Node-Level Fault Monitoring in WSN using Deep Neural Networks with Real-Time Graph Mapping and Performance Analysis

Technical Summary: Node-Level Fault Monitoring in WSN using Deep Neural Networks with Real-Time Graph Mapping and Performance Analysis

Problem Statement
Wireless Sensor Networks (WSNs) are critical for real-time monitoring in sectors such as agriculture, healthcare, and environmental surveillance. However, sensor nodes deployed in harsh, unpredictable environments are prone to node-level failures caused by hardware degradation, battery wear, software bugs, and communication issues like packet loss. These faults compromise data integrity and network reliability. Traditional fault detection methods, which rely on static thresholds or rule-based systems, often fail to adapt to changing sensor baselines or complex, non-linear data patterns. Furthermore, these conventional approaches frequently generate false alarms or miss subtle fault indicators, particularly in large-scale or dense networks where computational and communication overheads are significant.

Methodology
The authors propose a deep learning-based framework for real-time, node-level fault detection and visualization. The system is implemented as a web-based portal using Python, Streamlit, TensorFlow, Keras, Scikit-learn, Pandas, and NumPy. The methodology follows a structured pipeline:

1. Data Ingestion and Preprocessing: The system accepts sensor datasets in CSV format containing timestamps, sensor readings, battery levels, ambient temperature, and binary fault indicators. Data is shuffled to remove serial effects, normalized using Min-Max scaling to ensure feature uniformity, and split into an 80% training and 20% testing set.
2. Model Architecture: Three distinct deep learning models are constructed and trained to classify sensor nodes as either "faulty" or "normal":
   • Convolutional Neural Network (CNN): Designed to identify spatial patterns in structured data.
   • Recurrent Neural Network (RNN): Utilized for detecting sequential dependencies in time-series data.
   • Long Short-Term Memory (LSTM): An advanced RNN variant capable of learning long-term patterns and mitigating vanishing gradient issues.
     All models are compiled with the Adam optimizer and trained using binary cross-entropy loss.
3. Real-Time Monitoring and Visualization: Upon training, the system operates in a real-time mode where new sensor data is normalized and fed into the models. The interface provides:
   • Real-time prediction probabilities and confidence levels for each model.
   • Dynamic updates of performance metrics (Accuracy, Precision, Recall, F1-Score) via line charts.
   • Continuous updates of confusion matrices.
   • Visual analytics including histograms, correlation heatmaps, and bar charts to analyze feature relationships (e.g., the negative correlation between battery life and fault status).
   • An audio alarm system triggered when a fault is detected.

Key Contributions

• Comparative Deep Learning Analysis: The study provides a direct comparison of CNN, RNN, and LSTM architectures specifically for WSN fault classification, evaluating their efficacy in distinguishing between normal and faulty sensor behaviors.
• Real-Time Graph Mapping Interface: Unlike static analysis tools, this work integrates deep learning with a responsive, user-friendly web interface that visualizes model performance and node status dynamically. This includes real-time plotting of metrics and correlation matrices to aid in decision-making.
• Multi-Feature Fault Detection: The framework utilizes a temporal feature set including temperature, battery life, and sensor readings to enhance detection accuracy, moving beyond simple threshold-based checks.

Results
The models were evaluated using the Kaggle Node Status dataset. Key findings include:

• Performance Metrics: The RNN model demonstrated the superior overall performance, achieving the highest accuracy (0.9654), precision (0.9713), recall (0.8597), and F1-score (0.9121), outperforming both CNN and LSTM.
• Computational Efficiency: The CNN model was the fastest in terms of total prediction time (0.7994 s) and per-sample time (0.00008 s). However, it exhibited lower recall (0.6676) and F1-score (0.7792), indicating a trade-off between speed and the ability to correctly identify all faulty cases.
• LSTM Performance: The LSTM model showed competitive accuracy (0.9552) and recall but required the most computational resources, resulting in the longest total prediction time (1.3601 s).
• Visual Insights: Correlation heatmaps confirmed strong negative correlations between battery life and fault occurrence, validating the importance of battery monitoring. The system successfully visualized temperature peaks and voltage drops associated with fault conditions.

Significance and Claims
The paper claims that the proposed framework addresses the limitations of traditional fault detection by offering a scalable, adaptive, and self-adaptive solution. By combining deep learning with real-time visualization, the system enables early fault detection and provides operators with clear, actionable insights into network health. The authors conclude that while all three models are viable, the RNN offers the optimal compromise between classification performance and computational efficiency for this specific application. The study emphasizes that selecting the appropriate model depends on the specific requirements of the WSN deployment (e.g., prioritizing speed vs. detection sensitivity) and suggests that ensemble approaches could further enhance fault detection capabilities in real-world scenarios.