Detecting the Unexpected: AI-Driven Anomaly Detection in Smart Bridge Monitoring

Detecting the Une xpected: AI-Dri v en Anomaly Detection in Smart Bridge Monitoring Rahul Jaiswal ∗ , Joakim Hellum and Halvor Heiber g Smart Sensor Systems AS Oslo, Norway { rahul.jaiswal, joakim.hellum, halv or .heiber g } @smartsensorsystems.no Abstract —Bridges are critical components of national infras- tructure and smart cities. Therefor e, smart bridge monitoring is essential for ensuring public safety and pre venting catastrophic failures or accidents. T raditional bridge monitoring methods r ely heavily on human visual inspections, which are time-consuming and prone to subjecti vity and error . This paper proposes an artificial intelligence (AI)-driven anomaly detection approach f or smart bridge monitoring. Specifically , a simple machine lear ning (ML) model is developed using real-time sensor data collected by the iBridge sensor devices installed on a bridge in Norway . The proposed model is evaluated against different ML models. Experimental results demonstrate that the density-based spatial clustering of applications with noise (DBSCAN)-based model outperforms in accurately detecting the anomalous events (bridge accident). These findings indicate that the proposed model is well- suited for smart bridge monitoring and can enhance public safety by enabling the timely detection of unfor eseen incidents. Index T erms —Anomaly detection, bridge monitoring, iBridge sensor device, machine learning, and sensor data. I . I N T R O D U C T I O N Bridges are the critical component of smart cities, forming the backbone of modern transportation systems. W ith hundreds of thousands of bridges deployed w orldwide, for instance, ov er 18,000 in Norway [1], ensuring their safety and reliability is of paramount importance. T raditionally , the assessment of bridge conditions has relied heavily on human visual inspections. Howe ver , such inspections are often time-consuming and prone to error , as accurately ev aluating the extent of structural deterioration, particularly in concrete/wooden bridges, is chal- lenging due to the limitations of human visual observations. T o overcome these challenges, intelligent sensor-based, data- driv en machine-learning bridge monitoring techniques hav e emerged. It enables timely assessment of structural integrity , enhancing pedestrian and public safety , and mitigating the risk of catastrophic failures or accidents. Ef fectiv e bridge monitoring requires designing a smart model that can capture variations between normal and anomalous sensor data. Anomaly detection is an important application of data mining [2] and is commonly referred to as outlier detection. In the context of bridge sensor data, anomalies or outliers correspond to the measurements that deviate from the normal signal or data patterns. These deviations may arise from ∗ Corresponding author Anomaly Normal Time V alue Fig. 1: A toy example of anomaly detection in sensor data. structural damage, abnormal loading conditions, en vironmen- tal effects, or sensor malfunctions. The objectiv e of anomaly detection is therefore to identify such inconsistent or irregular sensor observations that dif fer significantly from typical bridge behaviour . A toy example illustrating anomaly detection in sensor data is presented in Fig. 1. As shown, the anomalous data point exhibits a significantly higher peak compared to the normal sensor measurements. Sev eral approaches have been proposed in the literature for detecting anomalies in sensor data. Statistical-based meth- ods [3], such as multiv ariate models [4] and mean-v ariance- based techniques [5], typically learn discriminativ e features from historical monitoring data representing normal structural behaviour . A data point is then classified as anomalous when it deviates significantly from these learned features. Ho we ver , the effecti veness of statistical methods relies on the prior empirical knowledge, including assumptions about data distributions and model parameters used for anomaly identification. In practice, the true distrib ution of real-world sensor data is often unknown, which can limit the accuracy of anomaly detection using statistical methods. Recently , artificial intelligence (AI)-driv en, particularly ma- chine learning (ML)-based, approaches