內藏式主軸智慧化診斷系統應用在品質控制之研究

Abstract

內藏式主軸為工具機內高單價且關鍵之核心元件，若能於出廠前確保其製造品質，便可顯著降低交付後因短期異常而引發之問題，進而減少整體經濟損失。因此，建立可於出廠前辨識主軸品質狀態與風險之方法，已成為製造端極需面對之關鍵課題。 本研究聚焦於內藏式主軸跑合工站之無切削負載振動訊號，提出一套可用於出廠品質管制之兩階段智慧化診斷方法，並以產線同型號內藏式主軸 13 支之固定轉速的外殼加速度量測資料進行建構與驗證。由於實務產線往往難以事先確知潛在不良模式，其對應之訊號特徵亦可能微弱且不明確，使得以預先定義故障類別並蒐集充足標註樣本為前提之監督式建模策略難以直接應用，故本研究採取先無監督及後監督之流程設計，以降低對標註資料的依賴並提升可部署性。 第一階段以振動訊號之頻域特徵為核心，透過降維與分群挖掘潛在品質狀態，並以 Davies–Bouldin 指標評估分群品質。為使分群結果能對應實務品管語意，進一步將分群結果與出廠前品質管制紀錄、抽樣耐久性試驗結果及出廠後一年內返廠維修資訊交叉比對，並結合工程師領域知識，建立具工程語意之品質標籤，作為第二階段監督式模型之訓練基礎。第二階段針對小樣本與類別不平衡之產線資料，將時域振動訊號建模為圖結構，引入圖注意力網路進行監督式多類別分類，並與 k-近鄰、支持向量機及圖卷積網路比較。為避免模型效能僅反映同一主軸資料分割之結果，本研究採留一交叉驗證法，以檢驗模型對未知主軸個體之可遷移性與泛化能力；並進一步透過不同訊噪比情境與資料平衡化策略，評估模型於量測雜訊顯著且故障樣本稀缺之條件下之穩健性。 研究結果顯示，在既有品質管制全數放行之前提下，所提流程仍能自出廠前短期振動資料中辨識出與耐久性失效及一年內返修風險相關之高風險個體，凸顯出廠測試訊號於風險辨識上之潛在價值。為促進方法落實於實務流程，本研究進一步建立智慧化品質診斷系統原型，整合資料庫、診斷模型與網頁介面，使檢測人員得以在既有出廠前檢測流程中，直接參考模型輸出之品質評估資訊。綜合而言，本研究以出廠前品質管控為核心，提出並驗證一套適用於標籤匱乏情境之兩階段診斷流程，並透過泛化與抗噪條件之設計評估其工業可用性，初步證實其於內藏式主軸出廠品質判定之應用潛力。

The built-in spindle is a high-value and mission-critical core component in machine tools. If its manufacturing quality can be assured prior to shipment, issues arising from early-life anomalies after delivery can be substantially reduced, thereby mitigating overall economic losses. Accordingly, developing a method capable of identifying spindle quality conditions and associated risks before shipment has become a critical challenge that must be addressed on the manufacturing side. This study focuses on no-load vibration signals acquired at the built-in spindle run-in station and proposes a two-stage intelligent diagnostic framework for outgoing quality-control (QC). The framework is developed and validated using fixed-speed housing acceleration measurements collected from 13 identical built-in spindles on a production line. In practice, potential defect modes are often unknown in advance, and their signal signatures may be weak or ambiguous, which limits the direct applicability of supervised modeling strategies that require predefined fault categories and sufficiently labeled samples. To reduce reliance on labeled data and improve deployability, this research adopts a sequential unsupervised-then-supervised design. In Stage I, frequency-domain features are extracted from vibration signals and used to uncover latent quality states via dimensionality reduction and clustering, with the Davies–Bouldin Index employed to assess clustering quality. To align clustering outcomes with practical QC semantics, the discovered clusters are further cross-validated against pre-shipment QC records, results from sampled durability tests, and post-shipment repair/return information within one year, together with domain knowledge from experienced engineers. This process yields engineering-interpretable quality labels that serve as the training basis for Stage II. In Stage II, to address small-sample and class-imbalance characteristics commonly seen in production data, time-domain vibration signals are modeled as graph-structured data and a Graph Attention Network (GAT) is introduced for supervised multiclass classification. The proposed approach is benchmarked against k-nearest neighbors (kNN), support vector machines (SVM), and graph convolutional networks (GCN). To ensure that performance does not merely reflect sample-level splits from the same spindle, leave-one-out cross-validation is conducted to evaluate transferability and generalization to unseen spindle units. Robustness is further examined under different signal-to-noise ratio (SNR) scenarios and data rebalancing strategies, reflecting conditions with substantial measurement noise and scarce failure samples. The results demonstrate that, even when conventional QC criteria would release all units, the proposed framework can identify high-risk spindles from short-duration pre-shipment vibration data—spindles that are subsequently associated with durability-test failures and increased repair risk within one year. This finding highlights the latent value of outgoing test signals for early risk identification. To facilitate practical adoption, this study also develops a prototype intelligent quality diagnosis system integrating a database, diagnostic models, and a web-based interface, enabling inspectors to directly reference model-derived quality assessments within existing pre-shipment inspection workflows. Overall, this research proposes and validates a two-stage diagnostic pipeline tailored to label-scarce manufacturing settings, and evaluates its industrial feasibility through generalization- and noise-aware designs, thereby providing initial evidence of its applicability for outgoing quality assessment of motorized spindles.