Fault-aware dual-head transformer for joint ammonium imputation and multi-horizon forecasting under sensor degradation
Results in Engineering
Daniel Voipan este asistent universitar la Universitatea „Dunărea de Jos” din Galați și lucrează la granița dintre învățarea automată și ingineria reglării automate. Direcția sa principală de cercetare sunt stațiile de epurare a apelor uzate: senzori software bazați pe date pentru calitatea efluentului și oxigenul dizolvat, modele surogat care rămân utile în bucla închisă de reglare și prognoze robuste la defectarea sau derivarea senzorilor. Lucrările recente tratează inteligența artificială explicabilă pentru mentenanța predictivă a sistemelor de aerare, modele transformer care completează și prognozează simultan concentrația de amoniu și cadre predictive neuronale, conștiente de constrângerile reglării, pentru gestionarea aerării.
În paralel, contribuie la aplicații ale inteligenței artificiale în medicină (screening pe baza ECG, monitorizare în cardio-oncologie, predicția riscului de accidentare în sport) și la studii despre reziliența organizațională, cu metode de procesare a limbajului natural și ale sistemelor complexe. A publicat în reviste precum Sensors, Water, Systems, Diagnostics și Results in Engineering și prezintă constant la conferințele IEEE ETFA, ICSTCC și Mediterranean Conference on Control and Automation (MED).
Senzori software și prognoze pentru calitatea efluentului, amoniu și oxigen dizolvat în stațiile de epurare.
Modele surogat și predictive neuronale gândite să funcționeze în bucla închisă de reglare a aerării.
Modele care rămân fiabile când senzorii se degradează și explicații pe care operatorii le pot folosi.
Screening pe baza ECG, monitorizarea cardiotoxicității și predicția riscului de accidentare, în colaborări clinice.
Analiză de limbaj și modele de dinamică colectivă pentru reziliența sistemelor om-cibernetic-fizice.
Results in Engineering
Water
11 citări
Sensors
31 citări
Results in Engineering
Biofouling, calibration drift, offsets, and signal loss in effluent ammonium sensors introduce structured bias into the measurements that drive aeration control in activated sludge systems. Current data-driven models treat reconstruction and forecasting as distinct tasks, leaving a robustness gap under degraded sensing. This paper presents a fault-aware dual-head Transformer-encoder/GRU-decoder framework that jointly reconstructs corrupted ammonium measurements and generates multi-horizon forecasts from a shared encoder. Seven engineered descriptors, comprising missingness masks, gap encoding, a gap-attenuated input state, rolling statistics, and local z-score deviations, are used as model inputs, together with adaptive gating and a residual correction pathway. Evaluated on Benchmark Simulation Model No. 2 (BSM2) under four fault types over five seeds, the framework achieves imputation R 2 = 0.996 ± 0.001 (MAE = 0.012 mg · L − 1 ) and forecasting R 2 = 0.884 ± 0.012 (MAE = 0.058 mg · L − 1 ) over a 4-hour horizon, reducing forecasting error by 17.8–48.3% relative to recurrent baselines, significantly so for both drift regimes, while also outperforming a learned imputation baseline (SAITS) and matching or exceeding a modern deep time-series forecaster (TimesNet) under a common protocol. Per-fault reconstruction ranges from R 2 = 0.999 (slow drift) to 0.44 (fast drift), and it retains a positive forecasting R 2 during multi-day outages, unlike the LSTM and GRU baselines, although the corresponding MAE difference is statistically unresolved. SHAP attribution, cross-validated against permutation importance, shows a shift from input-state-dominated reconstruction to deviation-dominated long-horizon forecasting. Degradation remains graceful up to approximately twice the training fault severity, while abrupt step components superimposed on gradual offsets expose an identifiability limit shared by all evaluated methods. Ablation attributes most forecast robustness to the engineered descriptors, which are the primary contribution; an architecture-only advantage is not established. The results support fault-aware feature engineering for resilient digital monitoring, within the scope of simulation-based validation.