ha ve gained significant attention for anomaly detection [6], as they automatically extract discriminative features from raw bridge sensor data to train models. Neural network-based methods are also ef fectiv e due to their strong representation and universal approximation capabilities [7], [8]. For instance, a K-nearest neighbour-based anomaly detection framework is proposed in [9] using raw vi- bration measurements collected from a bridge. A con volutional 979-8-3315-3672-5/26/$31.00 ©2026 IEEE neural network (CNN) is used in [10] to detect structural dam- ages in bridge systems. An autoencoder is employed in [11] for anomaly detection in bridge structural health monitoring. The re view of existing literature highlights that anomaly detection in bridge systems is critical for maintaining the structural performance of the bridge, ensuring pedestrian and public safety , and mitigating the risk of catastrophic failures or accidents. Motiv ated by these findings, this work focuses on machine learning-based anomaly detection for smart bridge monitoring. Specifically , we de velop an ef ficient ML-based anomaly detection framework to identify a bridge accident captured by iBridge sensor de vices (see Section IV -A) installed on a bridge in Norway . The proposed model can enable bridge monitoring authorities to promptly detect anomalous ev ents and take timely correctiv e actions, thereby reducing potential risks and prev enting further adverse consequences. The remainder of this paper is organized as follo ws. Sec- tion II describes the machine learning methods employed for anomaly detection. Section III introduces the proposed anomaly detection model. Section IV describes the experi- mental dataset. Section V discusses the results, and Section VI concludes the paper and outlines directions for future work. I I . B AC K G RO U N D This section describes various machine learning techniques used in our experimental study for detecting anomalies in the bridge monitoring system. A. Isolation F or est The Isolation Forest (IF) algorithm detects anomalies using an ensemble of binary trees [12]. In bridge monitoring applica- tions, sensor data are isolated by randomly selecting a feature and an associated split value at each node. The algorithm con- structs a forest, referred to as an iForest, comprising multiple isolation trees (iT rees) [13]. During training, the sensor data are recursi vely partitioned until the iT rees ef fectiv ely separate normal data points from anomalous data points. T ypically , anomalous data points are isolated near the root nodes of the iT rees, whereas normal operating conditions (points) require deeper trav ersals and are located farther from the root nodes. A simplified structure of an iTree [6] is illustrated in Fig. 2. Here, external nodes have no children, whereas internal nodes contain two child nodes. In this representation, anomalies are typically associated with external nodes. B. Autoencoder An autoencoder (AE) [14] is a neural network designed to learn a compact (latent) representation of input data while minimizing the error in reconstructing the original signal. In bridge monitoring applications, autoencoders are typically trained using sensor measurement data that capture normal signal behaviour . Consequently , when anomalous sensor mea- surements, such as those caused by structural damage, a vehicular accident, an unusual load on the bridge, or sen- sor malfunctions, are encountered, the reconstruction error Root Node Edge External Node Internal Node Fig. 2: A simple illustration of an iT ree structure. Input Data Latent Space Reconstructed Data Encoder Bottleneck Decoder Fig. 3: A simple illustration of an autoencoder . increases significantly . This reconstruction error can therefore be used as an effecti ve indicator for anomaly detection. An autoencoder consists of an encoder and a decoder . The encoder compresses input data into a low-dimensional latent representation (also called a bottleneck), and the decoder reconstructs the original input from that latent representation, as sho wn in Fig. 3. The autoencoder model is trained by minimizing the reconstruction error between the input and its reconstruction. Next, we briefly present the mathematical formulation of an autoencoder . Let a sensor device with d channels generate measurements