Machine Learning-Based Predictive Maintenance at Smart Ports Using IoT Sensor Data

Abstract

Maritime transportation plays a critical role in global containerized cargo logistics, with seaports serving as key nodes in this system. Ports are responsible for container loading and unloading, along with inspection, storage, and timely delivery to the destination, all of which heavily depend on the performance of the container handling equipment (CHE). Inefficient maintenance strategies and unplanned maintenance of the port equipment can lead to operational disruptions, including unexpected delays and long waiting times in the supply chain. Therefore, the maritime industry must adopt intelligent maintenance strategies at the port to optimize operational efficiency and resource utilization.

Towards this end, this study presents a machine learning (ML)-based approach for predicting faults in CHE to improve equipment reliability and overall port performance. Firstly, a statistical model was developed to check the status and health of the hydraulic system, as it is crucial for the operation of the machines. Then, several ML models were developed, including artificial neural networks (ANNs), decision trees (DTs), random forest (RF), Extreme Gradient Boosting (XGBoost), and Gaussian Naive Bayes (GNB) to predict inverter over-temperature faults due to fan failures, clogged filters, and other related issues.

From the tested models, the ANNs achieved the highest performance in predicting the specific faults with a 98.7% accuracy and 98.0% F1-score.

1. Introduction

The maritime industry is considered the backbone of international trade, moving more than 80% of goods around the globe [1–3]. The OECD "Ocean Economy in 2030" report predicts maritime industries will double their global value creation by 2030, driven by shipping, shipbuilding, marine equipment and offshore energy demand [2] — predictions fully dependent on technological innovation. Pre- and post-COVID-19 economic growth and rising cargo volumes drove ports to expand capacity unexpectedly, with shipping lines and alliances concentrating capacity in fewer ports serving more ships, and vessel sizes growing (from an average 1k TEUs in the 1970s to 7.7k TEUs ordered today [4]), shifting congestion pressure onto terminals and demanding optimal handling performance — amid persistent bottlenecks: inefficiencies, labor accidents, energy consumption and emissions.

In parallel, new digitization paradigms — cloud-edge computing, IoT, big data and machine learning — created new possibilities leading to Industry 4.0 [5], with Port 4.0 services connecting all equipment and systems in real time for seamless data exchange and more automated, interoperable solutions at lower risk, cost and lead time [6]. Timely fault detection is therefore crucial to avoid long downtime and performance degradation [7, 8]. With growing seaborne trade and ship numbers, congestion and safety problems arose in ports worldwide [9] — Figure 1 shows the EUROGATE Container Terminal Limassol (ECTL) at the Port of Limassol, Cyprus, which provided the operational data and real-world evaluation. Many terminal decisions and inspections remain manual (machine health, container seals), hindering efficiency [10], and unplanned maintenance of quay cranes and straddle carriers degrades terminal performance — demanding the latest fault-detection and fast-maintenance technologies. Intelligent fault detection and predictive maintenance reduce unplanned activities: predictive maintenance (PdM) gives up-to-date health knowledge, avoiding both waiting for shutdowns (reactive maintenance) and unnecessary servicing (preventive maintenance), especially for STS and RTG cranes, straddle carriers and reach stackers needing optimal maintenance under heavy loads, long hours and diverse weather.

