The construction industry's reliance on fixed, calendar-based inspections is inefficient and often misses gradual infrastructure degradation, causing costly reactive repairs. This paper presents a six-step framework combining a contextual graph model, real-time IoT data, and a hybrid physics-machine learning pipeline for predictive inspection scheduling. A physics-based degradation model simulates corrosion, fatigue, and stress to create a synthetic dataset for training. A Random Forest model achieved accuracy in detecting at-risk states and a forecasting algorithm predicting inspection needs. Applied to a static plan, the framework detects new high-risk periods, validates or adjusts scheduled inspections, and defers low-risk ones. By updating maintenance plans with data-driven insights, it enables proactive, condition-based asset management that improves safety and reduces costs.
Document, Multimedia,
[1]. A compelling example is the slow water ingress in a concrete foundation, which can lead to rebar corrosion. This degradation process may proceed undetected for years, and by the time a scheduled inspection occurs, the structural damage may be severe and prohibitively expensive to repair. This gap highlights the urgent need to transition from rigid, calendar-based schedules to intelligent, predictive maintenance methodologies.
The convergence of Digital Twin, the Internet of Things (IoT), and machine learning (ML) offers a solution to address this challenge. Yet, as identified in a comprehensive literature review [2], the field still requires more practical, implementable frameworks and advanced algorithms to realize its full potential.
This paper directly answers that call. We propose a novel framework for predictive inspection scheduling designed to preempt costly failures by detecting the degradation. Our methodology leverages a hybrid approach: it integrates real-time data from embedded sensors with the static, contextual knowledge of an asset, such as material specifications and structural load models. A machine learning model trained on this combined data forecasts the future need for an inspection. Upon a positive prediction, our framework automatically generates a targeted inspection task, transforming the maintenance process from a reactive, time-based cycle into a proactive, data-driven, and condition-based workflow.
The goal of this research is to design and validate a framework that facilitates a shift from calendar-based inspection schedules to a predictive maintenance strategy for civil infrastructure.
To achieve this objective, this study will resolve the tasks: 1. Develop a physical degradation model of a specific BIM element by translating sensor data into physical metrics. 2. Design an algorithm for forecasting Inspection Timeline. 3. Define a method to train and evaluate machine learning classification model.
The scope of this initial study is focused on demonstrating the methodology's viability by developing and applying the physical degradation model to a single, asset type: a structural steel beam.
The research in [3] proposes a Digital Twin framework for railway turnouts, with a core contribution centered on Explainable AI (XAI). By modeling the system with a Bayesian Network structured by expert engineering knowledge, their approach delivers not just predictions but also probabilistic explanations for impending failures. While this 'white-box' model is highly valuable, its reliance on pre-defined causal structures can be challenging to scale across the diverse and often less-understood failure modes of heterogeneous building assets. Our framework complements this by proposing a hybrid approach that leverages both physics-based models and data-driven machine learning, offering adaptability where a single, explicit causal model may be insufficient.
The framework developed in [4] for wind turbines focuses on real-time prognostics and active control. Their Digital Twin not only forecasts failures but also tests control strategies to bring deviating parameters back within operational thresholds, creating a self-regulating system. Our perspective is that this self-regulation addresses the symptom, it cannot replace physical intervention to resolve the root cause of degradation. Therefore, we treat the frequency and magnitude of such control actions as a primary input signal for our inspection prediction model.
The work in [5] addresses the pre-deployment problem of optimal sensor placement. Their results prove that a well-designed sensor network, derived from physics-based simulations, maximizes the signal-to-noise ratio for damage detection. Our research is complementary and focuses on the subsequent operational phase. Assuming such a well-placed network is providing data, our framework answers the question of how to integrate this live sensor data with the rich contextual information from BIM models to drive a predictive workflow. While simulation-based design is essential, it is insufficient on its own, as it operates under idealized conditions and cannot account for the operational complexities our framework is designed to handle.
The study in [6] successfully validates a real-time infrastructure monitoring system for a "Noise Barrier Tunnel", integrating a high-resolution BIM model with live IoT data. The system proved effective through a dual-layer anomaly detection strategy combining real-time hardware alerts with offline analysis of subtle irregularities. A key limitation, stated by the authors, is that these advanced data-driven analytics are not part of the real-time system. Our research addresses this exact gap by automating and integrating the predictive analytics layer into the operational workflow. We achieve this by adding a machine learning classifier to forecast inspection needs in real-time, transforming the system from a reactive monitoring platform into a proactive one.
