An IoT-Driven Intelligent Street Lighting System with Integrated Severity-Based Fault Classification

IoT-Enabled Smart Street Lighting Systems: Modern urban infrastructure has increasingly adopted IoT technologies to transform traditional street lighting into smart, connected systems. In a conventional setup, streetlights operate on fixed timers or manual control, requiring periodic human inspection for outages. IoT-based street lighting systems, by contrast, embed networked sensors and controllers in each lamp, enabling real-time remote monitoring and adaptive control. These systems can dynamically adjust illumination levels based on sensor inputs (e.g., ambient light, motion), substantially improving energy efficiency and road safety by providing lighting on demand. For example, Denardin et al. [4] demonstrated an intelligent street lighting control architecture where each lamp’s status and power consumption are reported to a central platform for active monitoring. Pantoni et al. [5] described an edge-assisted network for streetlights allows facility managers to control lights and detect anomalies at scale remotely. This connectivity reduces manual maintenance effort and supports advanced features like scheduling, dimming, and fault alerts. The literature agrees that IoT connectivity is a foundational enabler for smart street lighting in future cities. However, deploying such IoT infrastructure city-wide is not without challenges: high upfront costs, interoperability issues, and the need for stable network coverage (especially in remote areas) remain significant concerns [6]. For instance, Dahan et al. [7] noted that smart lighting projects often face reliability issues in underdeveloped regions due to inconsistent internet connectivity and power supply, underscoring the importance of robust network design. Despite these challenges, IoT-driven lighting systems have laid the groundwork for more intelligent maintenance and control strategies, as discussed next.

Rule-based logic relies on fixed thresholds and therefore struggles in real-world environments where sensor readings are noisy, traffic patterns are highly non-linear, and illumination levels fluctuate due to weather or vehicle headlights. It also assumes complete and consistent data, which is often not the case in IoT deployments where missing or intermittent sensor inputs are common. In contrast, the ML model learns the variability within the data and becomes more robust to noise, missing values, and complex multi-feature interactions that rules cannot easily encode. Our results further show that ML reduces both false positives and false negatives by capturing subtle patterns beyond the deterministic logic, thereby providing more reliable severity predictions. Thus, the ML component adds significant value by improving robustness, handling imperfect data, and enhancing decision accuracy in complex urban conditions where static rule-based algorithms alone are insufficient.

Deterministic IF–THEN rules alone are often insufficient for streetlight monitoring because they cannot reliably handle sensor noise, fluctuating illumination from vehicle headlights, or weather-induced variations. Fixed thresholds (e.g., “light level < X”) frequently misclassify normal conditions as faults or miss critical outages, especially in environments with non-linear and rapidly changing traffic patterns. These rules also lack the ability to account for contextual factors such as time of day, surrounding lamp conditions, or road usage intensity. ML overcomes these limitations by learning complex, non-linear relationships from data, filtering noisy signals, and producing more accurate severity assessments tailored to real-world conditions. Hence, ML enhances decision-making beyond what traditional rule-based approaches can achieve.

2.1 Fault detection and predictive maintenance approaches

Traditional streetlight maintenance: Crews respond to citizen complaints or periodic inspections to fix outages. This approach can lead to prolonged dark periods when lights fail, negatively impacting public safety and confidence [8]. Researchers have explored automatic fault detection using sensor data and predictive maintenance models to mitigate this. In IoT-enabled systems, each lamp can report its operational status (voltage, current, brightness, etc.), allowing immediate identification of failures or abnormalities [9, 10], For example, a telemetry manager for public lighting was developed that continuously monitors lamp health and flags faults in real time via a centralized dashboard. Such systems reduce the need for manual night patrols by generating automatic alerts when a bulb or driver malfunctions [11]. As an illustration, recent work [11] implemented a smart streetlight network where if any light goes faulty, it is detected without human intervention and reported through a web application. This immediate visibility into failures allows maintenance teams to respond faster, shortening dark outages.

