Smart Sensing for Safer Bridges: From Sensor Signals to AI-Driven Anomaly Detection

I. INTRODUCTION

• Bridge monitoring must distinguish normal measurements from anomalies that may be associated with structural damage, unusual loading, environmental influences, or sensor faults.The introduction connects reliable monitoring with public safety and reducing the likelihood of severe failures or accidents.Source: §I · p. 1
• An example anomaly is a sensor observation with a substantially larger peak than surrounding normal measurements.The paper also describes bridge anomalies as deviations that may be associated with unusual loading, environmental influences, or sensor-related faults.Source: §I · p. 1
• Statistical approaches often rely on prior assumptions about data characteristics, which may reduce reliability when real-world sensor distributions are unknown or complex.The paper cites statistical approaches, multivariate models, and mean-variance methods.Source: §I · p. 1
• The study investigates signal processing and machine learning for anomaly detection in measurements collected by an iBridge device installed on a bridge in Norway.Early identification of abnormal patterns may support timely maintenance and help reduce structural risks.Source: §I · p. 2

A. Signal Processing

• Signal processing uses find_peaks to identify significant local maxima and minima, with height, prominence, and minimum-distance parameters distinguishing deviations from normal fluctuations.The method also searches for negative valleys by applying the algorithm to the inverted signal [5].Source: A. Signal Processing · p. 2

B. Isolation Forest

• Isolation Forest identifies unusual observations by combining randomly constructed binary trees that recursively partition sensor observations.Anomalies generally require fewer partitioning steps and appear closer to an isolation tree’s root, while normal observations tend to lie deeper.Source: B. Isolation Forest · p. 2

III. PROPOSED MODEL

• The framework cleans and preprocesses sensor signals, extracts features, applies signal processing and Isolation Forest, then analyzes and visualizes detected anomalies.The analysis includes anomaly number, timing, magnitude, duration, detection time, and processing rate; visualizations are intended to support interpretation and monitoring decisions.Source: §III · p. 2

A. iBridge Sensor Data Acquisition Device

• The iBridge is a compact monitoring unit that integrates multiple sensor interfaces and communicates measurements to a cloud platform through 4G connectivity.It supports battery-powered operation and is primarily designed for strain-gauge applications.Source: A. iBridge Sensor Data Acquisition… · p. 2

B. Bridge Sensor Data

• The bridge dataset covers June 10–20, 2026, and records downward strain in mV at three locations designated ADC1, ADC2, and ADC3.The iBridge device was installed on the side of the bridge.Source: B. Bridge Sensor Data · p. 3
• The 11-day dataset contains 3,774,011 numerical samples collected at 5 Hz, or one measurement every 0.2 seconds.Source: B. Bridge Sensor Data · p. 3

V. RESULTS AND DISCUSSIONS

• The section introduces the experimental setup and a detailed analysis of the results.Source: §V · p. 3

A. Experimental Setup

• Experiments use Python 3.13.6, Keras with TensorFlow 2.20.0, and a MacBook with an Apple M4 processor and 16 GB of RAM.NumPy 2.3.2, Pandas 2.3.1, and scikit-learn 1.7.2 are used for preprocessing.Source: A. Experimental Setup · p. 3

B. Data Preprocessing

• 285 ADC1 and ADC2 entries and 290 ADC3 entries are missing, and the observations are estimated using time-based linear interpolation.The final bridge signals are shown in Fig. 6, Fig. 7, and Fig. 8.Source: B. Data Preprocessing · pp. 3–4
• The final ADC1, ADC2, and ADC3 bridge signals are shown in Fig. 6, Fig. 7, and Fig. 8, respectively.Source: B. Data Preprocessing · p. 4, Figure 6 · p. 4, Figure 7 · p. 4

• Ground-truth anomaly labels are unavailable, so precision, recall, and F1-score cannot be computed.Source: C. Performance Evaluation · p. 4
• Evaluation considers anomaly count, anomaly detection time in seconds, processing rate in samples/second, and anomaly rate in percent.Detected anomalies are also visualized in corresponding plots.Source: C. Performance Evaluation · p. 4
• Detection time is the elapsed time from the start to the finish of detection, while processing rate is input samples per detection-time unit.Anomaly rate is the percentage of input samples detected as anomalous.Source: C. Performance Evaluation · p. 4

D. Proposed Model Analysis

• Grid search determines suitable parameter configurations for each method, with the resulting settings summarized in Table I.The bridge sensor data is preprocessed before model development.Source: D. Proposed Model Analysis · p. 4

1) Signal Processing Analysis:

• 0.0546–0.1926 seconds and 19.59–69.09 million samples/second are the signal-processing detection-time range and processing-rate range, respectively.ADC3 has the shortest detection time and highest processing rate; the method detects 349 ADC1, 14 ADC2, and 1 ADC3 anomalies.Source: 1) Signal Processing Analysis · p. 4

2) Isolation Forest Analysis:

• Isolation Forest identifies 3,579, 29, and 11 anomalies in ADC1, ADC2, and ADC3, respectively.The detections are illustrated in the ADC1, ADC2, and ADC3 plots.Source: 2) Isolation Forest Analysis · pp. 4–5, Figure 12 · p. 5, Figure 13 · p. 5
• Detection time ranges from 14.52 to 31.79 seconds, while processing rate ranges from 118,722 to 259,891 samples/second.ADC1 has the shortest detection time and highest processing rate; ADC3 has the longest detection time.Source: 2) Isolation Forest Analysis · pp. 4–5
• Isolation Forest is computationally more demanding than signal processing but provides a data-driven approach to identifying anomalous observations.The paper describes this as a computational trade-off between the methods.Source: 2) Isolation Forest Analysis · p. 5

E. Temporal Agreement Analysis

• For ADC1, 100 detections coincide, representing 28.65% of signal-processing detections and 2.79% of Isolation Forest detections.ADC2 has two coincident anomalies, with agreement ratios of 14.29% and 6.90%, respectively; ADC3's one signal-processing detection also appears in Isolation Forest, giving 100% agreement relative to signal processing. This analysis measures temporal agreement, not detection accuracy.Source: E. Temporal Agreement Analysis · p. 5

F. Controlled Anomaly-Injection Analysis

• For 18 injected anomalies per signal, signal processing detects 0, 6, and 1 in ADC1, ADC2, and ADC3, while Isolation Forest detects 0, 18, and 13.The corresponding recall values are 0.00%, 33.33%, and 5.56% for signal processing, and 0.00%, 100.00%, and 72.22% for Isolation Forest; sensitivity varies across signals.Source: F. Controlled Anomaly-Injection Analysis · p. 6, Table IV · p. 6
• The injected perturbations are synthetic and may not represent actual bridge anomalies, so this is a controlled sensitivity analysis rather than evaluation on real bridge conditions.The experiment injects point spikes, short events, and level shifts at moderate, strong, and extreme strengths.Source: F. Controlled Anomaly-Injection Analysis · pp. 5–6

VI. CONCLUSIONS AND FUTURE WORK

• Signal processing offers a simple, computationally efficient detection mechanism, while Isolation Forest identifies unusual observations from patterns in bridge sensor data.The paper characterizes the approaches as complementary.Source: §VI · p. 6
• The methods identify some anomalies at approximately the same time, while controlled-injection sensitivity varies across sensor signals.Source: §VI · p. 6
• Future work will evaluate multiple bridges and additional sensor modalities and develop adaptive real-time detection for continuous bridge monitoring.Source: §VI · p. 6