The presented work considers three approaches for detecting the coordinates of contacts in printed circuit boards (PCBs). The objects of the approaches are contacts with holes, contacts without holes, and mixed contacts. The central procedure of the method is the flood-filling algorithm, which detects the component edges simultaneously. Overlaying the found coordinates of contacts with the standard PCB creates an image, which allows the determination of features characterizing defects, such as connectivity, redundant or missing components. Machine learning based on these identified image features is applied to classify printed circuit boards. Artificial neural networks (ANNs), which are available on the Internet, were used in the work. The developed approach combines two components: software for image processing and Machine Learning.
In this paper, a nonlinear analysis is provided for a simplified version of a ferro-resonant circuit with a non-linear hysteresis to exhibit sudden peaks in voltage/current/flux, which endanger and stress the components and shorten the lifespan of the component circuit. The approach relies on a toy model circuit (simplified version of a ferro-resonant power circuit), which is nonlinear due to magnetic hysteresis. Utilizing a well-known local analysis to evaluate the occurrence of this phenomenon, it is shown that the classical sufficient condition for local observability may fail around operating points associated with peak behavior when only voltage/current measurements are considered. This result is local and model-dependent and does not preclude the use of alternative measurements or other new practical detection methods.
Although semiconductor fabrication has advanced significantly in recent years, manufacturing lines frequently encounter operational inefficiencies when processing highly specialized or non-standard substrates. These disruptions degrade overall line performance and ultimately reduce product yield. This issue is equally pronounced at the single-device level, where performance degradation and elevated defect rates necessitate substantial resource allocation for post-fabrication screening and inspection to prevent faulty components from reaching end users. The better practice is to try and discover, and consequently, eliminate the bad devices before they run through multiple processes, costing time and money towards a useless end. In some cases, the solution to overcome these difficulties becomes as costly as it involves modification of properties at the individual device level. This in turn saves the product quality, but still consumes resources to a point making the production process marginally profitable. A better solution to such problem is one that approaches the wafer as a whole and not at the device level. Given that these wafers typically accommodate roughly a thousand individual devices, adopting a global, wafer-level approach becomes highly advantageous. This paper addresses a real-world industrial challenge encountered during wafer fabrication when attempting to meet strict target specifications. This issue consistently arises in wafers processed via evaporated metal thin-film deposition. A prominent example of such components includes passive filters designed around Surface Acoustic Wave (SAW) principles. A solution that replaces the existing custom-made one is proposed. This solution eliminates approaching every device on the wafer individually and looks at the wafer as a whole. The proposed solution applies a statistical approach to the problem using the concepts of statistical process control. Implementing this solution yields substantial improvements, including a marked increase in product yield, significant cost reductions, and the successful integration and promotion of process automation.
The Dendritic Cell Algorithm (DCA) is a classification algorithm inspired by the biological immune system that has shown promise for anomaly detection. However, the DCA relies on manual or domain expertise to tune signal weights which limits practical deployment and reproducibility. This paper addresses the DCA parameter optimisation challenge by applying the Clonal Selection Algorithm to optimise the weight matrix. The approach was evaluated using nine benchmark datasets achieving mean ROC-AUC of 96.48% compared to 66.56–75.33% for manually configured baselines (DCA-2 and DCA-1 individually, and best-of-the-two, respectively). CSA-DCA was also compared with alternative metaheuristic methods (GA-DCA: 92.25%; PSO-DCA: 90.83%), achieving numerically higher mean performance, though the difference was not statistically significant (Wilcoxon p = 0.082 vs. GA-DCA, p = 0.213 vs. PSO-DCA; small effect sizes, d = 0.34 and 0.41). The variance of CSA-DCA (CV = 7.88%) was lower than both manual baselines (CV > 35%), confirming more consistent performance across diverse problem domains.
Labour is a critical factor in the manufacturing sector, particularly in agricultural processing industries, which constitute a major industrial base in Thailand. Accurate forecasting of labour demand at the factory level is therefore essential for effective production planning and operational continuity. This study aims to identify an appropriate predictive approach to estimate the number of workers required in individual factories using fundamental production-related factors as predictor variables. Three machine learning models—Decision Tree Regressor, Random Forest Regressor, and Gradient Boosting Regressor—were developed and compared. The dataset was divided into 70% for model training and 30% for testing on unseen data. Based on the results of the 10-fold cross-validation performed on the training dataset, pairwise comparisons using paired t-tests showed that the Random Forest Regressor (RFR) achieved significantly lower MAE and MAPE values than both the Decision Tree Regressor (DTR) and the Gradient Boosting Regressor (GBR) (p < 0.05). However, for RMSE, although RFR significantly outperformed DTR, the difference between RFR and GBR was not statistically significant (p > 0.05). Evaluation on the independent test dataset further demonstrated that RFR achieved the best predictive performance and strong generalisation ability, producing the lowest RMSE, MAE, and MAPE values of 98.0732, 34.2280, and 1.5493, respectively. Furthermore, pairwise comparisons using the Wilcoxon signed-rank test confirmed that the absolute prediction errors of RFR were significantly lower than those of both DTR and GBR at the 95% confidence level (p < 0.05). Overall, these findings provide statistical evidence that RFR consistently delivered superior predictive performance across both the cross-validation phase and the independent test dataset. The combination of high predictive accuracy, statistical superiority, and robust generalisation capability indicates that RFR is the most appropriate tree-based regression model for labour demand forecasting in agricultural processing factories.