Systems
Surrogate models can reproduce wastewater effluent trajectories accurately, but their use in optimization also requires a suitable representation of the response to control actions. This paper proposes an actionability-aware framework for assessing and training surrogates for multi-timescale wastewater treatment control. The framework is evaluated on Benchmark Simulation Model No. 2 (BSM2), where a pre-training diagnostic shows that the measured four-hour aeration-to-ammonia gain is smaller than the asymptotic reference by more than two orders of magnitude. Two Mamba surrogates are constructed at 4 h and 45-day prediction horizons, respectively, and trained using trajectory, counterfactual-pair, and gain-margin losses. Closed-loop validation considers recirculation optimization at 5-day intervals and hourly dissolved-oxygen setpoint supervision as separate supervisory placements above the native PI aeration loop. The recirculation-floor and constant-setpoint sweeps show approximately 37-fold and 16-fold increases in marginal energy cost per violation day avoided, respectively. A constant-flow comparison indicates that the recirculation floor accounts for most of the recovered total-nitrogen compliance. At fixed recirculation, the learned setpoint supervisor reduces aeration energy by 3.0 ± 0.6% across four trained seeds relative to the interpolated constant-setpoint curve at equal summed ammonia and total-nitrogen violation duration. This matching does not imply equal individual durations: the supervisor shortens ammonia violations while lengthening total-nitrogen violations. This reduction corresponds to approximately 42,900 kWh over the evaluation period. A trajectory-loss-only ablation achieves approximately one seventh of the cost-index improvement. The results support the complementary use of horizon-dependent response analysis, actionability training, and matched-compliance validation when assessing learned surrogates for supervisory control.
2026 European Control Conference (ECC) · Reykjavík, Iceland
Mediterranean Conference on Control and Automation (MED) · Ancona, Italy
The extensive calibration burden of mechanistic process models is a common impediment to the deployment of Model Predictive Control (MPC) in wastewater treatment plants. In this paper, the Control-Aware Temporal Transformer (CATT) is introduced as a neural architecture that explicitly conditions effluent ammonium (SNHe) forecasts on candidate aeration sequences. This integration is made seamless with derivative-free MPC optimization. CATT outperforms LSTMbased baselines in both prediction accuracy and parameter efficiency when trained on Benchmark Simulation Model 2 (BSM2) data. The control-aware design is instrumental in ensuring reliable multi-step forecasting. Using probabilistic candidate rejection, a Cross-Entropy Method (CEM) optimizer enforces hard regulatory and actuator limits to ensure constraint satisfaction. Throughout the closed-loop simulation, CATT-MPC achieves zero regulatory violations, delivers substantial aeration energy savings, and maintains effluent quality within the target operational band in comparison to constant open-loop operation, all without requiring any mechanistic model calibration.
Applied Sciences
Supply chain disruptions during the COVID-19 pandemic exposed structural vulnerabilities of centrally controlled manufacturing systems, motivating renewed interest in organizational resilience within the context of Industry 5.0 human–cyber–physical systems. This study investigates how organizational decision-making paradigms and crisis-aware artificial intelligence (AI) jointly influence performance, crisis response, and recovery. An agent-based modeling (ABM) framework is developed to compare centralized, distributed, and self-organized organizational structures across 650 simulation runs under a controlled supply side disruption. A crisis-aware Q-learning architecture enables AI agents to shift from efficiency-oriented to stability-oriented strategies when resource scarcity is detected. To avoid baseline-dependent bias, resilience is evaluated using an absolute, capacity-normalized metric. Results indicate that self-organized systems consistently outperform centralized and distributed structures in baseline performance, crisis throughput, and recovery speed. The integration of crisis-aware AI further increases absolute resilience by approximately 10.7% and enables substantially higher throughput during disruption compared to hierarchical control. Enhanced performance is primarily driven by adaptive coalition formation, proactive resource conservation, and rapid post-crisis recovery supported by preserved coordination structures. These findings provide quantitative support for Industry 5.0’s human-centric principles and show that decentralized decision-making augmented by context-adaptive AI offers a robust organizational design strategy for volatile manufacturing environments.
