NILM (Non-Intrusive Load Monitoring)ã®è«æç´¹ä»ã§ãã
ã¿ã¤ãã«ããåããããã«ååç´¹ä»ããBOLTãçºå±ããããã®ã§ããBOLTã«ã¤ãã¦ã¯ååã®è¨äºããåç §ãã ããã
è«æ㯠(PDF) Dyna-Bolt: Domain Adaptive Binary Factorization Of Current Waveforms For Energy Disaggregation ããå ¥æã§ãã¾ãã
æ¦è¦
åé¡è¨å®
以ä¸ã®ããã«ãªã£ã¦ãããSemi-supervisedã®ã¿ã¹ã¯ã§ãã
- ãã建ç©
- 主幹è¨æ¸¬ã®æ³¢å½¢ãã¼ã¿ããã£ã¦ãã
- å家é»ã®åå¥è¨æ¸¬ãã¼ã¿ããã£ã¦ããï¼ã©ãã«ãã¼ã¿ããï¼
- å¥ã®å»ºç©
- 主幹è¨æ¸¬ã®æ³¢å½¢ãã¼ã¿ããã£ã¦ãã
- å家é»ã®åå¥è¨æ¸¬ãã¼ã¿ããã£ã¦ããªãï¼ã©ãã«ãã¼ã¿ãªãï¼
ã¨ããæ¡ä»¶ã«ããã¦ãããå¥ã®å»ºç©ã®ä¸»å¹¹è¨æ¸¬ãã¼ã¿ãã家é»ã®åé¢ãè¡ãããã ã¨ããã¿ã¹ã¯ã§ãã
åé¡æè
åã家é»ã§ãã£ã¦ãç°ãªã建ç©å士ã§ã¯é»æµæ³¢å½¢ãå¾®å¦ã«éã£ã¦ããããã£ã¦BOLTãåç´ã«é©ç¨ãã¦ããã¾ããããªã ã¨ããã®ãåé¡æèã§ãã
ä¾ãã°ã以ä¸ã®å³ã«ããã¦2ã¤ã®å®¶åºãHouse3ã¨House5ã®åã ã®Refrigeratorã¨Furnaceï¼é»æ°ãã¾ã©ï¼ï¼ã®é»æµæ³¢å½¢ã¨é»åã«ã¼ãã示ããã¦ãã¾ããäºãã«å¾®å¦ã«éã£ã¦ãã¾ãã
House3ã®ä¸»å¹¹ãã¼ã¿ã§å¦ç¿ãããã¢ãã«ããã®ã¾ã¾House5ã«é©ç¨ãã¦ã精度ãè½ã¡ã¦ãã¾ããã¨ã¯å®¹æã«æ¨æ¸¬ã§ãã¾ãã
DYNA-BOLTã§ä½ããã£ãã
åãã©ãã«ããã£ã¦ãããã©ç¹å¾´ï¼æ³¢å½¢ã®å½¢ç¶ï¼ãå°ãç°ãªã£ã¦ããã¨ãããã¨ã§ããã¡ã¤ã³ã·ããã®åé¡ã¨æãããã¨ãã§ãã¾ãã ãã ããæ¬è«æã«ããã¨Domain Adaptation Neural Network (DANN)ã®ææ³ãé©ç¨ããã ãã§ã¯å¹æã¯ä¸ååã¨ãããã¨ã§ããã
ããã§ãDyna-BOLTã§ã¯ã©ãã«ãã¼ã¿ã使ã£ã¦ã®æ©å¨ã®èå¥ã¿ã¹ã¯ã追å ãããã¨ã§æ¹åãå³ã£ã¦ãã¾ãã
ã¢ã¼ããã¯ãã£
ã¢ã¼ããã¯ãã£ã¯ä»¥ä¸ã®å³ã®éãã§ãã
Encoder~Binary Activation~Decoder ã®é¨åãBOLTã¨åãæ§é ã«ãªãã¾ãã Decoder 㯠åãªã1層ã®å ¨çµå層 $Wx$ ã§ããããã®éã¿è¡å $ W $ ï¼æ¬¡å : $(c, N)$ï¼ã cåã®Activationã«å¯¾å¿ããN次å 波形ã表ç¾ããå ¥å波形ã¨Decoderåºåã§ããåæ§æ波形ã¨ã®å·®ã åæ§æ誤差 $ L_{recon} $ ã«ãªãã¾ãã
Dyna-BOLTã§ã¯ã$ W $ ã¯ã½ã¼ã¹ãã¡ã¤ã³ç¨ã® $ W_{source} $ 㨠ã¿ã¼ã²ãããã¡ã¤ã³ç¨ã® $ W_{target} $ ã®2種é¡åå¨ããå ¥åãã¼ã¿ã®ç¨®é¡ã«ãã£ã¦ä½¿ãåãã¾ãã
$ W_{source} $ 㨠$ W_{target} $ ã¯å¯¾è±¡ã¨ãªããã¡ã¤ã³ãç°ãªã£ã¦ããããå¥ã®éã¿è¡åã¨ãªã£ã¦ãã¾ããã©ã¡ããåã家é»ã«å¯¾ãã波形ãã¿ã¼ã³ã表ãã¯ããªã®ã§éã¿ã¯äºãã«è¿ããªãã¹ãã§ãã
$ L_{difference} $ ã¯ãã®å¶ç´ã表ç¾ããæ失ã«ãªãã¾ãã
$$ L_{difference} = \sum_{i=1}^{m} d^2 (W_{source, i}, W_{target, i}) $$
