Battery locomotives used in underground coal mining operations require continuous maintenance, and failures occurring during these operations pose significant occupational safety and health (OSH) risks. Traditional Risk Assessment Methods (TRAMs), particularly the Risk Matrix Method (RMM), often fail to capture the uncertainty and subjectivity inherent in complex mining environments. This study develops a fuzzy logic-based risk assessment framework to improve the evaluation of accident risks associated with maintenance and repair activities in battery locomotive workshops of an underground coal mine in Turkey. Two fuzzy inference models (FL-Basic and FL-Advanced) based on expert knowledge and linguistic variables were designed using Mamdani-type inference with centroid defuzzification. The mathematical formulation of the fuzzy inference and defuzzification steps is presented explicitly, and a six-step algorithm formalises the proposed framework. The rule base of FL-Advanced systematically upweights the severity dimension relative to RMM through reassignment of 16 of the 25 consequent categories. The outputs of these models were compared with RMM to analyse their effectiveness in identifying critical hazards. Application results from Karadon Hard Coal Company show that the proposed FL-Advanced model significantly reduces ambiguity, prioritises high-severity risks more realistically, and provides a more consistent decision-making structure for OSH specialists. The study highlights the advantages of fuzzy logic for modelling uncertain, incomplete, and human-dependent data in hazardous underground mining conditions.

Hydrogen production is gaining increasing importance as a key component of the transition toward carbon-neutral energy systems. In this study, the prediction of hydrogen generation in membraneless alkaline water electrolyzers (MAWEs) is investigated using deep-learning-based time-series modeling. A single-input modeling framework is adopted, where only the system current is used as the input variable. Experimental current signals obtained from long-duration tests conducted at electrolyte concentrations between 5 and 35 g KOH (7200 s per experiment) are employed as the model inputs, while mass-based hydrogen production (in grams) is used as the output variable. Two recurrent neural network architectures, namely Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), are implemented, and their predictive performance is comparatively evaluated using RMSE, MAE, and R² metrics. In addition to deep learning models, classical approaches including Linear Regression, ARIMA, and Naïve Forecast are also considered for comparison. The results show that both models are capable of accurately reproducing the hydrogen-production dynamics across the entire concentration range. In particular, the prediction accuracy improves notably at medium and high electrolyte concentrations, where the coefficient of determination (R²) approaches 0.98. The residual distributions remain narrow and symmetric around zero, indicating the absence of systematic estimation bias. The results also show that classical models can achieve comparable performance under stable operating conditions, while deep learning models provide advantages in capturing nonlinear and dynamic behavior. While LSTM and GRU exhibit comparable accuracy, each architecture provides complementary advantages under different operating conditions. These findings indicate that deep-learning-based time-series modeling constitutes a lightweight and reliable framework for prediction and control applications in MAWE systems. Overall, this study demonstrates the applicability of data-driven models for the dynamic characterization of membraneless water electrolysis.

This paper introduces a novel cascaded softsign function-based PID (CSoft-PID) controller designed for precise pressure regulation in highly nonlinear shell-and-tube steam condenser systems. For the first time in the literature, the classical PID control structure is enhanced through a cascaded nonlinear transformation using the softsign function, which dynamically adjusts the controller input according to the magnitude of the error. This architecture allows for high sensitivity near the setpoint while gracefully limiting excessive control efforts during larger deviations, thereby improving stability and transient performance. To optimally tune the six parameters of the proposed controller, a new hybrid optimization algorithm, termed hGASO-PS, is proposed. This method synergistically integrates an adaptive gbest-guided atom search optimization (ASO) strategy with the precision of the pattern search (PS) technique, ensuring both effective global exploration and fine-tuned local exploitation. The controller parameters are optimized by minimizing the integral of time-weighted absolute error (ITAE), subject to a step change in the condenser pressure setpoint. Extensive simulations and statistical evaluations demonstrate the superiority of the proposed approach. The hGASO-PS-based CSoft-PID controller achieved the lowest ITAE value of 2.1608, with an average of 2.2746 across 30 runs. It also demonstrated the fastest settling time (12.51 s) and the lowest overshoot (1.98%) among all tested controllers. Comparisons with recent PI, FOPID, and cascaded PI-PDN controllers confirm the consistent outperformance of the proposed method in both transient response and control precision, making it a promising candidate for industrial condenser applications.

