使用迴路用電資訊進行非侵入式電器狀態監測與並行活動辨識

Non-intrusive Appliance Monitoring and Concurrent Activity Recognition from Circuit-Level Power Consumption

林谷原(Gu-Yuan Lin)

指導教授 ： 許永真

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2010年)

https://doi.org/10.6342/NTU.2010.02336

摘要

住家提供一個安全與舒適的環境讓人類可以在其中學習、休憩、通訊與遊樂。隨著家電產品的進步與生活品質的重視，住家的電器設備愈來愈多樣，家中的每個成員的用電量也隨之愈來愈高。家電的便利性與自動化，則造成愈來愈多使用者不重視用電習慣，甚至常常浪費能源而不自知。此外，由於現代的家庭生活中，大部份的居家活動皆與電器使用息息相關，這也顯示了利用電器運作狀態進一步推論居家活動之可行性。
我提出一套雙層方法架構，其以不打擾居住者之舒適度為前提，僅安裝迴路型電力計於家中電盤量測各迴路用電資訊，藉此監測家中各電器使用狀態並進一步推論使用者正在進行之並行活動(Concurrent activities)。利用以上兩者資訊可為家庭耗電情況進行分析，並提供更加有效與確實的節能服務。但實作此架構有兩個困難之處，首先，當有多種電器運作組合之總用電量非常相近時，僅根據用電波形容易造成誤判；再者，我們也需要證明僅採用電器使用狀態來推論多重並行活動是可行的。
針對第一個挑戰，我提出使用「階乘式條件隨機場模型」(Factorial Conditional Random Fields)以涵蓋多重電器使用狀態之同步關係(Co-temporal Relationship)加以區分多種相近總用電量之電器運作狀態組合;而針對第二個挑戰，除了同樣利用「階乘式條件隨機場模型」之外，並增加考慮先前電器狀態變化關係(Pre-local Relationship)以提升系統辨識並行活動之能力。
進行實驗時，於實驗室環境與真實個人宿舍布建電力計蒐集用電資訊與活動標記，以比較「階乘式條件隨機場模型」與「平行化條件隨機場」(Parallel Conditional Random Fields)及其他常用分類器(Classifiers)對於電器狀態監測與並行活動辨識之準確性，其中包括與「簡易貝氏分類器」(Naive Bayes)、「AdaBoost分類器」以及「支援向量機」(Support Vector Machine)之比較。實驗結果顯示，「階乘式條件隨機場模型」於以上兩種環境的電器監測實驗當中，皆獲得最高之聯合辨識率(Joint Accuracy)，如此驗證各電器使用之同步關係可幫助提升電器監測準確率；於真實宿舍環境中，雖然「階乘式條件隨機場模型」之辨識結果不如「支援向量機」，但藉由調整參考先前電器狀態變化之事件個數，可知向前參考多個歷史電器狀態變化可大幅提升並行活動之辨識能力，並得到98.35%之聯合辨識率，顯示了系統具有高度可行性。

關鍵字

階乘式條件隨機場模型 ； 平行化條件隨機場 ； 非侵入 ； 電器辨識 ； 活動辨識

並列摘要

With the innovation of appliances and the emphasis on the quality of life, the variety of appliances is increasing, which also results the increment of energy consumption. Due to the convenience and automation of appliances, residents are usually unaware of the waste of electricity. On the other hand, since most home activities are related to the use of appliances, this exhibits the possibility of using appliance operating states to further infer home activities.
I proposes a 2-layer framework which use the a non-intrusive way to install power meters to measure circuit-level power consumption and then use such consumption to monitor the appliance states and take one step ahead to recognize concurrent home activities. We can utilize such information to analyze the use of electricity and further provide more effective and practical power-saving services. However, there are two challenges for the implementation of the framework. First, when there are different appliance combinations consuming similar total power loads, distinguishing them only based on the power loads is easy to make erroneous prediction. Second, we want to verify the feasibility of just taking appliance states to infer concurrent activities.
To overcome the first challenge mentioned above, I advocate using the Factorial Conditional Random Fields(FCRFs) to accommodate the co-temporal relationships among multiple appliances and to discriminate those different appliance combinations with similar total power consumption. For the second challenge, I also employ FCRFs and additionally consider the pre-local relationships among the previous appliance state-change events to improve the capability of identifying concurrent activities.
In both a laboratory and a practical dormitory environment, I deploy power meters to measure the circuit-level power consumption and collect the annotations of activities for the comparisons of recognition accuracy between FCRFs, PCRFs, and other commonly used classiàers, including Naive Bayes, AdaBoost, and Support Vector Machine(SVM).
The experimental results of appliance recognition show that FCRFs has the highest joint accuracy in both environments, which demonstrates that the co-temporal relationships are helpful. On the other hand, in activity recognition, despite FCRFs perform badly than SVM in a real environment, the results still show that considering pre-local features dramatically improves the recognition accuracy. Furthermore, the best joint accuracy obtained from SVM is 98.35%, which demonstrates that the system has great feasibility in practical home environments.

並列關鍵字

Factorial Conditional Random Fields ； Parallel Conditional Random Fields ； Non-intrusive ； Appliance Recognition ； Activity Recognition

參考文獻

[30] H. Serra, J. Correia, A. Gano, A. de Campos, and I. Teixeira. Domestic power consumption measurement and automatic home appliance detection. In Intelligent Signal Processing, 2007 IEEE International Workshop on, pages 128–132, 2005.

[2] M. Berenguer, M. Giordani, F. Giraud-By, and N. Noury. Automatic detection of activities of daily living from detecting and classifying electrical events on the residential power line. In e-health Networking, Applications and Services, 2008. HealthCom 2008. 10th International Conference on, pages 29–32, 2008.

[3] M. Berges, E. Goldman, H. Matthews, and L. Soibelman. Learning Systems for Electric Consumption of Buildings. In Computing in Civil Engineering: Proceedings of the 2009 ASCE International Workshop on Computing in Civil Engineering: June 24-27, 2009, Austin, Texas, page 1. Amer Society of Civil Engineers, June 2009.

[4] J. Carmichael. Non-intrusive appliance load monitoring system. Epri journal, Electric Power Research Institute, pages 45–47, September 1990.

[5] C.-C. Chang and C.-J. Lin. LIBSVM: a library for support vector machines, 2001. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.