The research in [7] presents a comprehensive, six-layer data-driven Digital Twin framework designed to improve Facility Management for tunnels. A limitation of this otherwise holistic study lies in its treatment of the predictive analytics layer as a generic "data-driven" module, which may lack the specificity and interpretability required for the diverse and physically distinct assets found in a BIM model. Our research addresses this limitation by proposing a hybrid physical-ML pipeline. Instead of a single black-box model, we first use domain-specific engineering principles to compute explicit degradation metrics, which then serve as robust features for a data-driven classification model. This provides a more transparent, adaptable, and physically grounded approach to prediction.
This paper proposes six-process framework designed to enable predictive inspections for building assets by integrating heterogeneous data sources through a hybrid physical-machine learning pipeline (Figure 2).
At the core of our framework is a formal representation of the construction project as a heterogeneous property graph, denoted as G = (V, E). Here, V represents the set of all project object types O = {o1, …, on}, where n is the number of object types, and E represents the multi-relational connections between them. This contextual graph is managed within a native graph database to ensure high-performance traversal and querying, in line with modern data management principles.
Since each object o ∈ O is defined by different data types, the database for each must be selected as a "best-in-class" solution for that specific data type. For example, real-time time-series data from IoT sensors is best stored in a specialized database like InfluxDB [8].
Therefore, the framework includes processes to collect data about objects o ∈ O from their respective databases and storage systems. The process "Fetch Civil Building Construction Graph" retrieves the contextual graph G. The "Fetch Time-Series Data" process retrieves sensor measurements for specific BIM elements over a defined time window. In the "Combine Static and Dynamic Data" process, static metadata from the project graph G is merged with dynamic sensor measurements to form a dataset suitable for training a machine learning classification model. The "Create Physical Degradation Model of BIM Element" process defines functions to compute degradation metrics.
The "Design Binary Classification Model to Forecast Inspection Need" process involves designing and building a machine learning model to predict whether an inspection is required within a specified future time horizon. The "Integrate Predictions into the Graph" process dynamically updates the graph model G with inspection predictions by creating new nodes and relationships to represent forecasted risks and inspection schedules.
Corrosion reduction factor [9]:
(
The section modulus of a rectangular beam, adjusted for corrosion:
6⋅CF where b – beam width, m; h – beam height, m.
Maximum bending moment for a centrally loaded beam:
CF =1.0+ D ⋅ H 3650 100
⋅k c , S= bh2
, M = PL ,
4 where L – beam span length, m; P – weight of a single central load, kN.
Bending stress is calculated as the maximum bending moment divided by the section modulus: where α0 is an initial crack length, m.
Fatigue crack propagation over time in materials under cyclic loading, using Paris Law [12]:
where nc – number of cycles loadings the component per day, D – number of days the component has been in service.
Total crack length after N cycles starting from initial crack length α0 [13]:
N total=nc⋅D , α final=α 0+ ddαN ⋅N .
K I , final=σ √π α final .
A final stress Intensity Factor at the final crack lengthfinal:
The binary rule to identify when inspection is needed consists of 3 conditions, which are joined by logical “or” operations: a final crack lengthαfinal exceeds 15 mm; safety factor SF is less than 2.0 indicating that the material is closed to failure under the given conditions; the final stress intensity factor KI,final exceeds 90% of the material's fracture toughness 90% of the material's fracture toughness KIC [14]:
Inspection={1 ,α final>0.015∨SV <2.0∨ K I , final≥0.9 K IC .
0 , otherwise
(
To identify when a beam requires inspection, this paper proposes an algorithm that integrates
physical degradation modeling with predictive machine learning techniques. For the defined
forecast period the algorithm identifies the number of periods denoted asnsteps, and for each period
calculates the future date, denoted as tfuture. It updates the beam's parameters using a
time-dependent degradation model, denoted as Xfuture, using function f(∙) which performs
calculations according to equations (
Algorithm 1. Algorithm for forecasting Beam Inspection Timeline:
Random Forest Classifier is trained with 200 trees, a learning rate is set to 0.1 and a maximum tree
depth limited to 15. Additionally, 10 samples are required to form a node. Gini Impurity is used as
the splitting criteria:
k
G (t )=1−∑ ( pi)2 , (
i=1 where pi is the proportion of samples belonging to class i at node t.