Traditionally, CHE preventive maintenance used SCADA systems with human-coded thresholds, warning rules and configurations — a semi-manual approach ignoring complex dynamic machine behavior and whole-process contextual data. Thanks to AI and IoT advances, ML-based solutions are the next functional step toward cost savings, predictability and availability — training algorithms to detect anomalies and test correlations across IoT data streams from sensors, PLCs and weather stations; deep reinforcement learning such as DDPG has been effectively employed for joint time and energy management in intelligent sensors [11]. ML-based PdM now plays essential roles across aircraft engines [12, 13], smart vehicles [14], energy systems [15–18], elevators [19] and marine engines [3, 20, 21]. This work proposes and tests an ML framework for PdM of CHE at smart ports with threefold contributions: (1) a comprehensive data-collection framework designed and deployed at the Port of Limassol terminal — multiple sensors on straddle carriers and other CHE plus frameworks gathering real-time operational data, extensively preprocessed (cleaned, structured) for integrity and usability; (2) five ML models for CHE over-temperature faults (ANNs, DTs, RF, XGBoost, GNB) plus a tailored statistical model for straddle-carrier hydraulic anomaly detection; (3) extensive validation experiments in real-world PdM scenarios. The remainder: Section 2 literature review, Section 3 methodology, Sections 4–5 the two case studies (hydraulic anomalies; inverter over-temperature prediction), Section 6 discussion, Section 7 conclusions.

2. Literature Review

Predictive maintenance applied over recent decades in industrial settings improved maintenance scheduling, avoided downtime and balanced cost–benefit [22–24]; refined techniques with AI/ML algorithms represent a breakthrough — though AI/ML-led models are yet to reach terminal yard machinery, prior examples illustrate expected results. The automobile industry applied PdM to air-pressure systems, comparing two ML classifiers to save money by predicting faults [25]; Shafi et al. [14] predict faults in four vehicle subsystems (ignition, fuel, cooling, exhaust) from normal/faulty driving data with four classifiers (decision tree, SVM, random forest, nearest neighbor). Manufacturing benefits similarly: a production-line PdM system using IoT data with Random Forest and XGBoost detects failures ahead of time [26, 27], with deeper systematic reviews in [23, 28].

For aircraft, a PdM engine model using autoregressive moving average (ARMA) compared six data-driven models with SVM on top [12]; a novel hybrid DQINN (deep belief network plus quantum-inspired neural network) for fuel systems achieved higher accuracy [13]. For wind turbines, a Random Forest model with three modules (per-turbine predictor, a monitoring agent predicting next-hour failures every 10 min, and a visualization dashboard) was built [15], and a DNN on lubricant-pressure data for gearbox health beat counterparts on real Chinese wind-farm data [29]. In maritime, a systematic big-data/ML approach for marine-engine PdM was validated in a Norwegian-shipping-company case study with good accuracy [20]; studies [21, 30] predict ship-engine and component faults effectively; and an ML approach for ship electrical propulsion PdM [3] uses Balanced Random Forest with multiple-instance learning on event logs to predict failure likelihood, time-to-failure and explainable predictions for timely crew intervention over two years of seagoing events, outperforming alternatives.

Ref. [31] explored PdM feasibility for container-handling cranes with traditional ML (logistic regression, genetic algorithm) on four years of simulated synthetic data for hoist motor and gearbox — the genetic algorithm winning. Ref. [32] proposed AI-driven PdM for port handling equipment from maintenance logs: real data from a Tunisian port trained a supervised K-Nearest Neighbors model on 13 features to classify forklift status (working/failing) with only 64% average accuracy under 5-fold cross-validation on very small data. Overall, much PdM work exists across industries, but maritime still needs substantial work for higher port — especially container-terminal — performance: no literature methods specifically target port-machinery fault prediction from operational data. This study therefore investigates lightweight ML models (ANN, DT, RF, XGBoost, GNB) for container terminals, since advanced deep learning demands substantial compute, longer training and heavier tuning — challenging real-time edge inference and deployment — selecting models for computational efficiency, interpretability and deployability on resource-constrained edge devices such as industrial sensors and IoT gateways.

3. Proposed Methodology for PdM

The proposed IoT-, big-data- and ML-integrating predictive maintenance approach addresses key technical challenges for port machinery such as CHE: effective ML needs sufficient historical data capturing past-failure patterns (mechanical properties, usage rates, operating conditions). Labeled real-world-condition data is often hard to obtain from limited fault occurrences and sensor noise, and robust validation in such imbalanced scenarios is a key challenge since standard metrics may misrepresent rare-event detection. Even with comprehensive datasets, model selection, training and deployment must fit infrastructure compute constraints — cloud computing supports predictive analytics but remote-server models introduce high latencies and delayed responses depending on connectivity, while edge technology running models locally optimizes predictive-analytics speed and performance — necessitating lightweight models that may preclude complex deep architectures.