ov er time. At time t , the sensor measurements are giv en as: x t = x (1) t , x (2) t , . . . , x ( d ) t (1) where, x ( i ) t denotes the measurements of the i-th sensor de vice. The encoder compresses the input into a low-dimensional latent representation as: z = f θ ( x ) = σ ( W e x + b e ) (2) where, W e ∈ R k × D is encoder weight matrix, b e ∈ R k is the encoder bias vector , k < D is the dimension of the latent space which controls the compression strength, D is the dimension of the input vector , z ∈ R k is the latent vector , and σ ( · ) is a nonlinear acti v ation function (e.g., ReLU, tanh). Note that R denotes the set of real numbers. The decoder reconstructs the original input from the latent representation as: ˆ x = g ϕ ( z ) = σ ( W d z + b d ) (3) where, W d ∈ R D × k and b d ∈ R D . Next, the reconstruction error measures how well the model reproduces sensor signal patterns, and it is giv en by the mean squared error (MSE) [15] as: L ( x , ˆ x ) = 1 D D X i =1 ( x i − ˆ x i ) 2 (4) T o determine whether a given sample is normal or anoma- lous, an anomaly score based on reconstruction error (MSE) is computed and compared against a predefined threshold. If the anomaly score is below the threshold, the sample is classified as normal; otherwise, it is classified as anomalous. Note that normal samples exhibit low reconstruction error , whereas anomalous samples result in high reconstruction error . C. Density-based Spatial Clustering of Application with Noise The Density-based Spatial Clustering of Applications with Noise (DBSCAN) [16] is a density-based clustering technique that identifies data patterns by analyzing the local concentra- tion of samples. In bridge monitoring applications, DBSCAN assumes that sensor measurements corresponding to normal data points form dense regions in the feature space, whereas abnormal ev ents such as accidents or unusual loads appear as isolated points with low local density . The method relies on two key parameters. The neigh- bourhood radius eps , which defines the maximum distance between two sensor measurements for them to be considered neighbours and form a cluster . The minimum number of points minpts , which specifies the minimum number of neighbouring measurements required within the eps radius for a data point to be considered as part of the cluster [16]. During training, DBSCAN categorizes samples into three groups. The samples having a sufficient number of neighbours within the eps distance are treated as dense-region points and represent normal operating conditions of the bridge. The samples that lie close to these dense regions b ut do not independently satisfy the density requirement are considered adv an t a g es an d dis adv a n ta g e s , an d th e m o tiv at ion o f th e s e tech niqu es w e re dis c u s s e d in a com p arati v e approach . T a n g et al. [3] rev i e w ed t h e diff erent s o lu tio n m e th ods o f an o m al y det ect i o n probl em s , i . e., den s i t y -bas e d, conn ect i v i t y - based m et h ods , an d th e th eoretical s t ru ctu r e of th e a n o m al y detecti o n probl em s . Pet rov sk i y [4] di s c us s e d t h e a n o m al y det ect i on alg o rithm s , bas i c approach es , th e adv a n t a g es a n d di s a dv an t a g e s o f t h e s e approa ch es , an d propos ed a n e w fu z z y - bas ed an o m al y detectio n alg o r ithm . T h e an o m al y detection tec h n i qu es ca n be clas s i f i ed as s t atis t i cal approach es an d dis t an ce - b as ed approach es . Statis t i cal approach es ai m to dev elop a s t atis tical m odel of th e d a ta an d id en tify d a ta t h at d o n o t f it in to th e m o d e l. In dis t an ce - b as ed approach es , t h e dis t a n ces bet w ee n data are co n s id ered in d e tectio n o f a n o m alie s: th e d a ta at a d i sta n ce g r eater th a n a pre- def i n ed dis t an ce is called a n an o m al y . T h is stu d y f o c u ses o n d e tecti n g o u tliers u s i n g d i sta n ce- b a s e d approach es . D B S C A N , w h i c h w a s dev e l o ped by Est e r et al . [5], i s on e of th e di s t an ce - b as ed approach es , w h ic h h a s been w i de l y u s ed f o r s o lv i n g a n o m al y det ection proble m s . Es ter et al. [5] s h o w e d th a t D B SCA N i s m u ch m o r e p o w e r f u l a l g o r i th m th a n t h