即時金屬板導電率與導磁率及厚度估測之脈衝渦電流感測系統開發

Abstract

本研究提出即時脈衝渦電流感測系統，結合理論模型建立、參數估測、訊號處理與硬體設計，即時估測金屬板之厚度、導電率與導磁率。理論部分將傳統僅應用於頻域的截斷區域本徵函數展開模型擴展至時域應用，建立金屬材料電磁響應與激勵訊號之對應關係，並針對方波階躍響應引入簡單修正方法去除吉布斯效應對時域曲線邊緣的影響，透過有限元素法模擬進行驗證，證實此模型可有效描述材料參數與磁場變化之物理關聯，並可在短時間內生成大量資料，使用於建立機器學習模型的訓練資料集。在估測方法中探討多種機器學習方法，並整合其優點提出混合式神經網路架構，能直接由量測訊號進行厚度、導電率與導磁率的多參數估測，展現多元應用潛力。 訊號處理部分針對高頻雜訊與有限取樣率造成的誤差提出多項改良策略，大幅提升訊號穩定度並降低隨機雜訊影響，此序列式訊號處理流程確保了後續估測過程的可靠性，硬體設計方面，本研究以穿隧磁阻感測器搭配最佳化激勵線圈為核心，設計一體化線圈支架，使感測器盡可能靠近金屬表面，提升感測靈敏度與訊噪比，同時以資料擷取系統即時進行資料取樣與顯示，在厚度0.2至5.0毫米、導電率1.29至58.3 MS/m、導磁率1至150的範圍內，完成多參數即時估測，平均誤差僅 3.83%，而單次反算時間僅約80毫秒，兼具準確度與可攜性。

This paper proposes a real-time pulsed eddy current sensing system that integrates theoretical modeling, parameter inversion, signal processing, and hardware design for accurate estimation of thickness, electrical conductivity, and relative permeability of metallic plates. In terms of theory, the conventional Truncated Region Eigenfunction Expansion model, which has traditionally been limited to the frequency domain, is successfully extended to the time domain to establish a direct correspondence between the electromagnetic response of metallic materials and the excitation signals. A simple correction method is further introduced to eliminate the Gibbs phenomenon at the step edges of square wave excitation, thus improving the applicability of the time-domain response. Finite element analyses are performed for validation, validating that the proposed model effectively captures the physical relationship between material parameters and magnetic field variations. Moreover, the model can rapidly generate large datasets within a short computation time, which are subsequently employed to construct training data for machine learning models. In the inversion stage, multiple machine learning approaches are compared, and a hybrid neural network architecture that combines their advantages is proposed. This framework enables direct multi-parameter inversion of thickness, conductivity, and permeability from the measured signals, demonstrating strong potential for extensive applications. For signal processing, several stages are designed to mitigate high-frequency noise and sampling limitations, significantly enhancing the stability of the processed signals and reducing the influence of random noise. This sequential signal-processing workflow ensures the reliability of the subsequent inversion process. For hardware design, a Tunneling Magnetoresistance sensor combined with an optimized excitation coil is adopted as the core of the system. An integrated coil–sensor chassis is designed to minimize the distance between the sensor and the sample surface, thereby improving sensitivity and signal-to-noise ratio. The system utilize data acquisition system and MATLAB, enabling real-time sampling, processing, and visualization of the measured signals. Experimental validation demonstrates that the proposed system is capable of performing multi-parameter real-time estimations within the ranges of thickness from 0.2 to 5.0 mm, conductivity from 1.29 to 58.3 MS/m, and relative permeability from 1 to 150. The system achieves an average estimation error of only 3.83%, with each estimaion requiring merely 80 ms, thus presenting both high accuracy and portability.