車用電動天窗之異音偵測

Abstract

現行車用天窗品質檢驗多仰賴人員經驗、觸覺或目視判斷，易受主觀因素影響，評估結果缺乏一致性與客觀性。為提升檢測精準度，本研究提出一套以訊號處理為核心的異音偵測方法，建立客觀且可靠之異音指標，並實現檢測流程自動化。 本研究之檢測音訊係由車用電動天窗製造廠提供與標註。該製造廠在品檢程序中，天窗共進行四個動作，分別為天窗玻璃由關閉狀態以水平移動至開啟位置，由開啟位置以水平移動回到關閉狀態，關閉狀態上抬至通風位置，由通風位置回到關閉狀態。本研究所收集之音訊，天窗異音皆出現在天窗玻璃由關閉狀態以水平移動至開啟位置。此動作可細分為4個時段，第1，2時段可能因滑塊與滑槽搭配不佳產生異音，稱為NG1、NG2；第3時段可能因彈簧未定位好產生異音，稱為NG3；第4時段則可能因擋風板回彈產生異音，稱為NG4；另外還有橫跨各時段，因拉繩摩擦所產的異音，稱為NG5。 本研究比較三個常用的音訊時頻分析，包含短時傅立葉轉換(Short-Time Fourier Transform, STFT)、梅爾頻譜(Mel Spectrogram)及梅爾倒頻譜係數(Mel-Frequency Cepstral Coefficients, MFCC)。由於異音發生時，音訊能量會變大，故除了MFCC，STFT能量譜與梅爾頻譜皆可觀察到異音所造成的亮帶，其中以STFT能量譜的亮帶更為清晰，故本研究採用STFT對前述5類異音分別進行分析。 為掌握異音的特徵頻帶，我們先觀察各類異音最大聲響樣件之STFT亮帶所對應的頻帶，以此頻帶對訊號做帶通濾波，經品保師聆聽後確定此頻帶之聲響即為我們所關注的異音。接著，以帶通濾波之訊號繪製能量歷時曲線訂定異音長度。為精準的擷取實際異音發生區間，我們將掃描窗大小設定為異音長度，對時間域訊號做掃描，並計算所有掃描窗之特徵頻帶能量積分值。而當異音出現在某一掃描窗時，該掃描窗將具備最大能量值，本研究將此能量值定義為異音指標，作為後續分析異音的依據。 計算出所有訊號之異音指標後，我們觀察到正常(OK)、異常(NG)訊號會出現能量差異，而欲決定一音訊是否存在異音，必須先選擇合理的閥值，本研究採用兩種閥值：第一種是以NG訊號中最小的異音指標作為閥值，採用此閥值不會將NG訊號誤判為OK；第二種是以OK訊號中最大的異音指標作為閥值，採用此閥值不會將OK訊號誤判為NG。經測試，以第一種閥值做分類，NG1~NG5之準確率分別達97.7%、97.3%、98%、100%、100%；以第二種閥值做分類，NG1~NG5之準確率分別達99.2%、94.9%、98.4%、100%、100%。證實特徵頻帶能量積分值為有效之異音指標，證實此異音指標之有效性。 由於前述之異音特徵頻帶係由人為觀察STFT能量譜決定，為進一步優化異音特徵頻帶，本研究發展出一系統化方式，除能自動搜尋準確率最佳的特徵頻帶，還同時考慮了該頻帶的穩健性。結果顯示，以第一種閥值做分類，NG1~NG5之準確率分別達100%、96.9%、100%、100%、100%；以第二種閥值做分類，NG1~NG5之準確率分別達100%、99.6%、100%、100%、100%。整體而言，異音偵測的準確率確有提升。 最後，本研究對異音指標進行泛化性測試。發現將此異音指標套用於不同型號之天窗訊號時，檢測時可能產生誤判，這主要是因為不同型號天窗的異音特徵頻帶與本研究之異音特徵頻帶不同。因此對不同型號之天窗進行異音檢測，需以本研究之方法重新分析訊號。 針對特定機種，本研究所提出之異音偵測方法可有效分辨OK與NG訊號。未來在實務應用上，除了可以發展為自動檢測系統，輔助品保師有效率地執行品保工作，亦可利用各類異音的特徵生成大量的訊號，作為品保人員的訓練素材。

Current quality inspection of automotive sunroofs heavily relies on human experience, tactile feedback, or visual judgment, which are easily affected by subjective factors and lack consistency and objectivity. To improve detection precision, this study proposes a signal-processing-based anomaly detection method. The approach establishes an objective and reliable acoustic anomaly index and realizes an automated inspection workflow. The audio signals analyzed in this study were labeled and provided by the cooperating manufacturer. According to their inspection procedure, sunroof operation is divided into four phases: (1) horizontal movement from fully closed to fully open, (2) returning from open to closed, (3) tilting up from the closed position to a ventilation position, and (4) returning from ventilation to fully closed. All abnormal sounds in this study occurred during the first phase. This phase can be further segmented into four time intervals: NG1 and NG2 are associated with poor fitting between sliders and guide rails in intervals 1 and 2; NG3 arises from spring misalignment in interval 3; NG4 results from wind deflector rebound in interval 4; and NG5 spans all intervals, caused by rope friction. Three time–frequency analysis methods are compared: short-time Fourier transform(STFT), Mel spectrogram, and Mel-Frequency Cepstral Coefficients (MFCC). Since anomalies typically exhibit increased signal energy, both the STFT power spectrogram and Mel spectrogram reveal bright bands corresponding to these events. The STFT power spectrogram provides clearer visual features and is therefore adopted in this study for analyzing the five types of anomalies. To identify the characteristic frequency bands of each type of anomaly, STFT power spectrograms of the loudest samples were observed to locate bright bands. Then, bandpass filters were designed based on the characteristic frequency bands and applied to the audio signals. After auditory confirmation by quality assurance engineers, these bands were verified to represent the target anomalies. The filtered signals were used to construct energy envelopes for estimating anomaly durations. A sliding window, set to the estimated anomaly length, was applied to the time-domain signal to compute the energy within the characteristic band for each window. The window with maximum energy was defined to contain the anomaly, and this energy was defined as the anomaly index. After computing the anomaly index across all signals, a clear distinction between normal (OK) and abnormal (NG) signals was observed. Two thresholding strategies were employed: (1) the minimum anomaly index among NG signals, ensuring no NG is misclassified as OK; and (2) the maximum anomaly index among OK signals, ensuring no OK is misclassified as NG. Using the first threshold, classification accuracies for NG1–NG5 were 97.7%, 97.3%, 98.0%, 100%, and 100%, respectively; with the second threshold, the accuracies reached 99.2%, 94.9%, 98.4%, 100%, and 100%. These results confirm the effectiveness of the proposed anomaly index. Since the initial characteristic bands were determined manually, this study further develops a systematic approach to automatically search for optimal frequency bands, considering both classification accuracy and robustness. The optimized bands yielded improved accuracies: 100%, 96.9%, 100%, 100%, and 100% (first threshold), and 100%, 99.6%, 100%, 100%, and 100% (second threshold). The results demonstrate enhanced detection performance. Finally, a generalization test was conducted using signals from different sunroof models. Misclassification occurred due to model-specific differences in anomaly frequency bands, suggesting that the proposed method must be re-applied for each model. Nevertheless, for a given sunroof type, the method effectively distinguishes OK and NG signals. In practical applications, the proposed technique can be developed into an automated detection system to assist quality assurance engineers and generate diverse anomaly samples as training data for quality control personnel.