e ot h e r al g o ri t h m s o n m a ss an d dens e dat a s e t s f o r fi n d i n g an o m alie s . I n t he st ud y , t he D B S C A N a l go r i t h m i s use d fo r d e tectio n o f an o m alie s i n m o n t h l y te m p erat u r e d a ta a n d t h is m et h od is co m p ared w i th a s t ati s tical an o m al y detecti o n m e t h od. III. DB SC A N A LG O R I T H M DB SC A N i s a d e n s it y - b a sed sp atial cl u s teri n g alg o rithm t h at can al s o def i n e an o m alie s in t h e data s eries . It requ i res t w o us er- d e f i n ed param e t e rs , w h i c h are n e i g h borh o od dis t an ce ep silo n ( eps ) an d m i n i m u m num ber of poi n t s m i npt s . Fo r a gi ve n p o i nt , t h e p o i nt s i n t h e eps dis t an ce are called n e i g h bors of t h a t poi n t . If t h e n u m b er of n e i g h bori n g poi n t s of a poi n t i s m o re t h a n m i npt s , t h i s g rou p o f poi n t s i s cal l e d a clu s ter. D B S C A N l a bel s t h e dat a poi n t s as core poi n t s , border p o i n t s, an d o u tlier (an o m a lo u s ) p o i n t s. Co re p o i n t s are th o s e th at h a v e at leas t m i npt s num ber of poi n t s i n t h e eps dis t a n ce. Borde r poi n t s ca n be de f i n e d as poi n t s t h at are n o t core poi n t s , b u t are th e n e ig h b o r s o f co re p o i n t s. Ou tlier p o i n t s are th o s e t h at are n e i t h e r core poi n t s n o r borde r poi n t s . In DB SC A N , th e cl u s teri n g a pproach is dif f eren t f r o m t y pical cl u s teri n g approach es . DB SC A N can de f i n e ou tli er (an o m a lo u s ) p o i n t s th at d o n o t f it to a n y c l u s ters. In Fi g . 2 , th e g rou ps repres en t resu l t s o f a cl u s t e ri ng approach t h a t t a kes di s t an ce t h re s h ol d eps as a clu s ter m e tric. I n ad d itio n to di s t an ce m e t ri c , D B S C A N requ i res m i n i m u m num ber of poi n t s m i npt s i n s i de a g r oup t o l a bel i t as a cl u s t e r. F o r ex a m ple, i f t h e m i npt s v al u e i s s elected as 3, gr oup 1, gr ou p 2, gr oup 4 , an d gr oup 6 w i ll be m a r k ed as cl u s ters u s i n g DBSCA N . I n co n t ras t , gr oup 3 an d gr oup 5 w i l l be def i n e d as o u tliers si n ce th e y d o n o t co n t ain s u ff ic ien t n u m b er o f p o i n t s to f o r m a clu s ter. Si m ilarl y , if th e m i npt s v alu e i s s elec t ed as 5, gr oup 3, gr oup 5, an d gr ou p 6 w ill b e ass i g n ed as o u t l i e rs u s i n g DB SC AN. Fi g u r e 2 . A sam p l e dat ase t f o r d i st an ce - b ase d a n o m al y de te ctio n ap pr o ac h T h e ps eu do- code of D B S C A N al g o ri t h m i s g i v e n i n A l g o rith m 1 . T h e i n p u t s o f th e alg o rith m are d a ta set a n d u s er- def i n e d eps an d m i npt s para m e ter v a lu e s . Algo ri t h m 1. T h e p s eud o c o d e of DBSC A N a l go ri t h m In p u t s: D: the da ta se t Eps : t h e ne ig hb or h o o d dis t a n c e Minp ts : t h e m i nim u m num be r of poi nts Out put : Disc ov e r e d outlie rs a n d c luste rs Varia b les: m , n: row a nd c o lum n v a lue s of D m a trix , re spe c tiv e l y Di st : d i st an ce vect o r indic e s : i ndic e s tha t dis t a n c e of points is low e r tha n E p s c l a s s _no: i ndic a t e s the c l us te r s – de f a ult 1 Alg o rith m : 1. im port the da ta -se t i n to D 2. fo r i = 1 to m //r ow c ounte r 3. Di st = d i st an ce(i , D) 4. ne ig h bor s = f i nd( D i s t = < Eps ) 5. n e i ghb o r _ c ou n t = c o un t ( n e i ghbo r s ) 6. c o r e _ n e i g = c h e c k _c or e _ne ig hb or ( n e i g h bor s ) 7. if ( n ei gh bo r _ c o un t >=m i n p t s) 8. c l a s s ( i) = c l a s s _no / / c l us te r e d po int 9. w h ile ( m or e poi nts ne a r i) 10. c l a s s ( p o in t) = c l a s s _no 11. e nd w hi l e 12. cl ass_ no += 1 13. el se i f ( ne ig h bor _c ou nt< m inpts & c o r e _ne ig = = T r u e ) 14. c l a s s ( i) = 0 / / b o rde r p o int 15. el se if ( n e i g hbor _c o unt< m inpts ) 16. c l a ss(i) = -1 // ou tlie r poi nt 17. e nd i f 18. en d for 19. retu r n cl a s s 92 Fig. 4: A simple illustration of the DBSCAN method. AI-driven Anomaly Detection Model Detecting Anomaly Sensor Data Anomaly Detected Fig. 5: The proposed anomaly detection model. boundary points. All remaining samples that do not belong to any dense