Beyond detecting failures as they occur, an emerging trend is predictive maintenance – forecasting lamp failures before they happen. This is achieved by analyzing historical performance data to predict remaining useful life or failure probabilities. Accurate remaining-life prediction enables proactive replacement of lamps prior to burnout, thus eliminating unplanned outages. For instance, Segovia-Muñoz et al. [12] applied degradation models (LM-80/TM-21 standards) to LED streetlights and developed an exponential decay algorithm to estimate lumen depreciation and predict when a lamp will fall below acceptable brightness. By planning maintenance based on such predictions, cities can maximize equipment life and minimize downtime. Several studies report that predictive maintenance regimes, combined with IoT telemetry, drastically reduce maintenance costs and outage durations compared to reactive approaches. Nonetheless, implementing predictive maintenance requires reliable data collection over long periods and robust models; many municipalities are still in the early pilot stages of these technologies. The literature shows a clear evolution from reactive repairs to automated fault detection and prediction-driven maintenance in smart street lighting systems.

AI and ML Integration in Lighting Systems: Applying AI and ML techniques to streetlight data is a relatively recent area of research, but it is gaining momentum. Early smart lighting deployments focused on rule-based control and simple sensor thresholds (e.g., lights dim when no motion is detected). Now, with large volumes of sensor and usage data available, researchers are exploring data-driven AI models to optimize performance and detect complex fault patterns [9]. A systematic review [13] finds that the integration of AI in IoT-based lighting (often termed AIoT) is still in its infancy, with the first notable works appearing only around 2019. Their review highlights that most current smart lighting systems do not fully exploit ML and that incorporating edge-based ML could vastly improve adaptability and reliability.

One promising application of AI is anomaly detection in streetlight operations. Instead of waiting for a light to go completely dark, ML models can analyze subtle deviations in electrical parameters to predict failures or detect malfunctioning components. Śmiałkowski and Czyżewski [9] demonstrate this using energy consumption data from smart meters in a city lighting grid. Their system achieves real-time detection of anomalies such as lamp failures or schedule deviations by employing time-series forecasting (SARIMA models) and neural networks (LSTM). Notably, both statistical and deep learning approaches in that study enabled a self-learning, online fault detection mechanism deployable on edge devices, with the SARIMA model performing best for timely alerts. Another study by Chen et al. [14] developed a narrowband-IoT-based streetlight monitoring system that uses ML classifiers to distinguish regular operation from various fault conditions, improving detection accuracy over simple threshold methods. These works illustrate the potential of AI to enhance maintenance: instead of binary "on/off" sensing, smart algorithms can estimate the severity of degradation, filter out transient anomalies, and prioritize genuine issues for repair. Additionally, AI has been applied for predictive control. For example, the study by Sun et al. [15] built an ML-driven predictive model (using time series analysis) to forecast lighting needs and adjust lamp brightness pre-emptively, achieving significant energy savings while maintaining service quality.

Despite these advances, incorporating AI in municipal lighting is not yet widespread. Key barriers noted in the literature include the need for large labeled datasets for training, system complexity concerns, and legacy infrastructure integration. Nevertheless, the trajectory is clear: AI and data analytics are set to play an increasing role in making street lighting systems more intelligent, from fault diagnostics to adaptive illumination strategies.

Vehicle Presence and Traffic-Aware Lighting: Another vital aspect of smart street lighting is integrating traffic data – specifically, detecting vehicles or pedestrians to modulate lighting or assess maintenance criticality. In urban environments, traffic density varies widely; a streetlight on a busy highway serves far more road users at night than on a quiet lane. Numerous studies have proposed traffic-responsive lighting, where streetlights brighten or turn on only when vehicles or pedestrians are present and dim otherwise, thereby saving energy without compromising safety. Qaisar et al. [16] built an early prototype using infrared sensors to sense vehicle movement and then activating lights sequentially as vehicles passed by, which demonstrated the basic feasibility of on-demand illumination. In recent years, more sophisticated approaches have leveraged computer vision and radar. These systems highlight that coupling traffic sensing with lighting control can maintain safety (adequate illumination when cars or pedestrians are present) while optimizing energy use when roads are empty.