A complete data-cycle management toolset provides end-to-end processing by design, adding value at every step through well-defined APIs and user interfaces that cut data-management complexity and improve information reliability, decision certainty, operational performance and profitability. This article's methodology integrates the following data operations per [33]: Generation (organic with digitization — here operational data from PLCs on each SC); Capture (incorporating data into an analytics platform via industrial/computer protocol reads — here Node-RED custom flows on wired IoT gateways per SC acquiring PLC/sensor data and transmitting via MQTT to an edge server); Processing (preprocessing for quality and robustness: deduplication, empty/invalid-field removal, missing-value filling, timestamp synchronization, plus calibration converting raw binary/unitless PLC data into meaningful physical units, mitigating sensor degradation and calibration inconsistency); Storage (reliable, accessible, scalable FAIR-principle systems — time-series in InfluxDB; machine-health, fault-log and maintenance records in LIMBLE, a private cloud-based CMMS at EUROGATE); Analysis (advanced extraction: aggregations, statistics inference, meaningful-block determination, inferred-data association — here InfluxDB plus LIMBLE sourcing ML development: statistical aggregations, event detection, correlation inference, fault classification); Visualization (clear human-interpretable presentation — one graph can encode tens of gigabytes — balancing aesthetics, functionality and visual perception).

Understanding the ML-model life cycle is crucial for port applications; following [34]: Problem Definition and Data Acquisition/Collection (port experts set clear objectives; collect goal-relevant data); Data Preprocessing (exploration of distributions, relationships and bias; cleaning noise/errors; feature engineering for predictive power; train/validation/test splitting for unbiased evaluation); Model Selection and Training (suitable classification/regression algorithms with hyperparameter tuning and cross-validation for robustness); Model Evaluation (accuracy, precision, recall, F1-score against terminal needs); Deployment (integration into existing systems with scalability, efficiency and post-deployment feedback refinement); Monitoring and Maintenance (continuous performance monitoring, drift-triggered retraining, operational-challenge handling). For unsupervised models the approach shifts: objectives around inherent-structure discovery; clustering, autoencoders, statistical or dimensionality-reduction algorithms; silhouette-score, cluster-validity or qualitative evaluation; monitoring for data-pattern shifts rather than predictive accuracy.

The proposed architecture for implementing and testing ML-based PdM over CHEs at EUROGATE Container Terminal Limassol (ECTL) — supporting over 90% of Cyprus trade at ~500,000 TEUs/year with four quay cranes and 36 straddle carriers, 800 m quay length and 16 m depth — is shown in Figure 2: deploy ML algorithms on historical data to detect (1) anomalies and (2) faults in straddle carriers (SCs) performing horizontal port movements, since efficient SC fault prediction boosts performance and cuts machine downtime. As Figure 2 shows, generation, capture and processing run in far-edge gateways collecting sensor and STS/straddle PLC data; processed information flows to the data-lake/storage block, retrieved by the analysis component embedding the algorithms and ML models refining maintenance issues to predict before any failure — aligned with the cloud-edge-IoT continuum concept distributing intelligence across edge and cloud layers, as part of the aerOS EU Project [35]. The two case studies ensure representativeness and generalizability: hydraulic anomalies are subtle, gradual and unlabeled (suited to unsupervised techniques), while over-temperature faults are abrupt with clear failure signatures (allowing supervised learning) — demonstrating framework flexibility across fault types and learning paradigms. Four electric SCs are monitored with Omron CS1G-CPU44H PLCs (1.5 MB program work memory, 5 MB data storage) generating operational time-series, plus Danfoss MBS 3000 hydraulic pressure transmitters (≤±0.5% typical accuracy), IMUF99 inclination/acceleration sensors, VIM32PL vibration sensors, a Kinco HMI and Teltonika RUT240 4G routers — captured via low-code Node-RED [36] custom flows on SC-mounted IoT gateways, transmitted over MQTT [37] to an edge server and onward to data lakes, stored in open-source InfluxDB [38] time-series database with health/fault/maintenance records in LIMBLE (directly from PLCs plus manual technician input) — both platforms exploited for experiments.

