Magnetic anomaly detection often encounters high false alarm rates or detection failures due to the rapid decay of magnetic anomaly data with distance and its susceptibility to environmental interference. A fully parameter-adaptive magnetic anomaly detection method based on a multi-stable stochastic resonance system is proposed to address this issue. This multi-stable stochastic resonance model exhibits higher noise utilization than mono-stable and bi-stable models, demonstrating superior detection performance for weak signals under strong background noise. However, as the complexity of the well function increases, the severe coupling of system parameters complicates the adjustment of the potential well function. This study proposes an adaptive parameter selection method, which transforms the signal into the frequency domain using continuous wavelet transform (CWT). We propose a combined fitness function comprising peak signal-to-noise ratio (PSNR) and kurtosis index, calculated between the two-dimensional frequency domain image of the magnetic anomaly signal and the reference frequency domain image. Particle swarm optimization (PSO) is then employed to optimize the system parameters, achieving optimal alignment between the multi-stable stochastic resonance system and the input signal, improving the detection probability. The experimental section tests the proposed method on synthetic and field data and compares it to traditional detection methods. The results indicated that the method maintains a high detection probability when the signal-to-noise ratio (SNR) is below -10 dB with colored noise condition and outperforms traditional methods in detection performance.

Magnetic anomaly detection (MAD) serves as a method for detecting ferromagnetic objects via magnetic data. However, MAD typically confronts issues such as low signal-to-noise ratio (SNR) data and lack of prior information, resulting in the failure of conventional anomaly detection methods. Additionally, distinguishing the effective signal from the noise proves challenging, leading to difficulties in subsequent target location and further interpretation. Most of the current denoising algorithms are only applicable to the Gaussian noise, and their performance is poor for geomagnetic noise (generally colored noise). To solve this issue, we propose a new adaptive denoising method for magnetic anomaly data based on continuous wavelet transform (CWT), which elevates the SNR of magnetic anomaly data by integrating the stochastic resonance (SR) and pixel connectivity thresholding. For the problem that the noise wavelet coefficients are difficult to separate from the effective signal in the existing denoising methods in the wavelet domain, we present an adaptive pure background noise estimation method based on SR, which can adaptively remove the noise wavelet coefficients in the wavelet domain. After that, the residual noise wavelet coefficients that still exist are suppressed by using the method of pixel connectivity thresholding in view of the poor continuity of the residual noise wavelet coefficients. In experiments, we compare our proposed method with traditional denoising algorithms on simulated and real data. The results show that our method has better denoising performance with a high SNR, structural similarity (SSIM), and correlation coefficient (CC).

In this work, the influence of laser power (LP), scanning speed (SS), and powder feeding speed (PF) on the porosity, dilution, and microhardness of lightweight refractory high-entropy alloy (RHEA) coatings produced via laser cladding (LC) was investigated. Variance analysis (ANOVA) was deployed to ascertain the effect of LP, SS, and PF on performance metrics such as porosity, dilution, and microhardness. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) was then applied to optimize these processing parameters to minimize porosity, achieve suitable dilution, and maximize microhardness, enhancing the mechanical properties of RHEA coatings. Finally, machine learning models—Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Genetic Algorithm-enhanced GBDT (GA-GBDT)—were developed using orthogonal experimental data, with GA-GBDT demonstrating superior predictive accuracy. The proposed approach integrates statistical analysis and advanced ML techniques, providing a better understanding into optimizing LP, SS, and PF for improved RHEA coatings performance in industrial applications, thereby advancing laser cladding technology.