Training and Results of Neural Network

Một phần của tài liệu Machine learning and intelligent communications part i 2017 (Trang 148 - 155)

As mentioned earlier, increasing the training samples is one way to reduce over- fitting. Another is L2 normalization (weight decay). The idea of L2 normalization is to add an additional item to the loss function (see [13]). The regularization term shown in Eq.2is added to the cross entropy loss function which is used in this thesis.

C=1 n

xj

yjlnaLj + (1−yj) ln(1−aLj) + λ

2n

w

w2 (2)

wherexis the input sample,ais the actual output vector of the neural network, y is the desired output vector for the neural network,wis the weight. The first term in Eq.2is just the usual expression for the cross-entropy. The second term, namely the sum of the squares of all the weights in the network is added to the cross-entropy. This is scaled by a factor 2nλ, where λ > 0 is known as the regularization parameter.

In the practical training process, the dropout method is also applied [3]. The dropout rate is 50%, which means that the connection layer neuron output is set to 0 randomly. In this way, the corresponding weight will not be updated, which can help reduce over-fitting. The training method applies the Nesterov gradi- ent acceleration (NAG) method, which is slightly different from the momentum update and stochastic gradient descent and has become more popular now [2].

Nesterov can guarantee a stronger theoretical convergence for convex function and the practical performance is better than momentum and stochastic gradient descent. The core idea of Nesterov is to accelerate in the same gradient direction and decelerate in the changing gradient. The momentum update formula is Eq.3.

ΔVt=ρΔVt−1−η[∇C(Vt−1)]T (3) The momentum update includes two parts: one is the updated valueρΔVt−1, which has been calculated at last moment and is known at this moment; the other is calculated gradient η(∇C)T based on the current position. Nesterov accelerated gradient points out: since it is known that the update at this time will goρΔVt−1, then goρΔVt−1first and correct according to the gradient there.

Therefore, the gradient calculated in this method is not on a basis of old position, but is forward-looking.

ΔVt=ρΔVt−1−η[∇C(Vt−1−ρΔVt−1)]T (4) The NAG update formula is shown in Eq.4. In Eq.4, the learning rate is set to 0.005, the momentum factor is set to 0.9, the batch size is set to 64, with a total of 2000 steps, that is 25 epochs.

Figure8shows the relationship between the accuracy of the classification and the number of training steps. The solid line indicates the accuracy rate of the training set. The test is to extract 1000 images randomly from the training set

0 200 400 600 800 1000 1200 1400 1600 1800 2000

Step 0

10 20 30 40 50 60 70 80 90 100

Accuracy(%)

Accuracy on the training data Accuracy on the validation data

Fig. 8.Accuracy on the training and validation data set.

0 200 400 600 800 1000 1200 1400 1600 1800 2000 Step

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Loss

Loss value on training set Loss value on the validation set

Fig. 9. The loss function is changing with the number of steps on the training and validation data set. The value of the loss function is small, indicating that the neural network has achieved better performance.

Fig. 10.Testing a single picture. The horizontal coordinate represents the probabilities predicted by the neural network and vertical coordinate represents the types of cracks.

to test the accuracy every 50 training steps. At Step 2000, the accuracy of the training set is 98.40%. The dotted line indicates the accuracy of the validation set. During the training process, there is validation on validation set every 200 steps. The final accuracy is 98.35%. As is shown, the network have achieved relatively good results on the test set and the validation set. The results indicate that the network is reasonable and is of strong generalization ability without over-fitting. This point can be seen from Fig.9. The loss function finally reached a very small value of 0.2. Figure10 shows the probabilities of different defects when testing a single image.

4 Acquirement of the Battery Unit EL Images

As is shown in Fig.3, the first thing is to obtain crack images from the camera, and then use the neural network to identify the crack. The camera captures the EL image of the entire PV module, but the neural network requires an image of a single cell. So it is necessary to pre-process the acquired EL images. The processing steps are shown in Fig.11.