Systems
Organizational resilience and long-term success have become essential capabilities in turbulent and uncertain environments. This study redefines these concepts by applying a natural language analysis to a corpus of 1597 peer-reviewed publications retrieved from Web of Science and Scopus. The methodology adopts a three-level framework: first, a thematic clustering of literature into strategic domains; second, a semantic comparison of classical and emerging terms; and third, the mapping of artificial intelligence (AI) opportunities. The results identify five overarching domains: Health and Wellbeing; Organizations, HR and Leadership; Strategy, Innovation, and Culture; Education, Knowledge and Communities; and Society, Environment and Development. These domains illustrate how resilience and success are addressed at micro, meso, and macro levels. Semantically, the discourse expands from traditional notions such as robustness, risk management, and performance towards more human-centered, systemic, and digitally enabled perspectives. The study further highlights how AI functions both as a methodological tool and as a strategic enabler, with applications ranging from predictive health analytics and leadership support systems to foresight tools and sustainability monitoring. The findings contribute to organizational resilience theory and offer practitioners actionable pathways to strengthen resilience and competitiveness in the face of volatility, uncertainty, complexity, and ambiguity.
International Conference on System Theory, Control and Computing (ICSTCC) · Cluj-Napoca, Romania
This paper presents an AI-based solution for a data-driven control framework in the field of wastewater treatment, highlighting surrogate modeling and its advantages. To approximate the ammonium (SNHe) concentration in the effluent, a long short-term memory (LSTM) neural network (NN) was trained, using selected reference simulations from Benchmark Simulation Model no. 2 (BSM2) as input variables. To identify the most relevant predictors, recursive feature elimination (RFE) was performed. The surrogate model was first trained on open-loop simulation data and then integrated into a closed-loop control system by generating oxygen transfer rate (KLa) inputs based on SNHe predictions. This data-driven KLa signal was injected into the BSM2 simulation in place of the standard controller. A feedback-based retraining process was used to refine the model’s performance under controlled conditions. Experimental results demonstrate high prediction accuracy, a coefficient of determination (R2) with an average value of 0.974, and a substantial reduction in effluent SNH levels, with SNHe values decreasing from an average of 7.62 mg/L to 0.19 mg/L after control. These findings support the integration of AI-based surrogate models into supervisory control strategies for wastewater treatment plants.
International Conference on System Theory, Control and Computing (ICSTCC) · Cluj-Napoca, Romania
This paper presents a comparative study of data-preprocessing techniques for developing a gated recurrent unit (GRU)–based soft sensor to estimate dissolved oxygen (DO) levels in wastewater treatment plants (WWTPs). Simulation data from the Benchmark Simulation Model No. 2 (BSM2) were used to evaluate three feature-selection methods—SHAP, Boruta, and recursive feature elimination (RFE)—to identify the most informative input variables. Concurrently, three normalization schemes (MinMaxScaler, StandardScaler, and RobustScaler) were assessed for their impact on model accuracy. Each GRU network was trained and tested on datasets preprocessed according to every combination of feature-selection and normalization method. Model robustness was further evaluated under controlled input-noise conditions. Under nominal conditions, the best performance was achieved by pairing StandardScaler normalization with RFE-selected features, yielding the lowest prediction error. Under noisy inputs, models trained with RobustScaler demonstrated superior resilience to perturbations. Finally, the top-performing GRU models were integrated into a MATLAB Simulink environment for practical validation. The results underscore the critical influence of preprocessing choices on soft-sensor accuracy and robustness and confirm the feasibility of GRU networks for reliable DO estimation in wastewater treatment applications.
