工業 4.0 推動製造業朝向智慧化與數據驅動的生產模式發展，設備維護策略也逐漸由傳統的事後維修與定期保養，轉向能夠提前預警設備異常的預測性維護。預測性維護的核心在於透過感測器持續蒐集設備運行過程中的工業時間序列訊號，並利用資料分析與深度學習模型辨識潛在的劣化趨勢。此類訊號同時包含時域中的瞬時波形變化與頻域中的特徵頻率資訊，兩者共同構成機械故障診斷的重要物理基礎，也使時間序列故障診斷成為智慧製造與設備健康管理中的關鍵研究議題。 然而，深度學習方法由實驗室走向真實工業場域時，仍面臨多項限制。實際產線雖能累積大量無標籤感測資料，但高品質故障標籤稀缺且取得成本高；既有半監督學習方法多源自影像領域，其資料增強與一致性假設不易直接套用於具有明確物理意義的時間序列訊號；在少量標籤與高雜訊條件下，類別原型也容易受到錯誤標籤或極端樣本影響而產生偏移；此外，時域與頻域資訊雖具互補性，現有方法仍未充分利用頻域較高的物理穩定性來校準時域模型。為解決上述問題，本研究提出中心點雙域半監督故障診斷模型（Medoid-based Multidomain Consensus， MDC），以頻域為教師、時域為學生建立非對稱時頻對齊架構，並結合跨域共識閘門、Top-K Medoid 穩健原型更新，以及關注點分離的樣本-原型對比損失，提升少標籤條件下故障診斷模型的可靠度。 本研究於 CWRU 軸承故障公開資料集與 D-Electronics 半導體封裝廠 wire-bonding 製程真實感測資料集上驗證 MDC。CWRU 用於檢驗模型在標準且相對平衡之故障診斷任務上的競爭力；D-Electronics 則具有原始 Pass 與 Fail 比率高達 1308:1 的極端不平衡特性，用於評估模型在真實產線少數類偵測情境下的實用性。實驗結果顯示，MDC 在 CWRU 上與主流半監督方法表現相當，並於 D-Electronics 上在 macro F1 指標較七個基準模型領先 6.7 至 12.9 個百分點。上述結果顯示，MDC 能在標籤稀缺與類別不平衡的工業情境中有效提升少數故障類別的辨識能力，並具備推論階段僅需部署時域模型的實務落地可行性。;Industry 4.0 has accelerated the shift toward intelligent and data-driven manufacturing， making predictive maintenance an important strategy for detecting equipment abnormalities before failure occurs. Predictive maintenance relies on industrial time-series signals collected from sensors and uses data analysis and deep learning models to identify potential degradation patterns. These signals contain both time-domain waveform variations and frequency-domain characteristic information， forming an important physical basis for mechanical fault diagnosis. However， applying deep learning to real industrial environments remains challenging. Production lines generate large amounts of unlabeled sensor data， but high-quality fault labels are scarce and costly. Existing semi-supervised methods are mostly designed for images and are difficult to transfer directly to time-series signals with clear physical meanings. In addition， class prototypes may drift under limited labels and noisy conditions， while the physical stability of frequency-domain information has not been fully used to calibrate time-domain models. To address these issues， this study proposes the Medoid-based Multidomain Consensus (MDC) model， an asymmetric semi-supervised framework that uses the frequency domain as the teacher and the time domain as the student. MDC integrates a cross-domain consensus gate， Top-K Medoid robust prototype updating， and a separation-of-concerns sample-to-prototype contrastive loss to improve fault diagnosis reliability under limited-label conditions. MDC is evaluated on the CWRU bearing fault dataset and the D-Electronics wire-bonding dataset collected from a real semiconductor packaging line. CWRU is used to test performance on a standard and relatively balanced benchmark， while D-Electronics， with an original Pass-to-Fail ratio of 1308:1， evaluates minority fault detection in a real production setting. Results show that MDC performs comparably to mainstream semi-supervised methods on CWRU and outperforms seven baselines by 6.7 to 12.9 percentage points in macro F1 on D-Electronics. These findings demonstrate that MDC improves minority fault recognition under label scarcity and class imbalance， while requiring only the time-domain model during inference.
