International Journal of AI and Machine Learning Innovations in Electronics and Communication Technology

NetBot-AI: A Deep Reinforcement Learning Framework for Real-Time Threat Detection and Mitigation in VoIP Networks Using Call Detail Record Analytics

2026-08-27T04:39:21+00:00

The rapid expansion of Voice over IP (VoIP) has enabled scalable communication solutions; however, it has also introduced significant security vulnerabilities, particularly in developing regions that rely on legacy infrastructure. Traditional intrusion detection systems remain largely reactive, struggling to mitigate evolving threats such as Session Initiation Protocol (SIP) flooding and toll fraud within dynamic network environments. To address these limitations, this study proposes NetBot-AI, an artificial intelligence-driven framework designed for real-time threat detection and adaptive mitigation in VoIP networks. The methodology integrates Call Detail Record (CDR) analytics with a Deep Q-Network (DQN), enhanced by a novel dynamic scaling mechanism termed the Ekolama Constant (E_k). Evaluated on a consolidated dataset comprising over ten million real-world and NS-3 simulated call records, the proposed model achieved a binary classification accuracy of 98% and a multi-class accuracy of 95%, maintaining a false positive rate below 5%. Furthermore, integrating the E_k accelerated model convergence by 33% and reduced validation loss by 31% relative to conventional loss functions. These findings demonstrate that NetBot-AI offers a robust, scalable, and proactive cybersecurity solution capable of enhancing VoIP resilience across both advanced and resource-constrained telecommunications environments.

Development of an AI-Based Driver Drowsiness Detection and Alert System for Accident Prevention

2026-09-16T10:48:19+00:00

Driver drowsiness is a significant safety concern that can impair driving performance and increase the risk of road accidents. This research proposes a Driver Drowsiness Detection System with Location Using Artificial Intelligence as a non-invasive, real-time framework for detecting signs of driver drowsiness through facial and eye movements. The proposed system uses a camera to continuously capture the driver's facial features, while computer vision and facial landmark techniques are employed to localize the face and analyze the eye region. The Eye Aspect Ratio (EAR) is used as a measure of eye openness, and temporal analysis of eye closure is applied to distinguish normal blinking from prolonged eye closure associated with drowsiness. When the extracted features meet the predefined drowsiness criteria for a specified period, the system generates an audible and/or visual alert to notify the driver. In addition, a location and emergency-information module can provide the driver's current location and relevant event information to a designated contact or emergency service. By integrating artificial intelligence, computer vision, real-time monitoring, alert generation, and location-based emergency communication, the proposed system aims to provide a practical and non-intrusive approach to reducing the risks associated with driver fatigue and drowsiness and enhancing overall road safety.

The ZeroHAL Framework: Targeting Zero Hallucination

2026-10-08T08:32:53+00:00

Hallucination in large language model systems constitutes a documented and operationally significant reliability risk, with fabricated or contextually inconsistent outputs reported across healthcare, legal, and financial deployments. The article proposes a Zero Hallucination (ZeroHAL) framework comprising a unified thirteen-stage hierarchical framework for retrieval, evaluation, deployment, unified control, and enforcement for systematically diminishing hallucination in agentic retrieval augmented generation pipelines, encompassing thirteen functional stages: (i) clean ingestion, (ii) preprocessing, (iii) parent-child chunking, (iv) ontology-relative embedding, (v) indexing strategy, (vi) storage architecture, (vii) graph orchestration and loop engineering, (viii) retrieval modality selection, (ix) parameter tuning, (x) LLM configuration and context engineering, (xi) human oversight, (xii) evaluation metrics, and (xiii) observability and guardrails. The framework addresses a critical gap in the published literature: while individual mitigation strategies exist in isolation, no unified pipeline-level methodology has been proposed that integrates all of these concerns into a single reproducible workflow. The thirteen stages are instantiated for major cloud platforms including Amazon Web Services, Microsoft Azure, Google Cloud Platform, Databricks, Snowflake, and Palantir Foundry, and Python-native libraries including LangGraph and the SAMF framework (which is based on the MoSCoW – must-have, should-have, could-have, and won’t-have method). An enterprise use case demonstrates application to a 100-million-item heterogeneous corpus spanning seven modality types like documents, spreadsheets, visual content, plain text, tabular data, video, and audio, with index configuration trade-off analysis showing that hallucination-free response generation is achievable across all configurations through faithfulness gating and grounded non-answer protocols.

Event-Triggered Neuromorphic Vision Sensing for Energy-Autonomous Structural Health Monitoring in Smart Cities

2026-10-17T11:34:27+00:00

This paper proposes a novel event-triggered neuromorphic vision framework for energy-autonomous structural health monitoring (SHM) in smart cities, addressing the fundamental limitations of conventional frame-based vision systems: data redundancy, high power consumption, and latency bottlenecks that render them impractical for battery-operated infrastructure deployments. The proposed system fuses Dynamic Vision Sensors (DVS) with MEMS accelerometers and piezoelectric energy harvesters, exploiting the intrinsic alignment between event generation and structural vibration physics: event rate is proportional to vibration velocity, while the same vibration provides harvestable energy. Four-layer Spiking Neural Network (SNN) architecture employs Resonate-and-Fire neurons with resonance frequencies log-spaced from 0.5 to 10 Hz, enabling frequency-selective feature extraction and ultra-low-power damage classification at the edge. Event streams are processed directly without frame reconstruction using contrast maximization and time-surface representations, achieving millimeter-level displacement precision (RMSE < 0.3 mm, R² > 0.98) and millisecond-level response latency. Experimental validation demonstrates 94% weighted-average accuracy for multiclass damage classification across five structural states (healthy, crack, loosening, corrosion, and delamination), with misclassifications occurring exclusively at low excitation amplitudes where event density is insufficient. For binary damage detection (healthy vs. damaged), the system achieves 97% overall accuracy and 94% damaged-class recall. Energy autonomy is achieved under typical bridge vibration conditions (≥ 0.1 g at 2 Hz), with harvested power exceeding consumption and an energy neutrality threshold of approximately 0.08 g. The system achieves approximately a 6× reduction in processing time (95 ms vs. 550 ms per 10-second window) and a 9× reduction in data volume (8 MB vs. 72 MB per 10-second window) compared to conventional frame-based approaches, establishing a pathway toward truly autonomous, continuously vigilant infrastructure monitoring that transforms aging assets into self-aware systems.

AI-Driven Garbage Control: Smart Collection, Sorting and Recycling for Sustainable Waste Management

2026-10-23T12:06:00+00:00

This study examines how Artificial Intelligence (AI) is reshaping waste management, a function that is central to keeping cities clean and sustainable. Conventional approaches depend heavily on manual effort and fixed schedules, which makes them slow and costly. Using a comparative analysis of 12 published sources, the study reviews AI applications in waste collection, sorting, recycling, and environmental monitoring. A smart bin fitted with fill-level sensors tells collectors where pickups are actually needed, so vehicles skip half-empty containers and overflow becomes less frequent. Sorting robots such as the ZRR2 deployed at the Ecoparc4 treatment plant in Barcelona rely on machine vision to separate mixed waste, which improves both recovery and the purity of the output. Citizen-facing tools such as the BinBin Helper applet in Hangzhou, China, extend the benefit to households through image recognition and text search, and the data they gather helps city planners understand waste generation. The comparison shows that AI can shorten collection routes, lower operating costs, and reach high sorting accuracy, although high set-up costs, uneven data quality, and limited policy support still slow adoption. The study concludes that AI can support more efficient, sustainable, and hygienic waste systems.