region are treated as outliers or anomalies. A simplified illustration of the DBSCAN method [6] is presented in Fig. 4. As shown, six distinct groups are formed based on the clustering process, where the neighbourhood radius eps defines the cluster boundaries. A group is identified as a valid cluster only if it contains at least minpts data points. For instance, when minpts is set to 3, Groups 1, 2, 4, and 6 satisfy the clustering criteria and are classified as clusters, while Groups 3 and 5 are treated as outliers. In contrast, increasing minpts to 5 results in Groups 3, 5, and 6 being labelled as outliers. I I I . P R O P O S E D A N O M A L Y D E T E C T I O N M O D E L The primary goal of the proposed anomaly detection model is to accurately identify abnormal ev ents for effecti ve bridge monitoring. T o achieve this goal, the proposed system ar- chitecture integrates multiple cooperating components, as il- lustrated in Fig. 5. The architecture consists of three main stages. First, sensor data collected from the iBridge sensor device (see Section IV -A) are preprocessed and prepared as input to the anomaly detection model. Second, the AI-based anomaly detection model is trained using the processed data, during which it learns salient feature patterns. Finally , the trained model detects anomalous ev ents and outputs anomaly predictions, which can be used to identify potential failures or accident-related incidents in the bridge monitoring system. I V . E X P E R I M E N T A L D A TA S E T This section describes the sensor device used for data acquisition and the real-time bridge monitoring dataset. A. Sensor Data Acquisition Device: iBridge The bridge monitoring data are collected using a sensor- based device, iBridge, as illustrated in Fig. 6. The iBridge [17] is a compact and smart de vice equipped with a multifunctional sensor-based system that transmits sensor data to a cloud platform via 4G communication. It supports battery operation and can be interfaced with a wide range of sensors and transmitters. The iBridge is specifically designed for use with strain gauges and can measure v arious structural responses, Fig. 6: A simple illustration of the iBridge sensor device. Fig. 7: Illustration of installed iBridge devices on the bridge. such as acceleration, tensile and compressi ve stresses, crack- ing, displacement of mass and collapse, strain on the metal beams, weight, pressure, vibration, angular variation, torque, temperature, liquid lev el, and moisture. The iBridge device incorporates an integrated accelerometer and temperature sensor , functioning as a multifunctional data logger . It is easy to install on v arious materials and surfaces, can be relocated and reused, and supports adaptable battery configurations based on deployment needs. The direct sensor connectivity enables cable-free installation and no power con- nection. The device is connected to an intelligent cloud-based platform for configuration and data storage, providing contin- uous 24/7 access to monitoring data or seamless integration with the enterprise resource planning (ERP) system. The iBridge device is suitable for a wide range of monitor- ing applications. For example, in the b uilding and construction sector , it can be used to monitor cracks, settlement-related damages, vibrations, and moisture. In transportation infrastruc- ture, including roads, ports, bridges, and rail ways, it can enable the assessment of changes in load-bearing capacity , ov erload conditions, structural forces, and unforeseen incidents. B. Bridge Monitoring Dataset The bridge monitoring data in this study are collected from a bridge located in Inland, Norway . T wo iBridge sensor devices (iBridge A and iBridge B) are installed beneath the bridge, approximately 70 meters apart, as shown in Fig. 7. For accident detection, real-time sensor data from both devices is considered ov er ele ven days, from August 15 to August 25, 2025. The sensor data includes measurements of acceleration (in mG) in X, Y and Z directions, respectively , and strain on the metal beams (in mV). It contains a total of 3,775,112 samples. These measurements consist of numerical v alues. The data is sampled at a frequency of 5 Hz, corresponding to one measurement ev ery 0.2 seconds. C. Data Pr epr ocessing The collected raw sensor data contain