Beyond energy optimization, traffic data can inform maintenance decisions. A smart maintenance system might prioritize fixing a streetlight on a high-traffic thoroughfare over one on a seldom-used street since the impact of an outage differs. Some frameworks have begun to incorporate such considerations. For example, Gagliardi et al. [17] developed a multi-sensor streetlight node that monitors lamp electrical parameters and environmental and usage indicators, such as motion or traffic volume, to provide a holistic view of each lamp's importance and condition. In one case study, traffic-aware lighting reduced complaints, as lights on busy routes were rarely allowed to fail unnoticed. Moreover, adaptive systems can change lighting levels based on real-time traffic density. If a normally quiet street suddenly experiences heavy traffic (e.g., due to a detour), smart lights can increase brightness to enhance visibility. This on-demand adaptability improves efficiency and safety and is a key selling point of IoT-based street lighting. However, integrating continuous traffic monitoring comes with data and communication overhead. Recent works suggest using edge computing to process video or sensor data locally at the lamp post, which minimizes latency in reacting to vehicles and reduces bandwidth usage [9]. In summary, vehicle detection and traffic awareness have become integral to advanced streetlight systems, enabling context-aware operation where lighting and maintenance prioritize locations with more excellent human activity or vehicular flow.

Safety Implications and Severity Assessment in Lighting Failures: Street lighting is closely tied to public safety, and research indicates that insufficient lighting can elevate both accident and crime risks. Welsh and Farrington’s meta-analysis of crime prevention (2008) reported that improved street lighting was associated with a significant (~20%) reduction in nighttime crime in the areas studied. More recent field experiments support this: a randomized trial in New York City public housing [18] found that adding temporary lighting led to a 36% reduction in serious nighttime outdoor crimes. Conversely, lighting outages tend to increase fear and potential opportunity for crime. City incident data show that residents are acutely sensitive to lighting failures – in New York City's 311 service request system (2010-2016), streetlight outages were the third most common complaint type (about 5% of ~15 million complaints). Maintaining consistent illumination is critical for perceived and actual safety. A dark street not only hampers visibility for drivers and pedestrians, raising the risk of accidents, but also may become a locus for opportunistic crime. This is especially true in remote or sparsely populated areas, where a single failed light can leave a long dark stretch with no nearby light sources.

Literature on smart lighting increasingly calls for severity-aware maintenance strategies that prioritize repairs based on risk factors. Key factors identified include the number of adjacent lights out, the location's characteristics (crime rate, traffic volume), and the availability of alternate lighting. If consecutive streetlights fail, the combined dark span is more significant, compounding safety risks exponentially. However, most technical studies have focused on detecting faults rather than ranking their urgency. City engineering guidelines sometimes note that multiple contiguous outages should be the highest priority (as seen in municipal maintenance policies), but academic research on automated severity scoring is sparse. One relevant thread is in power distribution networks, where faults are classified by criticality to prioritize grid repairs; analogous concepts are now applied to smart city lighting. For instance, Yoomak and Ngaopitakkul [19] conducted a feasibility analysis of solar street lighting and noted the importance of fast response in critical outage scenarios to ensure public safety – essentially highlighting that not all lamp failures are equal in impact. Despite such insights, there is a gap in the literature: few, if any, AI-driven streetlight maintenance models explicitly combine crime data, traffic density, and outage patterns to compute a composite severity index for each failure incident. This represents an open research challenge and an opportunity for innovation, as our proposed system aims to address.

Research Gaps and Emerging Challenges: The above review shows that while significant progress has been made in IoT-based street lighting and smart maintenance, several gaps remain for future work. First, the integration of multi-modal data for severity analysis is largely missing. Prior studies treat energy efficiency, fault detection, and public safety independently. An ingenious maintenance system would ingest diverse data – lighting status, sensor readings, traffic counts, neighborhood crime statistics – and intelligently prioritize interventions. Such a holistic approach is only now becoming feasible with AIoT advancements. Some research notes that the convergence of AI with IoT for lighting is nascent, and many possibilities (like advanced decision-making algorithms at the edge) remain unexplored. Our review found no published framework that, for example, increases a fault's priority because it occurs in a high-crime, low-traffic area on a dark stretch – a scenario our proposed system explicitly targets.