Fig. 11.The process of images processing

First, the morphological closed operation is used to fill the small voids in the foreground, especially reducing the effect of the two main grid lines on each cell. Otherwise, the two main grid lines will cause negative effects on border extraction. Second, use the bilateral filter algorithm to filter the pictures (see [8]). Bilateral filtering algorithm is a nonlinear filtering method, the advantage of which is to retain the edge information when filtering. This is favorable for the edge extraction. Next, use Canny edge detection to extract the edges (see [6]). Finally, adopt the method in literature [10] to extract the boundaries of each cell. The whole result is shown in the Figs.12and13.

(a) Input image (b) Closing operation (c) Bilateral filtering

(d) Extract the edges

with canny algorithm (e) Extract the bor- der

Fig. 12.The images (a)–(e) in turn show the results obtained after each step of image processing

Fig. 13. Final recognition of results. To show clearer images and results, the images above are clipped from the input images. Thus, part information is missing and some cracks in the image are not detected.

5 Conclusion

The system first obtains the image data of each battery cell through a series of image processing algorithms, and then puts it into a well-trained neural network for classification. The accuracy of recognition reaches 98.40%. The characteristic of the whole system is that the algorithm can accurately extract the correspond- ing information and accurately identify the defects in the EL image collected from the camera no matter how much angle the PV modules rotate or when the

modules tilt in the range of 20 20. Thus, the system is very suitable for auto-inspecting the photovoltaic power plant, such as using a unmanned aerial vehicle (UAV) equipped with cameras to inspect power station.

References

1. China National Energy Administration: 13th five year plan for solar energy devel- opment. Solar Energy12, 5–14 (2016)

2. Bengio, Y., Boulanger-Lewandowski, N., Pascanu, R.: Advances in optimizing recurrent networks. In: 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8624–8628. IEEE (2013)

3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep con- volutional neural networks. In: Advances in neural information processing systems, pp. 1097–1105 (2012)

4. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521(7553), 436–444 (2015)

5. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proc. IEEE86(11), 2278–2324 (1998)

6. Rong, W., Li, Z., Zhang, W., Sun, L.: An improved canny edge detection algorithm.

In: 2014 IEEE International Conference on Mechatronics and Automation (ICMA), pp. 577–582. IEEE (2014)

7. Suzuki, S., et al.: Topological structural analysis of digitized binary images by border following. Comput. Vis. Graph. Image Process.30(1), 32–46 (1985) 8. Zhang, B., Allebach, J.P.: Adaptive bilateral filter for sharpness enhancement and

noise removal. IEEE Trans. Image Process.17(5), 664–678 (2008)

9. Fuyuki, T., Kitiyanan, A.: Photographic diagnosis of crystalline silicon solar cells utilizing electroluminescence. Appl. Phys. A96, 189–196 (2009)

10. Wu, Z., Huang, H.: Research on the effects of PV modules’ transformation on the performance. Mech. Eng. Autom.4, 107–109 (2011)

11. Xu, Z., Wang, H., et al.: Research progress in crack features of PV modules: part 1. Solar Energy10, 47–51 (2015)

12. Kajari-Schr¨oder, S., Kunze, I., Eitner, U., K¨ontges, M.: Spatial and orientational distribution of cracks in crystalline photovoltaic modules generated by mechanical load tests. Solar Energy Mater. Solar Cells95, 3054–3059 (2011)

13. Krogh, A., Hertz, J.A.: A simple weight decay can improve generalization. In:

NIPS, vol. 4, pp. 950–957 (1991)

14. MNIST database of handwritten digits.http://yann.lecun.com/exdb/mnist/

Construction Method Based on MEMS

Mu Zhou(B), Xiaoxiao Jin, Zengshan Tian, Haifeng Cong, and Haoliang Ren Chongqing Key Lab of Mobile Communications Technology,

Chongqing University of Posts and Telecommunications, Chongqing 400065, China {zhoumu,tianzs}@cqupt.edu.cn, jxxcq 836235528@foxmail.com,

