| 摘要: | 本文針對市售電動車提出一套建構於有限訓練資料條件下之智慧故障診斷系統建置方法，旨在因應電動車實際故障案例稀少之情況，發展線上監測之智慧故障診斷系統。該智慧故障診斷系統可透過車載物聯網資訊，即時監控車輛上各次系統之運作狀態，當偵測出異常時則立即發出告警訊號，並提供具體之診斷維修建議。為實現此目標，本研究設計兩組診斷模組，分別稱為次系統工況診斷模組(Subsystem Operating Condition Detection Module, SCDM)與故障源診斷模組(Root Cause Diagnosis Module, RCDM)。SCDM採用動態響應監測法，針對各次系統之關鍵訊號建構響應模型，並與實際量測值比較，以獲得響應誤差。再者，同時以響應誤差與原始訊號數據作為時序特徵，進行關鍵訊號之工況判別。其判別方法係以時序資料作為特徵，經由特徵擷取技術整合成單一綜合指標，接續進行工況分類。最終，綜合上述分類結果進行最終的異常判定，並發出對應之故障警報訊號。另一方面，RCDM亦採用動態響應監測法，但著重於多組關鍵訊號之響應誤差，透過時序特徵擷取評估訊號之異常機率。接續，基於異常機率進一步評估次系統所有潛在故障模式之風險，計算對應之風險分數。最終，依據風險分數產出最終的診斷維修建議。
本研究中使用總計224筆盲樣資料進行診斷性能測試。其中，每筆資料對應車輛於單一日內的完整運行紀錄，彙整後所涵蓋總樣本數約為60.2萬個。根據測試結果，SCDM在馬達散熱系統工況分類任務中的Macro F1分數為98.36%。另一方面，根據20筆異常工況資料作為盲樣測試集進行RCDM診斷測試，結果顯示所有案例皆被優先判定為水路氣阻過大，即使其中包含1筆實際回報為水泵故障之案例。此現象說明目前所收集的故障案例仍有所不足，且案例類型單一、數量比例亦不均衡，已對RCDM的泛化能力造成影響。此外，本研究亦針對所開發之智慧故障診斷系統，設計一套使用者互動介面。該介面整合前端網頁互動功能與後端診斷伺服器架構，以協助使用者即時掌握車輛上各次系統的運行狀態與潛在異常情形，進一步提升診斷系統於實務層面的應用價值。綜上所述，本研究提出之智慧故障診斷系統架構，即使在可取得之訓練資料有限的條件下，仍能展現優異的工況分類表現，未來可進一步應用於實際故障維修場域中，作為輔助第一線工程師進行異常判別與故障排查之工具。
;This study proposes an intelligent fault diagnosis system tailored for commercial electric vehicles (EVs) under conditions of limited training data, aiming to address the scarcity of real failure cases and enable online monitoring capabilities. The system utilizes vehicle onboard IoT information to continuously monitor the operational status of various subsystems in real time. When an anomaly is detected, the system promptly issues warning signals and provides specific diagnostic and maintenance suggestions.
To achieve this objective, two diagnostic modules are developed: the Subsystem Operating Condition Detection Module (SCDM) and the Root Cause Diagnosis Module (RCDM). The SCDM adopts a dynamic response monitoring approach to construct response models for critical signals within each subsystem. These models are compared with actual measurements to compute response errors, which, together with raw signal data, are treated as temporal features for operating condition classification. These temporal features are further extracted and integrated into a unified indicator, which is then used for condition classification. The final abnormality decision is made by aggregating the classification results, upon which the system triggers corresponding fault alarms. Conversely, the RCDM also employs dynamic response monitoring but focuses on the response errors across multiple key signals. It extracts temporal features to evaluate the probability of signal anomalies, then estimates the risk associated with all potential subsystem fault modes based on these probabilities and calculates corresponding risk scores. Diagnostic recommendations are ultimately generated based on these risk scores.
A total of 224 blind test samples were used to evaluate diagnostic performance, where each sample corresponds to a full day of vehicle operation. The compiled dataset includes approximately 602,000 instances. Experimental results show that the SCDM achieved a Macro F1-score of 98.36% in motor cooling subsystem condition classification. For the RCDM, 20 blind samples of abnormal operating conditions were tested, with all cases being primarily diagnosed as “excessive air resistance in the cooling circuit,” even though one case was actually reported as a water pump failure. This suggests that the current failure case dataset is insufficient in both diversity and quantity, which limits the generalizability of the RCDM. |