Electroluminescence (EL) imaging is widely used for photovoltaic (PV) defect inspection, yet fair comparison of deep learning backbones remains difficult because datasets, labels, and protocols vary across studies. This work presents a controlled image-level benchmark of six architectures (ConvNeXt-T, ViT-B/16, DeiT-B/16, Swin-T, DenseNet121, and MobileNetV3-Large) across five hierarchical tasks for monocrystalline and polycrystalline cells with binary and multi-class labels. A balanced proprietary dataset of 20,000 single-cell EL images was evaluated with identical preprocessing, augmentation, training, and stratified five-fold cross-validation, yielding 150 runs. ConvNeXt-T achieved the highest mean macro-F1 (93.12%) while using about one-third of the parameters of base ViT/DeiT models. On the four-class polycrystalline task, it reached 84.94 ± 0.45% macro-F1, compared with 70.08 ± 1.19% for DenseNet121 and 59.43 ± 1.71% for MobileNetV3-Large. Error analysis revealed conservative missed-defect behavior in lightweight CNNs, especially for surface-level degradation and crack categories. The results provide image-level cross-validation evidence for controlled benchmarking and motivate future module-level grouped validation.

Accurate temperature regulation is essential for ensuring product quality, operational safety, and energy efficiency in industrial electric furnace systems. However, the inherent thermal inertia, time-delay effects, and nonlinear dynamics of furnace processes often make precise temperature control a challenging task. Motivated by these challenges, this study proposes an optimization-based control framework aimed at improving the temperature regulation performance of electric furnace systems. The proposed approach integrates a proportional–integral–derivative (PID) controller with the recently developed gazelle optimization algorithm (GOA) for automatic tuning of the controller parameters. First, a mathematical model of the electric furnace is established to describe the dynamic relationship between the control input and the furnace temperature output. Based on this model, a PID controller is implemented to regulate the furnace temperature. The parameters of the PID controller are then optimized using GOA, a nature-inspired metaheuristic algorithm that mimics the adaptive predator–prey survival strategies observed in gazelle herds. In order to achieve a balanced improvement in both steady-state and transient performance, a composite objective function is introduced. The proposed performance index combines the integral of absolute error with additional transient performance indicators related to maximum overshoot and settling time. The effectiveness of the proposed GOA-based tuning framework is evaluated through extensive simulation studies and statistical analyses conducted over multiple independent optimization runs. The results demonstrate stable convergence behavior, with the optimization process achieving a minimum objective value of 2.4251, a maximum value of 2.5347, and an average value of 2.4674 across 25 runs. The optimized control system exhibits improved dynamic characteristics, including a rise time of 1.8509 s, a settling time of 3.6834 s, and a low overshoot of 1.5104%. To further assess its effectiveness, the proposed GOA–PID control strategy is compared with several widely used controller tuning methods reported in the literature, including genetic algorithm, Ziegler–Nichols, Cohen–Coon, Nelder–Mead, and direct synthesis approaches. Comparative results indicate that the proposed method achieves a superior balance between response speed, stability, and temperature tracking accuracy.

Automated visual inspection of photovoltaic (PV) cells is an important component of solar-module quality assurance. However, adapting modern pre-trained vision backbones to PV defect classification remains challenging because full fine-tuning requires substantial memory, naturally imbalanced datasets can reduce sensitivity to rare defect classes, and edge-oriented inspection workflows impose computational constraints. Parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA), have been widely studied in natural language processing, but their use for PV defect classification remains underexplored. This study presents a controlled benchmark of LoRA and QLoRA against full fine-tuning for PV cell defect classification. Four adaptation strategies—full fine-tuning, LoRA with rank 8, LoRA with rank 16, and 4-bit QLoRA with rank 16—are evaluated using a ConvNeXt-Tiny backbone on a 17,377-image polycrystalline PV cell electroluminescence dataset referred to as POLY, covering five classes: intact, cracked, broken, surface-diffuse, and surface-point. The natural 6.7× class imbalance is preserved without synthetic resampling, and a group-aware StratifiedGroupKFold protocol based on available cell or panel-image identifiers is used to reduce identifiable leakage across folds. All PEFT variants slightly outperform full fine-tuning in macro-F1 while training 26–52× fewer parameters. QLoRA_r16 achieves the highest macro-F1 score of 79.92 ± 0.75%, compared with 78.26 ± 0.94% for full fine-tuning, while training the same number of parameters as LoRA_r16 (1.060 M; 3.67% of the adapted model). QLoRA_r16 also improves F1 on the intact (+4.75 points) and surface-diffuse (+2.62 points) classes relative to full fine-tuning. This class-wise pattern suggests that quantized low-rank adaptation may influence minority and visually ambiguous categories; however, the present experiments do not isolate the independent effect of NF4 quantization from adapter rank, batch size, or optimization dynamics. Under the training configuration used, QLoRA_r16 records the lowest observed peak training GPU memory, approximately 30% below full fine-tuning (1727 MB versus 2478 MB). Because QLoRA_r16 was trained with batch size 16 whereas the other methods used batch size 32, this reduction should be interpreted as an end-to-end configuration effect rather than as the isolated effect of 4-bit quantization. Overall, the results indicate that PEFT is a promising and resource-efficient alternative to full fine-tuning for PV defect classification, although batch-matched memory experiments, direct embedded-device profiling, and cross-dataset validation remain necessary before making deployment-level claims.