Intelligent Fault Detection in Fog Computing Using Machine Learning for IoT Applications

The Internet of Things (IoT) revolution has brought about tremendous data output from device-associated data sources, for which efficient low-latency processing architectures are required. Fog computing, as the middle-layer between centralized cloud data centers and IoT sensors, fulfills the requirements of such applications through edge-based processing, storage, and decision-making capabilities. It reduces latency, bandwidth consumption, and dependence on centralized cloud facilities and is therefore best suited for time-critical and mission-critical applications. Despite its advantages, fog computing infrastructures are highly prone to faults such as node failure, network failure, and lack of resources since they are highly distributed and resource-constrained environments. Classic fault detection mechanisms such as Rule-Based Systems (RBS), Statistical Process Control (SPC), and Simple Threshold-Based Detection (STBD) are not very efficient when it is required to deal with dynamic fog environments [1]. These are not highly scalable and flexible and are prone to high rates of false positivity, thereby compromising the reliability as well as the safety of IoT service delivery. Today's trends hold the integration of Machine Learning (ML) concepts within Fog architectures with the aim of increasing fault detection accuracy and adaptability. Supervised and unsupervised ML models are gaining favor to handle complex patterns and capture anomalies in real-time. Approaches such as Support Vector Machines (SVM), Decision Trees (DT), and Ensemble Learning are gaining favor for their prediction accuracy and fault diagnosis strength [2]. Fault-tolerant fog computing applications span numerous industries, including smart healthcare monitoring, smart transport systems, industrial automation, and intelligent energy grids—whose applications require real-time action and persistent service. Aspects of providing intelligent and reliable fault detection for Fog-IoT environments are critical to sustaining present contemporary digital infrastructures [3].

Fog-IoT environments require intelligent mechanisms capable of detecting faults in distributed nodes with minimal latency and high reliability. Although several traditional detection approaches exist, they often fail to adapt to dynamic IoT environments. Therefore, this study proposes an Intelligent Fault Detection Model (IFDM) based on ML to enhance anomaly detection in fog computing infrastructures. The next section reviews related research work in fog computing fault detection and highlights the limitations of existing approaches.

1.1 Research gaps

Whereas fog computing is today regarded as a promising architecture to support low-latency and real-time applications for IoT scenarios, fault detection for such decentralized scenarios is not yet able to surmount many open research issues [4, 5]. Older schemes, such as RBS and threshold-based detection, are not flexible enough and are not efficient enough to respond well to the dynamic and heterogeneous fog environment. Further, most existing models miss the full potential of the available real-time data from sensors and the network and consequently incur delay for fault identification and more chances of missed anomalies. Most serious is the high false positive rate faced by many ML-based models due to noisy or class-imbalanced data. On top of this, most existing models are built on computationally intensive algorithms that are not suitable for fog-limited nodes and, therefore, practical deployment is limited [6]. There is not much extensive evaluation done for the existing models on different IoT scenarios to further limit the generalizability and robustness of the models. Most existing detection mechanisms are primarily standalone and are not frequently integrated with fault recoveries or self-healing mechanisms. These open research gaps urgently call for lightweight fault detection frameworks that are adaptive and intelligent to promote the reliability and efficiency of real-time Fog-IoT scenarios [7].

The next section presents the related work on fault detection in fog computing environments and highlights the limitations of existing approaches.

1.2 Major contributions of the work

The main things this work adds are as follows. First, we suggest a lightweight IFDM for IoT environments that use fog and telemetry-driven ML. Second, a multi-stage method is created that includes preprocessing, feature reduction, supervised classification, and fault decision fusion. Third, we create a resource-aware decision-making process by combining the anomaly score, classifier confidence, and resource-adjusted detection latency (RADL) into one Fault Decision Equation (FDE). Fourth, the proposed framework is better suited for real-time fog deployment than static traditional methods and more advanced options that require more processing power. Finally, the experimental analysis shows that the new method has a higher detection rate, fewer false positives, and a more reliable system than the old method.