$ L_{domain} $ 㯠Domain Classifier ã®èå¥æ失ã§ãã Domain Classifierã¯ã½ã¼ã¹ãã¡ã¤ã³ã®å ¥å波形ã¨ã¿ã¼ã²ãããã¡ã¤ã³ã®å ¥å波形ãèå¥ããããã«è¨ç·´ããã¾ãããããããã¨ã¯å»ºç©éã®å ±éã®ç¹å¾´ãã¨ã³ã³ã¼ãã¼ããæ½åºãããã¨ãããã¨ã§ãããã£ã¦ãGradient Reversalã¨æ¸ãã¦ããããã«å¾é ãå転ããã¦éä¼æ¬ããã¾ããããã«ãããã¨ã³ã³ã¼ãã¼ã¯ç°ãªããã¡ã¤ã³ã®å ¥åãèå¥ãããã¨ããã®ã§ã¯ãªããå ±éã®ç¹å¾´ããã¤ããã«å¦ç¿ããã¾ãã
æå¾ã«$ L_{class} $ ã§ãã Appliance Classifier ã«ããæ©å¨èå¥ã®æ失ã§ããæ©å¨ã®ã©ãã«ãå¿ è¦ãªã®ã§ã½ã¼ã¹ãã¡ã¤ã³å ¥åã«å¯¾ãã¦ã®ã¿è¨ç®ããããã®ã«ãªãã¾ãã
ãããã足ãåãããæçµçãªæ失é¢æ°$ L $ã¯ä»¥ä¸ã®ããã«ãªãã¾ãã
$$ L = L_{recon} + \gamma L_{class} + \lambda L_{domain} + η L_{difference} $$
å®é¨
å®é¨ã§ã¯ãREDDãã¼ã¿ã»ããã®2家åºã使ç¨ãã¦ãã¾ãã
- House3ï¼ ã½ã¼ã¹ãã¼ã¿ãã©ãã«ãã
- House5ï¼ã¿ã¼ã²ãããã¼ã¿ãã©ãã«ãªã
çµæã¯ä»¥ä¸ã®è¡¨ã®ããã«ãªã£ã¦ãã¾ãã
Supervised 㯠House3ã«BOLTãé©ç¨ãããã®å¾ãã©ãã«ãã¼ã¿ã使ã£ã¦æ©å¨åé¡ããçµæã§ããLogistic Regressionã«ããBinary Activationãå ¥åãã¼ã¿ã¨ããã©ãã«ã使ã£ã¦å¦ç¿ãã¦ããã¨èãããã¾ãã
ããã¦ãSemi-supervised㯠House3, House5ã«é¢ãã¦BOLTãé©ç¨ããå¾ãHouse3ã®ã©ãã«ãã¼ã¿ã使ã£ã¦æ©å¨èå¥ãå¦ç¿ããçµæã§ãã
ã©ã¡ãã®ã±ã¼ã¹ã«ããã¦ãHouse3ã®F1ã¹ã³ã¢ã¯ããªãé«ããã©ï¼0.98~0.99ï¼ãHouse5ã®F1ã¹ã³ã¢ãã¨ã¦ãä½ãï¼0.21~0.23ï¼ãã¨ããããã¾ãã
çç±ã¨ãã¦ã¯ t-SNEã«ããåãè¾¼ã¿ãã¯ãã«ç©ºéã2家åºéã§ã»ã¨ãã©ãªã¼ãã¼ã©ãããã¦ãªãã家åºéã®å ±éã®ç¹å¾´ãæãããã¦ãªãããã§ãã
以ä¸ã®å³ã®(a)ãè¦ãã°ãããã¾ãããHouse3ã¨House5ã«ã¯ç©ºéçãªéãªããã»ã¨ãã©ããã¾ããã
Vanilla DANN㯠DANN (Domain Adaptation Neural Network)ãé©ç¨ãããã®ã«ãªãã¾ãããHouse5ã®F1ã¹ã³ã¢ã¯0.23â0.32ã®å¾®å¢ã«çã¾ãã¾ãã
(b)ã®å³ãè¦ãã¨å ¨ä½çãªåå¸ã¯Domain classifierãå°å ¥ããå¹æã«ããä¼¼ãããã«ãªã£ã¦ã¾ãããç¢å°é¨åã«çç®ããã¨House3ã¨House5ã¨ã§å¥ã®è²ï¼å®¶é»ï¼ã«ãªã£ã¦ãã¾ã£ã¦ãã¾ãã
ãã㧠Appliance classificationã®ã¿ã¹ã¯ãå ããDyna-BOLTã¨ãªãã¾ããDyna-BOLTã§ã¯House5ã®F1ã¹ã³ã¢ã0.55ã¾ã§æ¹åãã¦ãã¾ãã
Appliance classificationã®ã¿ã¹ã¯ã«ã¤ãã¦ãMSEãã¹ãããã«ãã©ãã«ã§ã®ã¯ãã¹ã¨ã³ãããã¼ãã¹ã®ã»ããæ§è½ãè¯ãããã§ãã
ææ
BOLTã®è«æã§è¨åããã¦ããããã«ãã§ã¼ã³å±éãã¦ãããã¡ã¼ã¹ããã¼ããã¹ã¼ãã¼ãã¼ã±ããã¯åºèéã§ä¼¼ããããªå®¶é»ã»æ©å¨ããã£ã¦ããã¨èããããã¾ããæ°åºèã«ããã¦æ£è§£ãã¼ã¿è¨æ¸¬ããããã¨ã¯ç¾å®çã§ãã
å®é¨ã®çµæããSemi-supervisedã¨Dyna-BOLTã®ã¿ã¼ã²ãããã¡ã¤ã³å®¶åºã§ã®æ§è½ã®å·®ã¯æç½ã§ãããä¸è¿°ã®è¨å®ã§ããã°Dyna-BOLTã使ããªãæã¯ãªãã§ãããã