4. Case Study 1: Hydraulic Anomalies on Straddle Carriers

Most mechanical actuators inside straddle carriers are hydraulic — parking brake, steering, spreader — so monitoring and predicting hydraulic-system status and health is crucial. Every industrial hydraulic system uses burst-working pumps to hold pressure above a threshold: under normal SC operation the pump activates below a lower threshold and runs continuously to an upper threshold, creating a roughly inverted sawtooth pressure–time curve with ~60 s period when idle. Peak frequency strongly depends on momentary crane behavior — spreader operation demands higher pressure and fires the pump more often. Events such as leaks or clogged fittings noticeably impact system pressure or pump duty cycle (e.g. continuous pressure requests replacing leak-lost air), so monitoring the normally pulsing pump duty cycle can catch issues before significant damage.

Anomaly detection uses a custom variation of the usual Kolmogorov–Smirnov test with the 95th percentile (95 chosen experimentally as the highest value giving satisfactory robustness) instead of the supremum for robustness against extreme values in long-tailed distributions — a traditional statistical approach deemed more suitable given insufficient labels for supervised ML, comparing pressure signals against verified-correct-operation periods (behavior characterization). After simple cleanup and formatting, each observation is binned into states by which hydraulic-actuator subset is in use: spreader pressure measurement (PM7), parking brake (PM33), service brakes (PM25) — binary combinations yielding eight states, some never occurring (service brakes with parking brake engaged — states 1 and 5), with only states 0, 2, 3 and 6 regular (Figure 3: pump-output pressure PM2, the three state-determining signals PM7/PM25/PM33, the color-bar state, and the combination table). Segregating behavior into states homogenizes pump duty cycle in pressure demand, so relevant deviations indicate anomalies — observed by comparing key hydraulic-signal statistics against per-state reference distributions from guaranteed-working-order periods, focusing on the consecutive-peak period (Figure 4: cumulative consecutive-peak-period distribution of PM2 differs across states yet stays consistent within each).

Comparison computes residuals between reference quantiles and most-recent-hour distribution quantiles per state across different quantiles of absolute residuals. The two anomaly-flagging estimators are q0.1 and q0.9 from the distribution of |Fref(x) − Fobs(x)| for x ∈ [0,1] — higher values of either indicating larger deviation from regular operation. Given the small hand-labeled, unbalanced dataset, a decision rule on q0.9 alone was chosen to avoid overfitting: an observation is anomalous if q0.9 exceeds a ROC-curve-determined threshold (Figure 5); threshold 1.5 (red circle) gives the Section-4.3 results — the best (lowest) False Positive Rate (precision) without compromising the True Positive Rate (1 − sensitivity), crucial to miss as few true positives as possible. Experimental setting and results: lacking high-quality labels, unsupervised statistical-test anomaly detection ran on four hydraulic-signal time series sampled every 100 ms from June to October 2024 (4 GB total), generating potentially anomalous timestamps with rarity likelihood scores — statistical analysis validated outlier significance, and a domain expert manually reviewed a flagged subset over a continuous 5-month period (June–October 2024) under precision, recall and accuracy metrics. As the left confusion matrix (Table 1) shows: 499 timestamps considered, 45 flagged potentially anomalous, against six actual anomalous events reported by the engineering team — 83% of actual anomalies correctly flagged, with binary metrics of 91.8% accuracy, 99.78% specificity (True Negative Rate) and 83.3% precision — precision being most relevant since a false negative (undetected anomaly) costs far more than a false positive — and low recall (11.1%) from many false positives. Results show the unsupervised technique effectively identifies meaningful anomalies without labeled training data.

