結合馬氏田口系統與類神經網路分析法改善多感測器火災異常偵測績效

Abstract

為了改善多感測器火災偵測器的警報速度及準確度，本研究提出結合馬氏田口系統與類神經網路的LSTM-MTS方法。LSTM-MTS方法以馬氏田口系統作為主體進行火災異常偵測，但由於火災煙霧資料在正常狀態下，具有數值穩定且無高低變動的特性，導致無法建立正常狀態馬氏空間，因此本研究提出結合類神經網路中的長短期記憶，利用其善於處理時間序列資料的特性，將正常狀態資料進行轉換，以順利建構正常狀態馬氏空間。為了驗證本研究提出的LSTM-MTS方法，是否能夠有效提升多感測器火災偵測器的警報速度及準確度，本研究使用煙霧偵測器研發製造商A公司，以及美國國家標準暨技術研究院的火災實驗資料進行分析後，證實本研究提出的LSTM-MTS方法能夠有效提升多感測器火災偵測器的偵測績效，提早在火災偵測器實際警報之前偵測到異常，並且相較於單獨使用馬氏田口系統及類神經網路的方式，具有較佳的警報速度與準確度。To improve the alarm speed and accuracy of multi-sensor fire detectors, this study proposes the LSTM-MTS method, which combines the Mahalanobis-Taguchi System with Long Short-Term Memory (LSTM) neural networks. The Mahalanobis-Taguchi System serves as the primary method for fire anomaly detection. However, due to the stable and non-volatile nature of fire smoke data under normal conditions, it is challenging to establish a normal state Mahalanobis space. This study addresses this issue by integrating LSTM, which excels in handling time series data, to transform normal state data, thereby successfully constructing the normal state Mahalanobis space. To validate the effectiveness of the proposed LSTM-MTS method in enhancing the alarm speed and accuracy of multi-sensor fire detectors, this study analyzes data from smoke detector manufacturer Company A and fire experiment data from the National Institute of Standards and Technology (NIST). The results confirm that the proposed method significantly improves the performance of multi-sensor fire detectors, enabling earlier anomaly detection before the actual alarm triggers. Compared to using the Mahalanobis-Taguchi System or neural networks alone, the LSTM-MTS method demonstrates superior alarm speed and accuracy.