inconsistencies and missing v alues, which can adversely af fect accuracy and model performance. Therefore, the dataset is preprocessed to enhance data quality , reliability , and suitability for ML models analysis. The preprocessing stage in volv es data cleaning and handling missing entries. As the number of missing values is limited, these samples are discarded from the dataset. Alternativ ely , transfer learning [18] techniques can be employed to address missing data or data scarcity challenges. The processed dataset is then used to train ML models for anomaly detection. Next, we denote the notations for the features corresponding to iBridge A and iBridge B. For iBridge A, the accelerations in X, Y , Z directions, and strain on the metal beams are denoted as acx A, acy A, acz A, and adc2 A, respectiv ely . Similarly , for iBridge B, the corresponding features are denoted as acx B, acy B, acz B, and adc2 B, respecti vely . Since fiv e samples are recorded per second, the data are resampled at a one-minute interval to enhance data pattern clarity , highlight anomaly peaks, improve temporal uniformity , and stabilize model learning, thereby enabling more reliable ML models. For visualization, the final resulting dataset is sho wn in Fig. 8. Since the data ranges of acx A and acx B are significantly larger than those of the other features, they are illustrated separately in Fig. 9. The remaining features, excluding acx A and acx B, are shown in Fig. 10. A ug 16 2025 A ug 17 A ug 18 A ug 19 A ug 20 A ug 21 A ug 22 A ug 23 A ug 24 A ug 25 −100 −50 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000 1050 1100 variable acx_A acy_A acz_A adc2_A acx_B acy_B acz_B adc2_B Timeline Value Fig. 8: All features in the bridge monitoring dataset. A ug 16 2025 A ug 17 A ug 18 A ug 19 A ug 20 A ug 21 A ug 22 A ug 23 A ug 24 A ug 25 1000 1010 1020 1030 1040 1050 variable acx_A acx_B Timeline Value Fig. 9: acx A and acx B of the bridge monitoring dataset. A ug 16 2025 A ug 17 A ug 18 A ug 19 A ug 20 A ug 21 A ug 22 A ug 23 A ug 24 A ug 25 −40 −20 0 20 40 60 variable acy_A acy_B acz_A acz_B adc2_A adc2_B Timeline Value Fig. 10: All features except acx A and acx B of the dataset. 01:14:00 Aug 24, 2025 01:14:15 01:14:30 01:14:45 01:15:00 01:15:15 01:15:30 01:15:45 01:16:00 0 200 400 600 800 1000 acz_A acz_B acx_B acx_A acy_A Time Value Accident Point: Detection Peak (01:15:01) Fig. 11: Actual (ground-truth) accident timestamp. V . R E S U LT S A N D D I S C U S S I O N S This section presents the experimental environment and a comprehensiv e analysis of the results. A. Experimental En vir onment All algorithms are implemented in Python 3.13.6 using the Keras framew ork built on T ensorFlow 2.2.0 and executed on a MacBook with an Apple M4 chip and 16 GB of RAM. B. Gr ound T ruth The exact timing of the accident is subjectiv ely determined through a detailed manual analysis of the sensor data, which serves as the ground truth in this study , as shown in Fig. 11. The accident timing is identified as August 24, 2025, at 01:15 AM. Although this manual inspection enables accurate identification of the accident ev ent, it requires expert kno wl- edge and substantial human ef fort, making it impractical for continuous and real-time bridge monitoring. T o address this limitation, we subsequently present ML-based anomaly detec- tion models designed to automatically identify such events in a smart bridge monitoring system. C. Pr oposed Model Analysis The bridge monitoring dataset described in Section IV -B is employed to train three dif ferent ML models (see Section II) for detecting the precise timing of the accident as an anomaly in the smart bridge monitoring system. The dataset is first scaled to a common range to ensure fair feature contribution and reduce false alarms. A grid search [19] is then employed T ABLE I: Parameters used in each ML model. Models Parameters Isolation F orest n estimators 1 =1000, contamination=0.0001, n jobs=-1 Autoencoder input layer dimension=8, encoder input dimension=4, activ ation function=tanh, bottleneck layer dimension=1, activ ation function=linear , decoder input dimension=4, activ ation function=tanh, output layer dimension=8, activ ation function=linear , optimizer=adam, loss=mse, epochs=100, batch size=512 DBSCAN eps 2 =0.8, min samples=3, n neighbors=2 2025-08-16 