UNSUPERVISED VEHICLE CLASSIFICATION USING A STRUCTURAL HEALTH MONITORING SYSTEM

Amirhossein Moallemi, Luca Zanatta, Alessio Burrello, Mattia Salvaro, Mónica Longo, Paola Daró, Francesco Barchi, Davide Brunelli, Luca Benini, Andrea Acquaviva2023 · DOI

TL;DR

This paper presents a novel, cost-effective method for classifying vehicles by weight using only raw vibration data from existing Structural Health Monitoring (SHM) systems equipped with MEMS accelerometers. By leveraging unsupervised learning, they achieve high accuracy in distinguishing light, heavy, and super-heavy vehicles, offering a scalable alternative to expensive Weight-in-Motion (WiM) systems for bridge load monitoring.

What Problem Does This Paper Address?

Managing the structural integrity of bridges is a critical concern, especially with increasing traffic loads and aging infrastructure. Accurately assessing the weight of vehicles crossing bridges is crucial for preventing damage and ensuring safety. Traditional methods like Weight-in-Motion (WiM) systems provide precise data but are prohibitively expensive to install and maintain on a large scale. This creates a need for more affordable and scalable solutions that can leverage existing infrastructure. The paper addresses this gap by proposing a method to classify vehicles by weight using readily available data from SHM systems, which are often already deployed for structural monitoring.

Methods

The core of this work is an innovative pipeline that processes raw vibration data from MEMS accelerometers to classify vehicles into three weight categories: light (<10 tons), heavy (10-30 tons), and super-heavy (>30 tons). The process involves several key steps:

  • Data Acquisition: Raw vibration data is collected from MEMS accelerometers already part of an SHM system, alongside ground-truth data from a WiM system.
  • Preprocessing: The 2D vibration data (x and z axes) is converted into a 1D informative trace using L2 normalization. A 4th-order Butterworth filter is then applied to isolate low-frequency signals relevant to vehicle vibrations (0-15 Hz).
  • Energy Analysis: The energy of the filtered vibrations is computed using overlapping windows, followed by exponential smoothing to reduce amplitude variability and improve peak detection.
  • Vehicle Identification: A two-threshold approach (high for initiation, low for termination) is used to define the bounding box of a vehicle passage event, improving upon single-threshold methods.
  • Feature Extraction: Four key statistical features are extracted for each identified vehicle passage: Maximum Amplitude, Mean, Standard Deviation, and Line Length.
  • Data Labelling: Timestamps from the SHM and WiM systems are used to align and label the extracted features, creating a ground truth for unsupervised learning. A 2-minute time interval is used for this alignment.
  • Vehicle Classification: Unsupervised clustering algorithms (K-means, Mean Shift, and Gaussian Mixture Models - GMM) are employed to group the extracted features into the three defined weight classes. The paper specifically highlights the performance of the Mean Shift algorithm.

Datasets

The study utilizes data collected from a real-world viaduct scenario. This includes:

  • Raw vibration data: Acquired from 142 MEMS triaxial accelerometers deployed across 18 spans of the viaduct.
  • Ground-truth data: Obtained from a Weight-in-Motion (WiM) system located approximately 600 meters before the viaduct. This data includes timestamps, gross weight, velocity, class ID, and number of axles for each vehicle.

The dataset comprises five days of data, with four days used for training and one for validation. The data collection was primarily focused on nighttime periods due to lower traffic volumes.

Key Insights and Ideas

  • Leveraging Existing Infrastructure: The most significant insight is the demonstration that existing SHM systems, often equipped with affordable MEMS accelerometers, can be repurposed for vehicle classification without requiring new, expensive sensor installations.
  • Vibration Signatures as Proxies for Weight: The research effectively shows that the subtle vibrations induced by vehicles on a bridge can be reliably translated into meaningful features that correlate strongly with their gross weight.
  • Unsupervised Learning for Scalability: The use of unsupervised clustering algorithms is key to the scalability of the approach. It avoids the need for extensive, pre-labeled datasets for every possible vehicle type, making it adaptable to different traffic conditions.
  • Mean Shift's Robustness: The study highlights the Mean Shift algorithm's superior performance and robustness compared to K-means and GMMs, particularly in handling variations across different bridge spans.

Key Results

  • The proposed framework achieves a classification accuracy of 96.87% in the best-case scenario, utilizing the Mean Shift unsupervised clustering model.
  • The Mean Shift algorithm outperformed K-means by an average of 3.91% and was more robust than GMM, showing a standard deviation of 3.60% in classification accuracy across bridge sections compared to GMM's 5.91%.
  • The classification into three categories (light, heavy, super-heavy) is based on gross weight, which the authors argue is a more critical metric for bridge integrity than traditional axle-based classifications.
  • The extracted features (Maximum Amplitude, Mean, Standard Deviation, Line Length) showed a strong correlation with gross weight, validating their utility for classification.

Limitations and Future Work

  • Data Imbalance and Capture Issues: The paper notes that certain light vehicles (less than 2000 kg) might be captured by the WiM system but not by the accelerometers due to their low vibration signature. This can lead to a discrepancy in the number of events recorded by each system.
  • Variability in Vibration Signatures: The quantity and clarity of vibration samples can differ across intervals due to factors like lane changes or velocity reductions, potentially impacting the discernible vibration signal.
  • Future Work: The authors suggest exploring more advanced deep learning models for classification, potentially improving accuracy further. They also imply that the framework could be extended to incorporate other vehicle parameters or to monitor bridge health more comprehensively. Further investigation into the impact of different environmental conditions on vibration data could also be beneficial.

Summary updated

Abstract

The structural performance assessment of bridges is a crucial issue for managing transportation infrastructure systems in EU countries as traffic loads and structural ageing continues to increase. Weight-in-Motion (WiM) systems have been developed to estimate the gross weight of vehicles over a bridge and keep the bridge load under control. However, WiM systems are costly in procurement and installation; alternative approaches that aim to be more scalable and cost-effective are needed to respond to the need to monitor large-scale infrastructures. This work explores an innovative zeroincremental cost approach based on raw vibration data extracted from a system already deployed for Structural Health Monitoring (SHM) and based on MEMS accelerometers. A novel signal processing and classification pipeline has been developed to differentiate vehicles into three categories: light, i.e., less than 10 tons; heavy, i.e., between 10 and 30 tons; and super heavy, i.e., above 30 tons, using only features extracted from vibration data. The results show that this framework can distinguish vehicles with an accuracy of 96.87%, utilizing the mean-shift unsupervised clustering model. This method has the potential to be a significantly cost-effective and scalable solution for monitoring bridge loads compared to WiM systems, as it leverages existing SHM infrastructure and affordable MEMS sensors to provide real-time information on vehicular loads.
BibTeX
@misc{moallemi2023unsupervised,
  author = {Amirhossein Moallemi and Luca Zanatta and Alessio Burrello and Mattia Salvaro and Mónica Longo and Paola Daró and Francesco Barchi and Davide Brunelli and Luca Benini and Andrea Acquaviva},
  title = {{UNSUPERVISED VEHICLE CLASSIFICATION USING A STRUCTURAL HEALTH MONITORING SYSTEM}},
  year = {2023},
  doi = {10.12783/shm2023/36839},
}
Figure