1.3 Related work

Tran-Dang et al. [8] proposed a reinforcement learning (RL)-based approach for resource management in fog computing, aiming to intelligently handle task offloading and resource allocation in dynamic environments. The innovation lies in the application of RL to adaptively make decisions, reducing latency and improving performance under high request loads. However, a significant drawback is the lack of practical deployment and performance validation in real-time fog networks, especially considering the complex heterogeneity and varying workload conditions. Mukherjee et al. [9] presented a comprehensive overview of security and privacy challenges in fog computing environments. The innovation is the categorization of fog-specific security threats arising from mobility, heterogeneity, and geo-distribution, which are not adequately addressed by traditional cloud security methods. The drawback is the absence of proposed solutions or frameworks; the paper is limited to a theoretical analysis without implementation or experimental insights. Wu et al. [10] introduced a six-layer IoT Cloud-Edge architecture equipped with a reconfigurable mixed-signal controller for smart meters, targeting fast arc fault detection. The innovation lies in the novel architecture design that enhances detection speed, reduces bandwidth usage, and lowers computational costs in varying electrical load scenarios. However, the drawback is the complexity of implementing a six-layer architecture in existing IoT infrastructures and the need for custom hardware components, which may hinder scalability. The paper [11] reviews the application of fuzzy theory in fog computing, offering a detailed taxonomy of techniques across domains like trust management, intrusion detection, and resource allocation. The innovation is a comprehensive classification based on performance evaluation tools and fuzzy methods used. The drawback is that it lacks focus on real-time performance or energy efficiency trade-offs when fuzzy logic is applied in constrained fog environments. The study [12] proposes a fog-based green VANET (Vehicular Ad Hoc Network) infrastructure using self-powered devices like solar routers and smart cameras to enhance network sustainability and traffic sensing. The innovation is the integration of green energy and fog computing for a self-sustained, secure VANET system. The drawback is the dependency on environmental conditions for power generation and the high cost of deploying solar-powered units across a city-scale network. Zhao et al. [13] proposed BPRM, a blockchain-based, privacy-preserving, and multifunctional data aggregation scheme for fog-assisted smart grids. The innovation includes support for statistical data analysis without relying on a trusted third party, using smart meter consensus and batch verification to ensure data integrity. The drawback is the increased computational and storage overhead introduced by blockchain, which may challenge fog nodes with limited resources.

This paper [14] presents a Kademlia-based Distributed Hash Table (DHT) framework for alert fusion in distributed intrusion detection systems (IDS) using fog computing. The innovation lies in the fusion of similar alerts at collector nodes, reducing redundant message generation by 62% while maintaining over 80% detection accuracy. However, the framework is tailored for a specific dataset (DARPA 1999), raising concerns about its adaptability to modern and diverse IoT attack vectors. Siddiqui et al. [15] conduct a systematic literature review on Software-Defined Networking (SDN)-based IoT frameworks, highlighting their role in enhancing scalability, fault tolerance, and load balancing. The innovation is a detailed taxonomy of SDN frameworks for IoT, categorized by functions like virtualization, OpenFlow, and blockchain integration. The drawback is that while the review identifies architectural improvements, it does not experimentally validate any SDN-based solutions or discuss real-time deployment constraints.

Liu et al. [16] proposed a novel approach for detecting incipient faults in medium voltage cables by considering the degradation behavior of solid insulation during fault progression. The study analyzes the influence of insulation carbonization on fault transient characteristics and performs incipient fault detection using the transient variation of the zero-sequence current waveform. A diagnostic method based on the distortion rate of the zero-sequence current waveform is developed to accurately identify fault states. Simulation and experimental results demonstrate the effectiveness of the proposed detection mechanism in identifying early-stage faults in power cables. The innovation of this work lies in integrating insulation degradation behavior with transient waveform analysis for more accurate fault diagnosis. However, the approach may require precise waveform measurements and controlled experimental conditions, which could limit its applicability in large-scale real-time monitoring environments.

Zhao et al. [17] introduced a dynamic fault accommodation framework for nonlinear uncertain systems affected by multiple process faults. The method utilizes information obtained from fault detection and isolation modules to dynamically update the fault accommodation controller. The proposed framework incorporates different fault isolation modes, such as zero isolation, partial isolation, full isolation, missed isolation, and initialized isolation, to handle multiple faults effectively. A nominal controller maintains system performance before fault detection, while an adaptive controller reconfigures the control strategy after faults are identified. The innovation of this work is the development of a dynamic fault accommodation strategy that adapts to the availability and quality of fault isolation information. However, the method increases system complexity and may require significant computational resources for real-time implementation in highly dynamic environments.

Kong and Nian [18] proposed a fault detection and location method for mesh-type DC microgrids based on an improved Pearson correlation coefficient. The method analyzes the similarity between sampled line currents and reference currents within a movable time window to detect faults. Fault identification and location are achieved by comparing transient current signals with estimated values generated through a genetic algorithm. The correlation coefficient is used as the key parameter for recognizing fault types and determining fault positions. The innovation of this work lies in combining correlation analysis with evolutionary optimization techniques to enhance the accuracy of fault detection and localization. However, the method may experience delays in fault detection due to the time window processing and optimization iterations required for estimating fault locations.