2025-08-17 2025-08-18 2025-08-19 2025-08-20 2025-08-21 2025-08-22 2025-08-23 2025-08-24 2025-08-25 T imeline 0.0 0.2 0.4 0.6 0.8 1.0 Nor malized Anomaly Scor e 08-16 20:56 08-23 14:11 08-24 11:02 08-24 11:03 08-24 11:04 Anomaly Str ength Nar r owed R esult Fig. 12: Detected anomalies using the Isolation Forest model. to obtain the optimal parameters for each model, which are presented in T able I. 1) Isolation F or est Analysis: The Isolation Forest model is trained with the parameters presented in T able I. A ranking- based anomaly selection strategy is employed to identify the most anomalous samples. The samples are ranked according to their normalized anomaly scores, where lower values indicate stronger anomalies and higher values correspond to more normal beha viour . This approach focuses on detecting the most extreme deviations. Based on this strategy , Fig. 12 illustrates the detected anomalies. It can be seen that although a small number of prominent anomaly peaks (fi ve in total) are ob- served, none coincide with the accident timestamp (August 24, 01:15 AM). This indicates that the Isolation Forest model cannot accurately determine the exact timing of the accident. 2) Autoencoder Analysis: The autoencoder model is trained using the parameters presented in T able I and learns to re- construct input samples based on the underlying data patterns. The reconstruction error (see Section II-B) is computed for the samples and used as an anomaly score to identify abnormal behaviour . The samples are subsequently ranked according to their normalized reconstruction error, where smaller values indicate accurate reconstruction and normal behaviour , while larger v alues correspond to poor reconstruction and anomalous behaviour . Based on this methodology , Fig. 13 presents the detected anomalies. It can be seen that although a limited number of anomaly peaks (five in total) are observed, none 1 In Isolation Forest, n estimators is the number of isolation trees, contam- ination is the expected anomaly proportion and n jobs is the av ailable CPU cores (n jobs=-1= all available). 2 In DBSCAN, eps is the maximum distance between two neighbour sam- ples and min samples is the minimum number of samples in a neighbourhood to form a dense region. 2025-08-16 2025-08-17 2025-08-18 2025-08-19 2025-08-20 2025-08-21 2025-08-22 2025-08-23 2025-08-24 2025-08-25 T imeline 0.0 0.2 0.4 0.6 0.8 1.0 Nor malized R econstruction Er r or 08-20 05:35 08-21 06:17 08-22 06:31 08-23 21:56 08-24 16:50 Anomaly P r obability Nar r owed R esult Fig. 13: Detected anomalies using the autoencoder model. 2025-08-16 2025-08-17 2025-08-18 2025-08-19 2025-08-20 2025-08-21 2025-08-22 2025-08-23 2025-08-24 2025-08-25 T imeline 0.0 0.2 0.4 0.6 0.8 1.0 Nor malized Outlier Distance 08-21 06:16 08-24 01:14 08-24 01:15 08-24 01:16 08-25 04:33 Outlier Intensity Nar r owed R esult Fig. 14: Detected anomalies using the DBSCAN model. coincide with the accident timestamp (August 24, 01:15 AM). This suggests that the autoencoder model cannot accurately determine the exact timing of the accident. 3) DBSCAN Analysis: The DBSCAN model is trained us- ing the parameters presented in T able I. T o identify anomalous samples, a nearest-neighbour distance-based anomaly selection strategy is adopted (see Section II-C). The samples are classi- fied as normal or anomalous based on their normalized outlier distances (anomaly scores), where larger distances indicate stronger anomalies and smaller distances correspond to weaker anomalies. Using this approach, Fig. 14 shows the detected anomalies. Notably , a pronounced anomaly peak is observed at August 24, 01:15 AM, coinciding with the actual accident timestamp. This indicates that the DBSCAN model performs better than the other models and is capable of accurately detecting the timing of the accident, making it suitable for anomaly detection in the smart bridge monitoring system. V I . C O N C L U S I O N S A N D F U T U R E W O R K This paper in vestigates the anomaly detection problem in bridge monitoring systems using real-world sensor data collected by the iBridge sensor de vices. The dataset is first analyzed and preprocessed to ensure data quality and relia- bility . Motiv ated by recent advances in artificial intelligence, we propose a simple machine learning model that learns