underlying aquifers based on variable weight model and its application. Water Resour Manag 2016; 30: 3331–3345.

11.

Yang

Yue

, et al. Study on PCA-LDA for fast identifying the type of coal mine water with LIF technology. Chem Engineer Trans 2016; 51: 1135–1140.

12.

Goodfellow

Pouget-Abadie

Mirza

, et al. Generative adversarial nets. In: Proceedings of the 27th international conference on neural information processing systems, Montreal, QC, Canada, 8–13 December 2014, pp.2672–2680. Cambridge, MA: MIT Press.

13.

Killoran

Lee

Delong

, et al. Generating and designing DNA with deep generative models. arXiv preprint arXiv:1712.06148, 2017, https://arxiv.org/abs/1712.06148

14.

Schlegl

Seeböck

Waldstein

, et al. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In: Proceedings of the international conference on information processing in medical imaging, Boone, NC, 25–30 June 2017, pp.146–157. Cham: Springer.

15.

Peng

Liu

Cheng

, et al. Feedforward chaotic neural network model for rotor rub-impact fault recognition using acoustic emission method. J Electr Comput Eng 2018; 2018: 9718951.

16.

Meng

Finley

, et al. Lightgbm: a highly efficient gradient boosting decision tree. In: Proceedings of the 31st conference on neural information processing systems (NIPS 2017), vol. 18, Long Beach, CA, 4–9 December 2017, pp.3149–3157, https://dl.acm.org/citation.cfm?id=3295074