Water
Aeration systems are among the most energy-intensive components of wastewater treatment plants (WWTPs), consuming up to 75% of total electricity while being prone to performance degradation caused by diffuser fouling and pressure losses. Traditional maintenance strategies are largely reactive or preventive, leading to inefficient interventions, higher operational costs, and limited fault anticipation. This study addresses the need for an advanced predictive maintenance framework capable of early detection and differentiation of multiple aeration system faults. Using the Benchmark Simulation Model No. 2 (BSM2), two representative degradation scenarios—acute airflow pressure loss and chronic diffuser fouling—were simulated to generate a labeled dataset. A hybrid machine learning approach was developed, combining Random Forest-based feature selection with Long Short-Term Memory (LSTM) neural networks for temporal, multi-label fault classification. To enhance interpretability and operator trust, SHapley Additive exPlanations (SHAP) were applied to quantify feature contributions and provide transparent model predictions. The results show that the proposed framework achieves over 94% detection accuracy and provides early warnings compared to static threshold-based methods. The integration of explainable AI ensures actionable insights for maintenance planning. This approach supports more energy-efficient, reliable, and sustainable operation of WWTP aeration systems and offers a benchmark methodology for future predictive maintenance research.
Diagnostics
The increasing prevalence of cardiovascular complications in cancer patients due to cardiotoxic treatments has necessitated advanced monitoring and predictive solutions. Cardio-oncology is an evolving interdisciplinary field that addresses these challenges by integrating artificial intelligence (AI) and smart cardiac devices. This comprehensive review explores the integration of artificial intelligence (AI) and smart cardiac devices in cardio-oncology, highlighting their role in improving cardiovascular risk assessment and the early detection and real-time monitoring of cardiotoxicity. AI-driven techniques, including machine learning (ML) and deep learning (DL), enhance risk stratification, optimize treatment decisions, and support personalized care for oncology patients at cardiovascular risk. Wearable ECG patches, biosensors, and AI-integrated implantable devices enable continuous cardiac surveillance and predictive analytics. While these advancements offer significant potential, challenges such as data standardization, regulatory approvals, and equitable access must be addressed. Further research, clinical validation, and multidisciplinary collaboration are essential to fully integrate AI-driven solutions into cardio-oncology practices and improve patient outcomes.
Sensors
Maintaining effluent quality in wastewater treatment plants (WWTPs) comes with significant challenges under variable weather conditions, where sudden changes in flow rate and increased pollutant loads can affect treatment performance. Traditional physical sensors became both expensive and susceptible to failure under extreme conditions. In this study, we evaluate the performance of soft sensors based on artificial intelligence (AI) to predict the components underlying the calculation of the effluent quality index (EQI). We thus focus our study on three ML models: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Transformer. Using the Benchmark Simulation Model no. 2 (BSM2) as the WWTP, we were able to obtain datasets for training the ML models and to evaluate their performance in dry weather scenarios, rainy episodes, and storm events. To improve the classification of networks according to the type of weather, we developed a Random Forest (RF)-based meta-classifier. The results indicate that for dry weather conditions the Transformer network achieved the best performance, while for rain episodes and storm scenarios the GRU was able to capture sudden variations with the highest accuracy. LSTM performed normally in stable conditions but struggled with rapid fluctuations. These results support the decision to integrate AI-based predictive models in WWTPs, highlighting the top performances of both a recurrent network (GRU) and a feed-forward network (Transformer) in obtaining effluent quality predictions under different weather conditions.