discriminativ e features directly from sensor data and employs a decision module to detect anomalies, with a particular focus on identifying a bridge accident in Norway . Experimental results show that the proposed approach based on DBSCAN accurately detects anomalous ev ents (bridge accident) and outperforms other ev aluated ML models. Our proposed model has the potential to significantly enhance anomaly detection capabilities in bridge monitoring systems, thereby reducing the risk of catastrophic failures or accidents and improving public safety . Future work will e xplore the integration of more complex feature representations and the dev elopment of an early warning system for proactive smart bridge monitoring. A C K N O W L E D G E M E N T This work was supported by Smart Sensor Systems AS through data provision and conference dissemination support. R E F E R E N C E S [1] The Large Norwegian Encyclopedia, https://snl.no/bro. [2] J. Han, J. Pei, and H. T ong, Data Mining: Concepts and T echniques . Morgan Kaufmann, 2022. [3] D. Samariya and A. Thakkar, “A Comprehensiv e Survey of Anomaly Detection Algorithms, ” Annals of Data Science , vol. 10, no. 3, pp. 829– 850, 2023. [4] R. Edirisinghe and G. Zhang, “Markov Model-based Building Deteri- oration Prediction and ISO factor Analysis for Building Management, ” Journal of Management in Engineering , vol. 31, no. 6, pp. 1–9, 2015. [5] E. J. Elton, M. J. Gruber, S. J. Bro wn, and W . N. Goetzmann, Modern P ortfolio Theory and In vestment Analysis . John W iley & Sons, 2009. [6] R. Jaiswal and A. Priya, “Leveraging Machine Learning for Fraud Identification in International Shipping, ” in International Conference on Signal Pr ocessing and Communication , 2025, pp. 470–474. [7] R. Jaiswal and A. Hines, “The Sound of Silence: How Traditional and Deep Learning based V oice Activity Detection Influences Speech Quality Monitoring, ” in AIAI Irish Conference on Artificial Intelligence and Cognitive Science , 2018, pp. 1–12. [8] R. K. Jaiswal and R. K. Dubey , “CA QoE: A Novel No-reference Context-A ware Speech Quality Prediction Metric, ” ACM T rans. on Multimedia Comp., Comms. and Apps , vol. 19, no. 1s, pp. 1–23, 2023. [9] A. Entezami, H. Sarmadi, B. Behkamal, and S. Mariani, “Big Data Analytics and Structural Health Monitoring: A Statistical Pattern Recognition-based Approach, ” Sensors , vol. 20, no. 8, p. 2328, 2020. [10] O. Abdeljaber , O. A vci, S. Kiranyaz, M. Gabbouj, and D. J. Inman, “Real-time V ibration-based Structural Damage Detection using One- dimensional Convolutional Neural Networks, ” Journal of Sound and V ibration , vol. 388, pp. 154–170, 2017. [11] M. Tian, “Anomaly Detection and W arning in Bridge Structure Health Monitoring: Method Based on Autoencoder, ” Pr ocedia Computer Sci- ence , v ol. 262, pp. 588–595, 2025. [12] Y . Chabchoub, M. U. T ogbe, A. Boly , and R. Chiky , “An In-depth Study and Improvement of Isolation Forest, ” IEEE Access , vol. 10, pp. 10 219– 10 237, 2022. [13] H. Xu, G. Pang, Y . W ang, and Y . W ang, “Deep Isolation Forest for Anomaly Detection, ” IEEE T ransactions on Knowledge and Data Engineering , v ol. 35, no. 12, pp. 12 591–12 604, 2023. [14] A. Abhaya and B. K. Patra, “An Efficient Method for Autoencoder Based Outlier Detection, ” Expert Systems with Applications , vol. 213, p. 118904, 2023. [15] Z. W ang and A. C. Bovik, “Mean squared error: Love it or leav e it? A New Look at Signal Fidelity Measures, ” IEEE Signal Pr ocessing Magazine , vol. 26, no. 1, pp. 98–117, 2009. [16] C.-F . Chien and B. Suwattananuruk, “Density-based Spatial Clustering of Applications with Noise (DBSCAN) for Probe Card Production for Advanced Quality Control of W afer Probing T est, ” IEEE T ransactions on Semiconductor Manufacturing , vol. 37, no. 4, pp. 567–575, 2024. [17] iBridge Sensor Device: A Smart Sensor-based Monitoring System, https://smartsensorsystems.no/en/products/ibridge-smart-data-logger/. [18] R. Jaiswal, M. Elnourani, S. Deshmukh, and B. Beferull-Lozano, “Location-free Indoor Radio Map Estimation using Transfer learning, ” in 97th V ehicular T echnology Conference . IEEE, 2023, pp. 1–7. [19] E. Elgeldawi, A. Sayed, A. R. Galal, and A. M. Zaki, “Hyperparameter T uning for Machine Learning Algorithms Used for Arabic Sentiment Analysis, ” in Informatics , vol. 8, no. 4. MDPI, 2021, pp. 1–21.