Yadegar et al. [19] presented a distributed high-impedance fault detection and protection scheme for DC microgrids. The proposed method considers the nonlinear characteristics of high-impedance faults, including low current amplitude, intermittence, and high-frequency components. A distributed detection mechanism is designed to identify the faulty section and provide both forward and reverse fault discrimination. The effectiveness of the scheme is validated through simulation and experimental evaluations under different operating conditions and noise scenarios. The innovation of this work is the development of a distributed protection mechanism capable of detecting high-impedance faults that are typically difficult to identify using conventional relay techniques. However, the distributed detection framework may require multiple sensors and communication coordination, which could increase system implementation cost and complexity.

Based on the limitations identified in the existing literature, a ML-driven fault detection architecture is designed to improve detection accuracy and system reliability in fog-IoT environments. The proposed framework integrates telemetry-based anomaly detection and intelligent decision mechanisms to identify abnormal device behavior in real time [20].

Several recent studies have explored the use of ML techniques for fault detection in fog computing environments. These approaches improve detection accuracy and adaptability in dynamic IoT systems. However, many existing methods suffer from limitations such as high computational complexity, a lack of real-time adaptability, and limited validation in heterogeneous fog environments. Therefore, there is a need for lightweight and IFDM that can operate efficiently in resource-constrained fog nodes. The proposed IFDM model addresses these challenges by integrating telemetry-based anomaly detection with ML-driven decision mechanisms [21].

1.4 Proposed methodology of Intelligent Fault Detection Model for fog-based IoT fault detection

Figure 1 shows how the proposed IFDM for fog-based IoT applications would work in practice. The first step is to collect real-time IoT telemetry data and clean it up so that it is of better quality. In the second step, important features are chosen and extracted to help with fault classification. In the third stage, a Random Forest model at the fog node looks at the processed data and finds behaviour that is out of the ordinary or wrong. At the end, a resource-aware decision mechanism uses the anomaly level, confidence score, and resource condition to make the final fault decision. This workflow helps find faults quickly and reliably in foggy conditions.

Figure 1. Four-stage Intelligent Fault Detection Model (IFDM) workflow for real-time fault detection in fog computing

1.5 Methodological novelty of Intelligent Fault Detection Model in fog-based IoT fault detection

Figure 2 shows the most important new methods that the proposed IFDM for fog-based IoT environments uses. The first part shows the integrated framework, which combines anomaly analysis, ML, and decision fusion into one structure. The second part shows telemetry-driven analysis, which uses real-time IoT data to find strange behaviour. The third part talks about lightweight fog deployment and makes it clear that the suggested method can be used on fog nodes with limited resources. The last part shows the resource-aware final decision, which uses the anomaly score, confidence score, and latency together to find the last fault. The model is meant to help with accurate, quick, and real-time fault detection in fog computing environments as a whole.

Figure 2. Four core methodological novelties of Intelligent Fault Detection Model (IFDM)

Figure 4 is the working flow of the subject-proposed IFDM for fog computing-based IoT fault detection. IoT sensors generate relentless telemetry, which undergoes Feature Extraction followed by Preprocessing. The preprocessed data is subjected to the Data Collection phase through the fog level. The data thus collected is provided as input to the IFDM, where data is checked based on the trained ML model. Based on the check-up, it detects abnormal behavior or fault and raises up Fault Alerts when necessary. The fault detected is stored and monitored through the Fault module. Optimal Feature Extraction is embedded through a Feedback loop for real-time adaptation as well as for improved detection model precision. Such is the architecture ensuring fog level intelligent decision-making that is solid and expandable for IoT-based intelligent environments.

Figure 4. Functional flow diagram of Intelligent Fault Detection Model (IFDM) for fault detection in fog-based IoT system

3.1 Weighted Fault Probability estimation

To quantify the likelihood of failure in an IoT node, the proposed model estimates a Weighted Fault Probability (WFP) using a weighted sum of critical features such as CPU load, temperature, and network delay and it is determined by Eq. (4):

$\mathrm{WFP}_i=\frac{w_1 \cdot C_i+w_2 \cdot T_i+w_3 \cdot D_i}{w_1+w_2+w_3}$   (4)

where, C_i is CPU utilization, T_i is temperature, D_i is network delay, and w₁, w₂, w₃ are their respective weights.