Diagnostics
Background/Objectives: Cardiac arrhythmias impact quality of life (QoL) and are often linked to psychological distress. This study examines the relationship between QoL, depression, and arrhythmias using AI-assisted analysis to enhance patient management. Methods: A total of 145 patients with arrhythmias were assessed using an SF-36 health survey (QoL) and a PHQ-9 questionnaire (depression). Statistical analyses included regression, clustering, and AI-based models such as K-means and logistic regression to identify risk factors and patient subgroups. Results: Patients with comorbidities had lower QoL and higher depression scores. PHQ-9 scores negatively correlated with SF-36 mental health components. AI-assisted clustering identified distinct patient subgroups, with older individuals and those with longer disease duration exhibiting the lowest QoL. Logistic regression predicted depression with 93% accuracy, and XGBoost achieved an AUC of 0.97. Conclusions: QoL plays a key role in arrhythmia management, with depression significantly influencing outcomes. AI-driven predictive models offer personalized interventions, improving early detection and treatment. Future research should integrate wearable technology and AI-based monitoring to optimize patient care.
Diagnostics
Background/Objectives: This study aimed to evaluate an artificial intelligence (AI)-enhanced electrocardiogram (ECG) screening protocol for improved accuracy, efficiency, and risk stratification across six sports: handball, football, athletics, weightlifting, judo, and karate. Methods: For each of the six sports, resting 12-lead ECGs from healthy children and junior athletes were analyzed using AI algorithms trained on annotated datasets. Parameters included the QTc intervals, PR intervals, and QRS duration. Statistical methods were used to examine each sport’s specific cardiovascular adaptations and classify cardiovascular risk predictions as low, moderate, or high risk. Results: The accuracy, sensitivity, specificity, and precision of the AI system were 97.87%, 75%, 98.3%, and 98%, respectively. Among the athletes, 94.54% were classified as low risk and 5.46% as moderate risk with AI because of borderline abnormalities like QTc prolongation or mild T-wave inversions. Sport-specific trends included increased QRS duration in weightlifters and low QTc intervals in endurance athletes. Conclusions: The statistical analyses and the AI-ECG screening protocol showed high precision and scalability for the proposed athlete cardiovascular health risk status stratification. Additional early detection research should be conducted further for diverse cohorts of individuals engaged in sports and explore other diagnostic methods that can help increase the effectiveness of screening.
MANAGEMENTUL CUNOȘTINȚELOR ȘI CUNOAȘTEREA MANAGERIALĂ ÎN ORGANIZAȚIILE COMPETITIVE MODERNE
Diagnostics
This review provides a comprehensive analysis of the transformative role of artificial intelligence (AI) in predicting and preventing sports injuries across various disciplines. By exploring the application of machine learning (ML) and deep learning (DL) techniques, such as random forests (RFs), convolutional neural networks (CNNs), and artificial neural networks (ANNs), this review highlights AI's ability to analyze complex datasets, detect patterns, and generate predictive insights that enhance injury prevention strategies. AI models improve the accuracy and reliability of injury risk assessments by tailoring prevention strategies to individual athlete profiles and processing real-time data. A literature review was conducted through searches in PubMed, Google Scholar, Science Direct, and Web of Science, focusing on studies from 2014 to 2024 and using keywords such as 'artificial intelligence', 'machine learning', 'sports injury', and 'risk prediction'. While AI's predictive power supports both team and individual sports, its effectiveness varies based on the unique data requirements and injury risks of each, with team sports presenting additional complexity in data integration and injury tracking across multiple players. This review also addresses critical issues such as data quality, ethical concerns, privacy, and the need for transparency in AI applications. By shifting the focus from reactive to proactive injury management, AI technologies contribute to enhanced athlete safety, optimized performance, and reduced human error in medical decisions. As AI continues to evolve, its potential to revolutionize sports injury prediction and prevention promises further advancements in athlete health and performance while addressing current challenges.