5. Case Study 2: Detecting Overtemperature Faults in Straddle Carriers

This section covers inverter over-temperature faults in straddle carriers — a critical fault occurring six times during 2024, directly affecting SC cooling systems with elevated inverter or engine temperatures that can escalate to complete engine or straddle-carrier shutdown with operational delays and safety risks. To develop and evaluate ML models, four of the six recorded incidents trained the models while two were reserved for testing — learning historical failure patterns and testing generalization to unseen events. Analysis uses real-world telemetry from 15 SCs at EUROGATE Container Terminal Limassol from straddle-carrier PLCs and sensors (hydraulic pressure, inclination, acceleration, vibration) — over 200 measurements every few operating minutes per SC: inverter, motor and engine temperatures, speed, torque, hydraulic pressure and error flags, plus on-site weather-station ambient temperatures. Data were manually labeled into two classes (normal/faulty): faulty from incident-involved SCs, normal from simultaneously operating SCs — six inverter over-temperature incidents from January to December 2024 (four training, two testing), with training size 8.1 MB (20,210 records: 13,860 normal, 6350 faulty) and testing 2.1 MB (5192 records: 3947 normal, 1245 faulty). Preprocessing filled missing values with per-metric means, one-hot-encoded categoricals, and standardized numericals (mean removal, unit-variance scaling).

Five popular ML models were evaluated for history-based PdM: (1) Artificial Neural Network (ANN — captures complex nonlinear relationships via neuron layers transforming inputs onward); (2) Decision Tree (DT — rule-based splitting by feature values into a decision tree); (3) Random Forest (RF — ensemble of trees on random data/feature subsets with combined outputs for accuracy and reduced overfitting); (4) Extreme Gradient Boosting (XGBoost — optimized sequential tree-ensemble gradient boosting with regularization for generalization); (5) Gaussian Naive Bayes (GNB — Bayes-theorem probabilistic classifier with conditionally independent Gaussian features). Feature importance came from per-model recommended approaches (permutation importance for ANN, impurity-based for DT/RF/XGBoost, log-coefficients for GNB), averaged per variable keeping non-zero features, plus SHAP (SHapley Additive exPlanations) analysis [39] on all models for per-feature prediction contributions; hyperparameters were grid-search-optimized under 5-fold cross-validation, with accuracy, precision, recall and F1-score evaluation.

This study uses full-year data (January–December 2024) to train and test inverter over-temperature fault ML models, developed in Python 3.12.2 with scikit-learn 1.5.2 on an AMD EPYC 7352 24-core (2.3 GHz), 64 GB RAM, Ubuntu 22.04.4 LTS machine. Table 2 lists the 21 selected features with average importance: top features are hoist (H) and drive (D) inverter and motor temperatures directly impacted by this failure (H1InverterTemperature leading at 0.612), with AmbientTemperature (0.198) helping distinguish weather-driven from fault-driven temperature rises; SHAP analysis visualized in the Figure-6 violin plot (x-axis SHAP value — positive pushing toward failure prediction, negative toward reduced risk — features ranked top-down, blue-to-red low-to-high values) shows heavy reliance on motor/inverter temperatures (H2InverterTemperature, H1MotorTemperature, H3InverterTemperature highest) with wide spreads indicating strong nonlinear target relationships critical to the decision boundary, moderate AmbientTemperature and EngineErrorSlot3 influence, and smaller torque/speed-reference effects. Hyperparameters were grid-search-tuned under 5-fold cross-validation (Table 3: ANN — batch size 64, 20 epochs, 0.0 dropout, 128 neurons, adam optimizer, 0.1 validation split; DT — ccp_alpha 0.001, gini, unlimited depth, min leaf 1, min split 2, random splitter; RF — entropy, unlimited depth, 10 max features, 100 estimators; XGBoost — 0.6 colsample_bytree, 0.1 learning rate, depth 3, 300 estimators, 0.6 subsample; GNB — 1.0 var_smoothing — spaces from literature and documentation [40–42]). Learning curves (Figure 7) diagnose convergence: ANN accuracy/loss (a, b) converge near zero within three epochs with sufficient capacity and no underfitting; XGBoost log-loss (c) declines smoothly with closely overlapping curves (strong generalization, minimal overfitting); DT and RF curves (d, e) show perfect training accuracy (overfitting) with validation fluctuating then plateauing ~90%; GNB (f) shows gradually declining training accuracy converging with validation ~97% (good bias–variance balance). Table 4 test results (Figure 8 confusion matrices): ANN leads on all metrics (98.73% accuracy, 97.5% precision, 99.0% recall, 98.0% F1) via nonlinear sensor-to-fault relationships and multilayered representations; RF and XGBoost tie (95.32% accuracy, 93.0% F1) — ANN misclassified no normal points as faulty beyond a few (2%) false positives, while RF/XGBoost had zero false positives — then GNB (94.78% accuracy, 92.5% F1) on its strong feature-independence assumption, and DT last (94.41% accuracy, 91.5% F1) from higher false negatives despite zero false positives. ANN effectively captures temporal dependencies, suiting it highly to predictive maintenance.