3.2 Fault Detection Confidence Score

To assess the confidence of the ML model in identifying a fault, a Fault Detection Confidence Score (FDCS) is computed from the softmax output of the model and it is represented by Eq. (5):

$\mathrm{FDCS}=\max \left(\sigma\left(z_1\right), \sigma\left(z_2\right), \ldots, \sigma\left(z_n\right)\right)$   (5)

where, $\sigma\left(z_k\right)$ is the softmax probability of class $k$, and $n$ is the number of fault classes.

3.3 Resource-adjusted detection latency

Eq. (6) calculates detection delay while considering available fog resources. It helps balance performance with computational cost and it is represented by Eq. (6):

$\mathrm{RADL}=\frac{T_d}{1+\alpha R}$   (6)

where, $T_d$ is raw detection time, $R$ is the available resource score (e.g., CPU, RAM), and $\alpha$ is a resource adjustment constant.

3.4 Anomaly score aggregation for multi-device

For environments with multiple IoT devices, this function computes a cumulative anomaly score to detect systemic faults and it is represented by Eq. (7):

$\mathrm{ASAMD}=\frac{1}{m} \sum_{i=1}^m\left(\frac{A S_i-\mu_{A S}}{\sigma_{A S}}\right)^2$   (7)

where, $A S_i$ is the anomaly score of device $i, \mu_{A S}$ is the mean, $\sigma_{A S}$ is the standard deviation, and $m$ is the number of devices.

3.5 Final proposed Fault Decision Equation

The proposed FDE integrates telemetry anomaly score, ML confidence score, and resource-adjusted latency to determine whether a device is in a fault state as determined by Eq. (8):

$\mathrm{FDE}_i=\gamma \cdot A S_i+\delta \cdot F D C S_i-\eta \cdot R A D L_i$   (8)

where, $F D E_i$ is the final fault score for device $i, A S_i$ is the normalized anomaly score, $F D C S_i$ is the ML model's confidence score, $R A D L_i$ is the RADL and $\gamma, \delta, \eta$ are empirically determined weights based on system tuning. If $F D E_i \geq \theta$, where $\theta$. is a predefined fault threshold, the device is classified as faulty.

The proposed IFDM can be applied in real-world IoT environments such as smart city infrastructure, healthcare monitoring systems, and industrial automation networks. In these applications, the model can support abnormal behavior detection, improve reliability, and enable timely fault identification in fog-based distributed systems.

3.6 Computational complexity and deployment feasibility of Intelligent Fault Detection Model

The proposed IFDM is designed for fog environments with limited computation, memory, and energy. Its main operations include data preprocessing, feature reduction, Random Forest-based inference, and final decision fusion. Preprocessing and feature handling are lightweight during runtime, while the Random Forest classifier remains suitable for fog deployment due to its moderate computational cost compared to deep learning models. The final decision stage adds very low overhead because it only uses weighted scoring and threshold comparison.

From a deployment perspective, IFDM is practical for resource-constrained fog nodes because it uses compact telemetry features and a lightweight supervised model. The memory requirement is mainly for storing model parameters, selected features, and intermediate decision values, which is lower than that of deeper architectures. The inclusion of resource-adjusted latency and gateway load analysis further improves runtime feasibility. Hence, IFDM is suitable for real-time fog-based IoT fault detection with low latency and moderate resource usage.

The proposed IFDM effectively addresses the challenges of fault detection in Fog-based IoT environments by leveraging ML techniques. Through comparative analysis with conventional methods—RBS, SPC, and STBD—the IFDM demonstrated significant performance improvements, achieving a 23.6% increase in detection probability, 18.2% reduction in false positives, and a 21.4% rise in system reliability. These results validate the robustness, adaptability, and precision of the IFDM in dynamic and distributed computing scenarios. For future enhancements, the model can be extended by incorporating federated learning to preserve data privacy and real-time adaptive learning to dynamically adjust to evolving fault patterns. Integration with blockchain can further ensure the security and integrity of fault logs. Additionally, deploying the model in real-time smart city infrastructure, healthcare systems, and industrial IoT platforms will validate its scalability and practical applicability. Further optimization for energy efficiency and computational resource management will make the solution more sustainable and deployment-ready for large-scale heterogeneous networks.

Although the proposed IFDM shows improved fault detection performance, real-world deployment may face challenges due to limited computation, memory, and energy availability in fog nodes. Hence, further optimization is needed for efficient large-scale implementation.