18008382985@163.com, 13108970732@163.com

Abstract. In indoor positioning system based on fingerprint, the tra- ditional fingerprint database construction method consumes much man- power and time cost. To solve this problem, we propose an effective method for constructing fingerprint database by using Microelectro Mechanical System (MEMS) to assist Bluetooth Low Energy (BLE), which overcomes the low efficiency of traditional methods. Meanwhile, the method achieves the comparable positioning accuracy and reduces workload more than 70%. In the optimization procedure, we use affine propagation clustering, outlier detection and filtering of Received Signal Strength Indication (RSSI) to optimize fingerprint database. Finally, the BLE positioning error conducted by the effective database is about 2 m.

Keywords: Indoor positioning

Fingerprint database construction methodãBLE

1 Introduction

With the development of wireless positioning technology, the demand for Loca- tion Based Service (LBS) is becoming popular. At present, the Global Positioning System (GPS) technology can not meet the needs of indoor positioning accuracy [1]. Aiming at the complex signal propagation environment, many indoor posi- tioning technologies have been proposed, such as Bluetooth [2], MEMS sensor [3]

and Wire Local Area Networks (WLAN) [4]. The cost of equipment, scalability and accuracy limit the development of WLAN positioning technology. Mean- while, the traditional fingerprint database constructing method receives RSSI at Reference Points (RPs), which is not suitable for large indoor scenes. The BLE technology is of low power consumption, low cost and short delay, which greatly reduces the cost of the BLE anchor. And compared with WLAN signal, the BLE signal is more stable.

To solve the inefficiency of traditional database constructing method, we pro- pose an effective BLE fingerprint database constructing system based on MEMS.

Firstly, the test staff holds the mobile phone and goes along the designated path, RSSI and MEMS data are uploaded to the server. The server tracks the change

c ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2018 X. Gu et al. (Eds.): MLICOM 2017, Part I, LNICST 226, pp. 133–141, 2018.

https://doi.org/10.1007/978-3-319-73564-1_14

of heading angle and signal peak of RSSI propagation model to determine the coordinate of the beacons, which are used for correcting Pedestrian Dead Reckon- ing (PDR). Then, we use affinity propagation clustering, outlier detection and RSSI filter to eliminate noise and generate fingerprint sub-database. Thus in positioning phase, we firstly determine the sub-database and then do fingerprint positioning, which also reduces the server load.

The remainder of the paper is organized as follows. Section2 reviews some related work about effective construction methods of fingerprint database. In Sect.3, we introduce the proposed algorithm in detail. Section4shows the exper- imental results. Finally, the conclusion is provided in Sect.5.

2 Related Work

In recent years, the effective construction methods of fingerprint database have been widely concerned. An automatic database construction system based on crowdsourcing is introduced in [5]. In the system, fixed landmark nodes, invisi- ble landmarks and particle filtering technology are used to correct crowdsourcing path. However, the system does not filter low quality data, which can not guar- antee that all fingerprints are valid. In order to reduce the labor cost of the off-line phase, a method is proposed to solve fingerprint identification in [6].

The off-line analytical fingerprint database is generated automatically by the server, which avoids the link of the actual survey and construction. However, this method requires a low SNR in environment, and the data from MATLAB simulation experiments are not verified by the actual project. In [7], the system uses a relative RSSI value vector group to replace the absolute RSSI value as the fingerprint data, but failed to solve the problem of low efficiency of traditional method. The main contribution of this paper is that we construct a more effec- tive and accurate fingerprint database by using beacons to constantly correct the PDR trajectory, which reduces the error of the fingerprint coordinate. At the same time, we also use clustering, sub-database generating and filtering algo- rithm to further optimize the fingerprint database, which improves the efficiency of online fingerprint matching.

3 Algorithm Description

Một phần của tài liệu Machine learning and intelligent communications part i 2017 (Trang 148 - 155)

Tải bản đầy đủ (PDF)

(722 trang)