Diagnostics
The application of artificial intelligence (AI) in electrocardiography is revolutionizing cardiology and providing essential insights into the consequences of the COVID-19 pandemic. This comprehensive review explores AI-enhanced ECG (AI-ECG) applications in risk prediction and diagnosis of heart diseases, with a dedicated chapter on COVID-19-related complications. Introductory concepts on AI and machine learning (ML) are explained to provide a foundational understanding for those seeking knowledge, supported by examples from the literature and current practices. We analyze AI and ML methods for arrhythmias, heart failure, pulmonary hypertension, mortality prediction, cardiomyopathy, mitral regurgitation, hypertension, pulmonary embolism, and myocardial infarction, comparing their effectiveness from both medical and AI perspectives. Special emphasis is placed on AI applications in COVID-19 and cardiology, including detailed comparisons of different methods, identifying the most suitable AI approaches for specific medical applications and analyzing their strengths, weaknesses, accuracy, clinical relevance, and key findings. Additionally, we explore AI's role in the emerging field of cardio-oncology, particularly in managing chemotherapy-induced cardiotoxicity and detecting cardiac masses. This comprehensive review serves as both an insightful guide and a call to action for further research and collaboration in the integration of AI in cardiology, aiming to enhance precision medicine and optimize clinical decision-making.
2024 28th International Conference on System Theory, Control and Computing, ICSTCC 2024 - Proceedings · Sinaia, Romania
This research paper delves into the application of Long Short-Term Memory (LSTM) neural networks within the Benchmark Simulation Model No. 2 (BSM2) to enhance the predictability and efficiency of wastewater treatment processes. The study aims to develop advanced predictive models that can simulate the dynamics of wastewater treatment more accurately and adjust operational strategies dynamically. By integrating LSTM networks, the research enables continuous prediction of Effluent Quality Index (EQI) variables under stochastic and deterministic scenarios, thereby improving the accuracy and efficiency of predicting pollutant levels. The research uses an LSTM model to learn from a comprehensive dataset derived from historical simulations of BSM2, where key parameters such as the oxygen transfer coefficient (KLa) are systematically varied to measure their impact on effluent quality. The LSTM's capability to handle complex, nonlinear data and its adaptability to time series forecasting significantly enhances model performance, offering a robust tool for real-time decision-making and process optimization in wastewater treatment facilities. This approach not only improves the accuracy and efficiency of predicting pollutant levels but also supports environmental compliance and operational sustainability, making it a valuable tool for environmental engineers and professionals in the field of wastewater treatment.
IEEE International Conference on Emerging Technologies and Factory Automation, ETFA · Sinaia, Romania
The management of wastewater is a significant global concern that calls for innovative solutions to lessen its negative effects on the environment. Conventional techniques of treating wastewater need improvement in order to deal with newly discovered contaminants, which highlights the importance of providing precise estimates of process performance and resource requirements. The worsening water shortage situation requires a paradigm shift in which wastewater is viewed as a useful resource. It is possible to create an economy that is both sustainable and circular by treating and recycling wastewater, putting less pressure on freshwater supplies, and leaving as little of an environmental footprint as possible. This study investigates the use of Artificial Neural Networks (ANNs) as software estimators in the treatment of wastewater, with a particular emphasis on predicting ammonium concentrations in effluent. In order to deal with imbalanced time-series data, the research introduces innovative data pretreatment strategies. These techniques include a Sliding Window protocol, Data Normalization, and a K-Fold training scheme. This illustrates the potential of ANNs to revolutionize wastewater treatment procedures and drive developments in this field. The suggested method demonstrates higher performance when estimating pollutant concentrations, showing the ability of ANNs to do so.
Nicio publicație nu corespunde filtrului.
Reykjavík, Iceland · Lucrare
Ancona, Italy · Lucrare
Cluj-Napoca, Romania · Lucrare
Sinaia, Romania · Lucrare
Prezentat: Enhancing Effluent Quality Predictions in Wastewater Treatment with LSTM Neural Network
Sinaia, Romania · Lucrare
Prezentat: Integration of ANN for Accurate Estimation and Control in Wastewater Treatment
Deschis la colaborări de cercetare, proiecte comune și coordonare de studenți.