麥克風陣列安裝角度誤差對刀具磨耗偵測之預測準確度的影響

Abstract

在銑削加工中，刀具的狀態是一個重要的指標，且會影響整體加工的品質，所以刀具磨耗的偵測系統就格外重要。雖然在過去的經驗裡，常透過有經驗的操作人員來聽聲音的變化以及其他現象進行判斷。不過為了提升加工便捷、操作簡單化以及提高可靠度，應透過感測器來收集相關訊號，以及透過機器學習來預測刀具磨耗的狀態。 本研究透過麥克風當作感測器，其目的是可以透過非接觸式的方式進行訊號偵測，易於安裝及調整感測器的位置，相較於接觸式感測器，限制較少。過去的研究為了提高實驗的重覆性，通常仔細校準麥克風的位置與角度，以降低實驗時的誤差。然而，實務上可能在安裝麥克風時會有些許角度的誤差，文獻中並未有研究討論到安裝麥克風的角度誤差對預測準確度的影響。本研究在實驗過程中除了單一麥克風外，會透過三支麥克風組成麥克風陣列進行聲音訊號處理，其方式稱為波束成型法(Beamforming)，分別為延遲加總波束成型法(Delay-and-Sum Beamforming, DSB)與最小方差無失真響應波束成型法(Minimum Variance Distortionless Response, MVDR)之方法進行實驗。除此之外，透過群組分離準則以及特徵選取，將刀具磨耗指標進行分類，再利用費雪線性區分法(Fisher linear discriminant)建立刀具偵測系統，預測刀具磨耗的狀態。 實驗結果顯示在小角度的變化之下，可有效地透過MVDR波束成型法補償。相對於對準聲源的情況下準確率可以到93%，麥克風與聲源夾角為5度下，透過MVDR可以將準確率最大從75%上升到88%，角度誤差為10度的情況下，透過MVDR有效將準確率從最大從68%上升到82%。不過觀察到角度越大，能補償的有限，但在10度以內，補償後的預測準確度可以來到80%以上。

In milling operations, the condition of the cutting tool is essential to determining the quality of machining. Therefore, an effective tool wear detection system is crucial. In the past, experienced operators would often rely on auditory cues and visual observations to assess tool conditions. However, to improve convenience, simplify operations, and enhance reliability, it is more effective to utilize sensors to collect relevant signals and apply machine learning techniques to predict tool wear. In this research, microphones are used as sensors for contactless signal collection. This approach allows for easy installation and adjustment of sensor positions, offering more flexibility compared to contact sensors. Furthermore, the study investigates the accuracy of tool wear prediction by analyzing acoustic signals captured at different angles using microphones. To this end, the experimental setup allows the microphone positions to be adjusted freely. The experiments involve both individual microphones and a three-microphone array, utilizing acoustic signal processing methods known as beamforming. The two beamforming techniques employed are Delay-and-Sum Beamforming and Minimum Variance Distortionless Response. In addition, group separation criteria and feature selection are applied to classify tool wear indicators, followed by the use of Fisher Linear Discriminant for tool wear detection to predict tool wear accuracy and establish tool life standards. The results indicate that small angular deviations can be effectively compensated using the MVDR beamforming method. Compared to the scenario where the microphone array is perfectly aligned with the acoustic source, achieving an accuracy of 93%, MVDR improves the accuracy from a maximum of 75% to 88% at a 5-degree angle error. For a 10-degree angle error, MVDR effectively raises the maximum accuracy from 68% to 82%. However, it was observed that the compensation becomes limited as the angle increases. Nevertheless, for deviations within 10 degrees, the post-compensation prediction accuracy remains above 80%.