The XGBoost classifier is trained with 300 trees, a learning rate of 0.1, and a maximum tree depth limited to 7. The objective function to be minimized is Binary Log Loss:
1 N L=− ∑ [ yi log ( pi)+(1− yi) log (1− pi)] .
N i=1
The hyperparameters for Random Forest Classifier and XGBoost were selected based on established best practices and findings from preliminary experiments. A standard scaling method is used to normalize feature values with the same scale.
The performance of each model is evaluated using the 5-fold cross-validation method and includes the following metrics:
The proportion of correctly classified samples among all samples:
The harmonic mean of precision and recall, balancing the tradeoff between false positives and false negatives:
Accuracy =
TP +TN TP + FP +TN + FN
.
F1= 2⋅Precision⋅Recall .
Precision+ Recall
Additionally, a comprehensive visualization of where errors occur in "No Inspection Required" vs. "Inspection Required" predictions is provided through a confusion matrix.
To enable predictive inspections for construction project assets, we propose an operational framework that integrates data from multiple specialized databases.
A comprehensive civil construction project model, represented as a graph G = (V, E), is stored in a Neo4j database (see Figure 3). The process of identifying data for a specific asset—in the scope of this paper, a construction beam—begins by querying this graph. A Cypher query, such as MATCH (s:Sensor)-[:ATTACHED_TO]->(e:Element {elementID: $target_id}) RETURN s.deviceID, is used to traverse the graph from a target :Element node to its associated :Sensor nodes, retrieving the unique deviceID for each sensor.
Static parameters required for the physical degradation model would then be retrieved from their respective sources. For this study, we assume these parameters are stored as properties on the :Element nodes within the Neo4j graph. For each structural beam asset (elementID) the parameters include: Geometric Properties: beam.width (b), beam.height (h), beam.spanLength (L). Material Properties: beam.yieldStrength (σy), beam.fractureToughness (KIC), beam.paris_C, beam.paris_m. Operational Parameters: beam.initialCrackLength (α0), beam.cyclesPerDay (nc). Provenance Data: beam.installationDate, from which "Days in Service" (D) is calculated.
Using the retrieved deviceIDs, queries would be sent to a time-series database (InfluxDB) to
fetch dynamic measurements such as Humidity (H), Temperature (T), and Load (P) over a given
time window. The combination of this static and dynamic data forms the complete feature vector
required for real-time prediction.
(
This study validates the proposed predictive models using a synthetically generated dataset. This approach allows for the creation of a balanced and comprehensive dataset to rigorously train and test the machine learning algorithms.
A dataset of 2000 observations was generated. A subset of parameters (Table 1) were sampled
from uniform distributions and the remaining parameters were derived according to Equations
(
To address the class imbalance inherent in generated datasets (where non-failure cases far outnumber failure cases), the SMOTE (Synthetic Minority Over-sampling Technique) was applied to the training data. The default sampling_strategy='auto' was used, which oversamples the minority class until it has the same number of instances as the majority class, resulting in a balanced 1:1 class ratio. The number of nearest neighbors used to generate synthetic samples, k_neighbors, was also kept at its default value of 5. This balancing technique helps prevent the machine learning models from being biased towards the majority class and improves their ability to detect the rare "inspection required" events [15].
The balanced dataset was then used to train two machine learning classifier algorithms: a
Random Forest Classifier (RF) and an Extreme Gradient Boosting Classifier (XGBoost), with the
goal of predicting the binary outcome defined in Equation (
The models were developed and trained using the Python programming language in the following software environment: PyTorch (version 2.7.0 + cpu), NumPy (version 2.2.4), Pandas (version 2.2.3), Scikit-learn (version 1.6.1), Matplotlib (version 3.10.1). Construction graph model was developed and stored in Neo4j desktop (version 2.0.3). Time series data loaded from InfluxDb.
Figure 5 visualizes physical relationships in the dataset. Red points denote “Need Inspection” and green “No Inspection.” Figure 5(a) shows applied load (P) vs. bending stress (σ): loads < 30 kN are mostly safe; > 60 kN correlate with inspection need. Figure 5(b) shows crack growth rate vs. daily cycles – low cycles (< 5,000/day) show low growth, while > 15,000/day strongly increase da/dN. Figure 5(c) links corrosion to humidity—low humidity (10–40 %) causes minor corrosion; high humidity (70–95 %) accelerates it.