6. Discussion

While developing ML predictive models for container terminals, this study also offers transformative terminal/port benefits as a cornerstone of operational efficiency and asset reliability: leveraging IoT, AI/ML and advanced analytics enables early anomaly and failure detection, minimizing unexpected downtime and ensuring seamless operations — with generation, capture and processing in far-edge gateways collecting sensor and straddle-PLC data, then transmitting to the analysis component embedding prediction algorithms and ML models ahead of any failure. This extends critical-machinery lifespan (cranes, forklifts, conveyor belts), optimizes maintenance scheduling, cuts labor costs and prevents costly emergency repairs, and aids policymakers and terminal managers in maintenance scheduling, optimized spare-parts inventory, safety and risk management, and minimized operational delays.

In that sense, current preventive tools at EUROGATE Container Terminal Limassol schedule maintenance by working hours — a sub-optimal approach insufficient against unexpected CHE-component failures, causing undesired operational-hour idle times; with total 2023 downtime estimated at 900 h, the proposed ML models should significantly cut failure-driven CHE idle time, delivering business value in cost reduction, minimal unexpected-failure delays and downtime, sustainability, increased throughput and strong reputation (attracting more shipping lines). Model development faced challenges led by scarce real-world data — particularly faulty records from the Port of Limassol — with operators' data-sharing reluctance hampering cross-environment/port validation; the significant limitation of scarce faulty records and heavy manual labeling of historical telemetry against recorded incidents; sensor-data reliance challenging autonomous fault detection through data quality and consistency effects on accuracy; and likely retraining needs for different environments with unique operational factors despite promising terminal-specific results.

7. Conclusions

This study investigates diverse AI/ML for a predictive maintenance service paramount to container-terminal efficiency and safety, with data collected at the EUROGATE Container Terminal, Port of Limassol, Cyprus — both faulty and normal operational data over two years for training and validating the developed ML models. On one hand, a traditional unsupervised statistical model for straddle-carrier hydraulics was implemented for lack of labels, achieving 83.3% precision — the most relevant metric given the disproportionate cost of false negatives (undetected anomaly) over false positives. On the other hand, five models (ANN, DT, RF, XGBoost, GNB) were developed on real-world historical data to detect inverter over-temperature faults from fan failures, clogged filters and related issues: ANN achieves 98.7% anomaly-prediction accuracy, followed by RF and XGBoost at the same 95.3%, while DT and GNB perform worse at around 94%.

As future work, further testing must validate and enhance the hydraulic-anomaly-detection approach since the dataset contained few anomalies — expanding it or applying the model to similar scenarios will follow. Furthermore, we plan to extend ML-based predictive maintenance to other fault types such as engine, brake and other inverter issues.