以深度學習應用於感測器數據預測及故障補值

Using Deep Learning Method for Data Prediction and Imputation in Sensor Fault Conditions

林沛寰(Pei-Huan Lin)

指導教授 ： 陳俊勳

國立交通大學/工學院/機械工程系所/碩士(2019年)

摘要

本論文採用深度學習應用在感測器來預測室內人數及感測器故障時之數據補值，並利用不同演算法比較和分析結果。數據源自於實驗場域(國立交通大學工程五館343室)自行架設各式感測器(可量測溫濕度、二氧化碳濃度及用電量)；首先以短期(一週)實驗及隨機森林(Random forest)演算法來調整感測器架設位置，接著進行長期(兩個月)實驗取得數據。在建立預測室內人數之模型中，使用支援向量機(Support Vector Machine)、遞歸神經網路(Recurrent Neural Network)、長短期記憶模型(Long Short-Term Memory)及Gated Recurrent Unit(GRU)，透過不同之超參數調整優化，並比較出最適合此研究之演算法，建立預測室內人數模型。接著以兩種情境模擬感測器故障狀況，以生成對抗網路(Generative Adversarial Network: GAN)及變分自編碼器(Variational Autoencoder; VAE)模擬停電情境並進行資料補值；以Gated Recurrent Unit訓練所有未故障感測器對單一故障感測器之相關度並進行補值，比較前述兩種故障情境以不同方式之補值的表現，並分析其結果。本文發現，相較於VAE，GAN擁有較好的穩定度，在不同的缺失時間下平均精度達到81%，而GRU對於單一故障感測器有很好的補值表現，平均精度可達到91%。除探討不同演算法在此研究中之表現，本研究也建立一流程，利用深度學習進行感測器架設、預測及補值。

關鍵字

感測器 ； VAE ； GAN ； GRU ； 補值

並列摘要

This thesis utilizes the deep learning techniques to predict the number of occupants in the experimental field and data imputation of the sensor failures and to compare and analyze the results obtained by using different algorithms. The sensors were set up in the experimental field, located at room 343, Engineering Building V at National Chiao Tung University, to measure the temperature and humidity, carbon dioxide concentration and electric power consumption. The random forest method was adapted to evaluate the sensor importance rankings according to one-week collected data in order to optimize the allocation of the sensors. After that, two months of data collection were followed. To establish a model for predicting the number of people in the room, Support Vector Machine(SVM), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) are used. The corresponding results were compared to determine the most suitable algorithm for this model. Then, two sensor fault conditions were simulated by using the Generative Adversarial Network (GAN) and the Variational Autoencoder (VAE), respectively, to impute the missing data in the scenario of the power break down. GRU was used to impute the missing data in the scenario of occurrence of a single malfunction sensor. Comparing and analyzing the performance of the above two scenarios and results, it is found that GAN has the better stability with an average accuracy of 81% compared to those of VAE, whereas GRU has a good performance for the condition of single sensor fault with an average accuracy of 91%. In addition to exploring the performance of several different algorithms used in this thesis, it also establishes a process by utilizing the deep learning techniques for the optimal location, prediction, and data imputation of sensors.

並列關鍵字

Sensor ； Variational Autoencoder ； Generative Adversarial Network ； Gated Recurrent Unit, ； Imputation

參考文獻

[1] 裴有恆、林祐祺, 物聯網的無限商機. 碁峯資訊, 2007.

[2] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, "A survey on sensor networks," IEEE Communications magazine, vol. 40, no. 8, pp. 102-114, 2002.

[3] K. S. Perera, Z. Aung, and W. L. Woon, "Machine learning techniques for supporting renewable energy generation and integration: a survey," in International Workshop on Data Analytics for Renewable Energy Integration, 2014: Springer, pp. 81-96.

[4] H. A. Kazem, J. H. Yousif, and M. T. Chaichan, "Modeling of daily solar energy system prediction using support vector machine for Oman," International Journal of Applied Engineering Research, vol. 11, no. 20, pp. 10166-10172, 2016.

[5] M. Caudill and C. Butler, "Understanding neural networks: computer explorations–volume 1 basic networks," ed: Massachusetts: The MIT Press, 1993.