Figure 6 confirms the dataset’s suitability for ML training. The XGBoost classifier misclassified two “Needs Inspection” cases (recall = 0.9901), while the Random Forest (RF) achieved perfect accuracy (Accuracy = 1.0, F1 = 1.0). Hence, Algorithm 1 was executed using the RF model.
Table 2 records less Accuracy of XGBoost classifier caused by two missed cased as was illustrated on Figure 6. Therefore, Algorithm 1 is executed with the trained Random Forest model.
The forecasted with Algorithm1 inspections requirements are illustrated on Figure 7. Figure 7 (a) depicts the probability of the beam requiring an inspection over the forecasted time horizon. The red dashed horizontal line indicates the decision threshold (0.5); when the probability of requiring inspection exceeds this threshold, an inspection is recommended. There are three peaks exceeding the threshold, each lasting for a short period before the probability drops: Early in the timeline (near t = 0). At approximately Year 2 and Year 4. On Figure 7 (b) the plot shows the change in the safety factor (SF) over time, with the red dashed line denoting the minimum safety factor threshold (SF = 2.0). When t = 2 years, 4 years SF approaches dips below the critical threshold coinciding with the inspected probability of the beam requiring an inspection peaks in Fig 7. (a) chart. Crack propagation has a spike near the start of the graph (Figure 7 (c)). The crack length grows beyond 10mm, crossing into the critical failure zone early in the forecast. After this initial crack growth, no further crack propagation is observed until Year 4, where another spike is visible. This second spike also crosses into the Critical Zone. Most of the time horizon sees stable or negligible crack growth—suggesting limited degradation during those intervals.
The results of the integration of the predictions for planned inspection are illustrated on Figure 8 (b) and include the following: 1. The static plan's first inspection is a routine visual check on March 2025. However, our predictive model (Figure 7.a) shows a high probability (>0.5) of inspection need starting earlier, driven by a significant initial crack propagation event (Figure 7.c). The framework automatically generates a task for a new, unscheduled "Detailed Inspection" in March 2025, specifically focused on crack assessment. 2. The static plan calls for a 'Detailed Inspection' on November 2026. Our model confirms the need for an inspection around this time, with the probability peaking in late 2026 (approximately Year 2). However, our framework provides additional insight: the driver for this inspection is the Safety Factor (SF) below the critical threshold of 2.0 (Figure 7.b). The scheduled 'Detailed Inspection' for November 2026 is validated and confirmed as necessary. The framework will update the description, specifying to focus on the factors affecting the safety factor. 3. The static plan includes a 'Routine Visual' inspection on June 2027. Our predictive model shows a very low probability of inspection need throughout 2027. The safety factor is above the threshold, and crack growth during this period. The framework updates the inspection's description recommending that the inspection scheduled for June 2027 is unnecessary and can be safely deferred. 4. The static plan schedules a 'Detailed Inspection' for November 2028. Our model concurs, showing a sharp spike in inspection probability around Year 4 (~ October 2028). The underlying drivers are identified as another significant crack propagation event and a corresponding dip in the safety factor. The framework updates the inspection's description by setting a high-priority.
This research developed and validated a novel framework for predictive inspection scheduling in construction projects. The core contribution is a hybrid physics-machine learning (ML) methodology, where a physics-based degradation model of a structural beam generated a highfidelity synthetic dataset for training ML classifiers.
Applying the forecasting algorithm to a real-world, static inspection plan demonstrated the framework’s practical value. It dynamically identified a new high-risk period requiring an added inspection, validated necessary scheduled inspections with causal insights, and flagged a low-risk period where an inspection could be safely deferred-optimizing resources while enhancing safety and efficiency.
The main limitation is that the ML models were trained on synthetic data derived from physical models, which cannot fully represent real-world noise, sensor errors, or operational variability. The study was limited to one asset type (a structural beam) and specific degradation modes (corrosion and fatigue), excluding others such as fire or impact damage.
Future work will expand the framework’s scope by developing a library of degradation models for various asset types and incorporating explainable AI (XAI) methods, such as SHAP, to increase transparency and user trust in predictive decisions.