Second, scalability and real-time performance pose challenges. City-wide deployments may involve tens of thousands of lights. Detecting faults and computing severity in real time requires efficient distributed computing. Edge computing has been suggested as a solution to reduce latency and network dependency, but deploying AI models on edge devices (like lamp-mounted controllers) must overcome constraints in power and processing. Early experiments with edge ML in lighting (e.g., using microcontrollers for anomaly detection) show promise, yet issues of the edge devices' model updating, security, and maintenance are open questions. Cybersecurity is another concern: as streetlights become networked endpoints, they could be targets of hacking or data manipulation, which can have dangerous implications if not addressed (e.g., false outage reports or malicious switching off of lights). Ensuring robust encryption and fail-safes in the maintenance system is thus crucial, though few papers in this domain delve into security beyond basic network encryption.

Lastly, there is the challenge of interdisciplinary integration. Smart lighting maintenance does not exist in a vacuum; it overlaps with urban planning, law enforcement, and transportation management. Lighting improvements have been linked to lower crime and increased community confidence and economic activity after dark. Thus, city authorities are interested in solutions that can justify maintenance investments in terms of broader social benefits. However, academic studies seldom quantify these broader impacts when proposing technical solutions. This gap suggests the need for collaborative research that combines engineering with social science to evaluate how AI-based maintenance scheduling might influence crime patterns or traffic safety over time.

The literature study shows a clear trend toward smarter streetlight infrastructure with IoT connectivity, automated fault detection, and preliminary uses of AI. Energy efficiency and basic adaptive lighting are well-documented benefits of such systems. The novel contribution needed – and which our proposed system aims to provide – is a unified AI-driven maintenance framework that factors in the severity of streetlight faults by analyzing their spatial context (e.g., consecutive dark lamps), safety implications (crime risk, remote area), and real-time usage (vehicle and pedestrian density). By addressing these gaps, the next generation of smart street lighting systems can save energy, reduce downtime, and improve urban safety and quality of life. The study by Mona et al. [20] has shown how the IoT plays a crucial role in the development of smart cities and smart homes, particularly in enhancing public safety and security. The open challenges identified (data integration, edge AI deployment, and cross-domain impact evaluation) form the basis for the research questions the proposed work will tackle as we strive to advance state-of-the-art smart city lighting maintenance.

4.1 ML model performance

We first compare the performance of the three ML models (Logistic Regression, Decision Tree, and Random Forest) on the severity classification task. The models were tested on the held-out 30% of the data (300 samples). Figure 4 shows the accuracy achieved by each model.

Figure 4. Accuracy of different ML models for severity classification

Figure 4 shows the accuracy of different ML models for severity classification. The Random Forest classifier attained the highest accuracy (~94.3%), slightly outperforming Logistic Regression (~94.0%) and notably outperforming the Decision Tree (~92.3%). All models performed relatively well, with accuracies above 92%, indicating that the features in the dataset contain sufficient information to distinguish the severity levels in most cases. The Random Forest's slight edge likely comes from its ability to handle the non-linear interactions between features (for example, the combination of "Nighttime + Power State Off + high traffic" indicating a certain severity) better than a single Decision Tree or linear model. Despite its simplicity, we also note that Logistic Regression did quite well; this suggests that the underlying decision boundaries between classes might be approximately linear or that a few key features dominate the prediction (indeed, whether a fault is isolated or not could be a primary separator). The Decision Tree's performance was lower, possibly due to limited depth (we pruned it to prevent overfitting) or some minor classes (Low, Medium) that were more challenging to identify with limited data.

To gain deeper insight into how well the model distinguishes the individual severity categories, we present the confusion matrix of the best-performing model (Random Forest) in Figure 5.

Figure 5. Confusion matrix for severity classification (Random Forest model)

This matrix breaks down the model's predictions versus the true labels for the four classes: Normal (no issue), Low, Medium, and High severity. The diagonal elements (in dark blue) represent correct predictions. We can see that Normal status (no fault) and High-severity faults are identified almost perfectly by the model. All 226 normal test instances (no fault) were correctly classified as Normal, and all 56 instances of truly High severity were predicted as High. This is important because High-severity cases are the most critical to catch (we want no false negatives in that category), and the model achieved 100% recall for High-severity cases in this test.

On the other hand, the matrix shows that cases of Low and Medium severity were often misclassified. In the test set, there were seven instances of Low and 11 of Medium severity (these are relatively small numbers). The model tended to predict most of those as High. For example, out of 11 true Medium cases, 10 were predicted as High, and only one was correctly as Medium; none of the 7 Low cases were predicted as Low (all were marked High). This indicates a bias of the classifier to err on the side of higher severity for ambiguous cases. From an application perspective, this bias is not entirely undesirable – it means the system would treat some moderate issues as if they were high priority, which is a safer error than underestimating a serious fault. However, it also means the model is not yet finely distinguishing the nuances between isolated minor faults (Low) and more significant single faults (Medium). The likely reason is the class imbalance and the rule-based nature of our labels: most fault instances were either High (adjacent outages) or labeled Medium due to an extreme context, with very few true Low examples. The model thus had limited data to learn what constitutes a "Low" severity fault, and it defaulted to classifying any fault as at least Medium/High. In a more balanced training scenario (or with further model calibration), we would aim to improve the precision for Medium and Low classes. As per-class performance summary (precision, recall, and F1-score) for the Random Forest classifier, which achieved the highest overall accuracy. Table 1 presents these metrics for the four severity classes.

Table 1. Per-class precision, recall and F1-score for the Random Forest model

Overall, the Random Forest model successfully detected all actual fault cases (no missed faults, since no Normal was misclassified as something else) and incorrectly flagged the truly urgent cases. The accuracy of ~94% reflects that out of 300 test instances, only ~18 were misclassified (and these misclassifications essentially treated some Medium/Low as High). The high accuracy and recall for critical cases demonstrate that an AI model can reliably automate the decision process encoded by our severity rules. In practical terms, this means the system can take the sensor inputs and immediately determine the appropriate level of response with high fidelity, matching what an expert might decide using the same information.

4.2 Visualization of severity vs. traffic

An important aspect of our system is integrating traffic (vehicle density) data into the severity determination. We performed an analysis to understand the relationship between traffic levels and the severity assigned to our results. Intuitively, one expects higher traffic to lead to higher severity classifications (because more people are affected by an outage). However, extremely low traffic (remote areas) can also lead to an elevated severity due to security concerns. Our model was designed to account for both extremes by bumping severity for high or low traffic conditions. We examined the distribution of Traffic Density values for each severity category in the dataset. The correlation between numerical severity level (treating Normal = 0, Low = 1, Medium = 2, High = 3) and traffic density was near zero (Pearson correlation ~0.005), indicating no simple linear relationship. This is not surprising because the rules were nonlinear: Medium severity occurs at both ends of the traffic spectrum (very high or very low traffic). High severity was dominated by the adjacency of faults (independent of traffic). We did observe the following trends: Medium severity incidents largely corresponded to cases with either very low traffic (bottom 10th percentile) or very high traffic (top 10th percentile) – consistent with our definition that those contexts raise a single outage's importance. High-severity incidents (multiple outages) occurred across a range of moderate to high traffic densities; in about 73% of high-traffic fault cases, the system ended up classifying them as High severity, and similarly, about 75% of low-traffic fault cases were High severity (those had multiple outages in remote areas). Meanwhile, Low severity was assigned only to a few faults under moderate traffic conditions without other exacerbating factors.

The dataset evaluation confirms that incorporating vehicle flow data into the decision process can differentiate scenarios that would otherwise look similar from a purely electrical perspective. In our ground truth labeling, a single lamp failure during rush hour got a higher severity (Medium) than an identical failure late at night with light traffic (Low severity). Our AI model learned this pattern to some extent (though it tended to over-predict High, as noted). This indicates that the model is sensitive to the features of traffic and timing that we intended it to be. In a deployed system, one could further refine this by providing the model with more direct criticality indicators (for example, a precomputed risk score for each location that considers crime rates and traffic together). Our results demonstrate that fusing IoT sensor data with contextual city data (traffic) effectively assesses maintenance priority. The classification performance achieved suggests that city maintenance departments could trust such a model to triage streetlight outages automatically.