As Structural Health Monitoring (SHM) systems migrate toward edge computing environments, challenges related to computational efficiency and energy consumption become increasingly critical. Although many diagnostic algorithms perform well under specific conditions, they often struggle to achieve high-precision multiclass damage identification, distinguishing different types or severity levels of damage on resource-constrained hardware. To address these limitations, this study proposes a lightweight diagnostic framework based on Hyperdimensional Computing (HDC), a brain-inspired computational paradigm characterized by high robustness and low complexity. The proposed framework consists of three stages: (1) Feature extraction, which employs an overlapping window segmentation to capture dynamic evolution and extracts time-frequency descriptors and autoregressive (AR) features for complementary structural representations; (2) HDC encoding, which proposes a joint mechanism to map channel positions, feature types, and quantized values into a unified hyperdimensional space for preserving spatiotemporal dependencies; (3) Classification, which implements a prototype-based strategy to achieve rapid damage identification through efficient inner-product similarity matching. Experimental evaluations on the ASCE benchmark dataset confirm that the framework captures essential vibration patterns with a high accuracy of 95.08%. The model maintains a minimal computational footprint of 250 KB with millisecond-level inference latency, demonstrating a superior balance between diagnostic accuracy and practical deployability for real-time SHM.

Detecting hidden detachments and subsurface discontinuities in multilayer architectural finishes, civil architectures, and industrial coverings is a critical step in structural diagnostics and conservation. In cultural heritage practice, conservators commonly perform auscultation — gently tapping and listening for “dull” versus “clear” sounds — a fast but subjective procedure that is poorly repeatable. To remove operator bias and provide quantitative in-situ screening, we present a compact, battery-operated instrumented impact probe for the rapid detection of delaminations and anomalies in local stiffness.
The prototype, whose block diagram is shown in Fig. 1, is based on an ESP32-S3 module (dual-core Xtensa LX7, up to 240 MHz, native I²S and rich digital peripherals, Wi-Fi + BLE transceiver, and low-power modes) enabling on-device signal processing and wireless streaming. The front-end integrates an STMicroelectronics MP23DB02MM digital pulse density modulation (PDM) MEMS microphone together with an Analog Devices ADXL355 high-resolution MEMS accelerometer up to 20-bit, ultra-low noise spectral density of 22.5 μg/√Hz, programmable range up to ±8 g, and user-selectable output data rates; in our setup, a sampling frequency of 4 kHz was selected to capture impact transients and early-time vibration energy. Finally, to provide repeatable impulse scanning functionalities, a low-voltage, low-power solenoid is integrated. The current system is implemented as a USB peripheral for PC operation and can be converted to a wireless node for untethered surveys. The signal-processing chain is optimized to preserve low-energy, material-dependent responses while suppressing the solenoid’s intrinsic excitation. The pipeline is engineered to be portable on the host microcontroller unit (MCU), from data pre-processing (building on robust lag estimation and temporal alignment) to automatic impulse detection. The latter encompasses an adaptive spectral-subtraction module necessary to cancel the actuator signature, which is based on short-time Fourier transform (STFT) analysis, inverse-STFT reconstruction with normalization, STFT delay correction and optional fine alignment. From the reconstructed vibration channel, we compute a compact set of features, including the peak and the root-mean-square acceleration, and the amplitude and frequency of the dominant FFT component, which are inexpensive to evaluate, thus, being suitable for on-device thresholding, rule-based screening, or as input to a tiny AI-based on-edge classifier.

Laboratory validation on three reference materials spanning thin/soft to thick/rigid behaviors (cardboard of 3 mm, plastic of 1 mm, solid wood 15 mm) shows clear differentiation in peak acceleration statistics: peak acceleration of 0.041 g for cardboard, 0.059 g for plastic, and 2.32 g for wood. The adaptive spectral-subtraction block effectively suppresses the solenoid-related spectral content (i.e., the actuator motion signature) and uncovers material-specific responses even when spectral overlap occurs.

The handheld platform is promising for rapid cultural-heritage surveys, civil substrate inspection and industrial NDT screening. Its modest computational footprint enables real-time execution on MCUs and makes the platform suitable for on-device TinyML models supporting noise subtraction, data fusion, and automated defect detection and mapping, alongside expanded multilayer test campaigns and localization capabilities.

Phased array ultrasonic testing (PAUT) provides high-resolution subsurface imaging through electronic beam steering and focusing and is widely used for internal defect diagnosis in structures. However, effectively interpreting PAUT data requires understanding complex spatial-temporal patterns, and deep learning has recently been introduced to address this challenge. While deep learning–based automatic interpretation has improved performance [1], most prior approaches rely on heavy CNN models that inadequately capture long-range dependencies and incur high computational cost, limiting real-time operation and on-device deployment.

This work addresses these limitations by proposing a PAUT defect-classification framework that minimizes false negatives by applying class-weighted Knowledge distillation (KD) to transfer a Vision Transformer (ViT) teacher’s diagnostic capability to an extremely lightweight linear student. The proposed approach preserves high reliability while achieving real-time performance.

The proposed method has three stages. First, we construct a dataset. PAUT responses are sensitive to beam angle, which can obscure or weaken defect echoes, so we pair left and right acquisitions and use them together as a two-channel input. Next, we train a ViT as the teacher model with class-weighted cross-entropy so that the model becomes more sensitive to defect classes. Finally, we transfer the teacher’s knowledge to a lightweight linear student through class-weighted KD. In this study, the lightweight linear student model consists of one convolutional layer and one linear layer. Standard KD combines a KL divergence term between teacher and student distributions with a ground-truth cross-entropy term, and the ground-truth term can reintroduce false alarms even when the teacher suppresses them. To avoid this issue and to preserve recall on defect classes, we apply the same class weights to the ground-truth term during distillation.

To validate the performance of the proposed method, we compared the student model trained with the proposed scheme against ResNet and ViT models. On 937 test samples, all three models achieved an accuracy above 90%. However, while both ResNet and ViT produced two false negatives, the proposed method resulted in zero false negatives. In addition, the inference throughput was 227 images/s for ResNet and 135 images/s for ViT, whereas the proposed method achieved 2,482 images/s, reducing the inference time to less than one tenth of that of the baseline models. Per-sample inference time was about 0.0004 s, confirming real-time operation.

The target computing platform for this work is an embedded inspection system under development at CEA, intended for integration with PAUT sensing hardware. While the proposed framework is ultimately envisioned to be deployed at the edge, directly on PAUT systems, the present study focuses on algorithmic feasibility rather than full system-level deployment.

References

[1] Guyon, R., Newson, M., Fisher, C., Miorelli, R., & Roué, D. (2025, July). Towards embedded AI models for welding defect detection in pipes. In 2025 IEEE Sensors Applications Symposium (SAS) (pp. 1-6). IEEE.

In the context of Structural Health Monitoring (SHM) based on ultrasonic guided waves, one of the major challenges lies in the strong sensitivity of signals to environmental and operational variations, particularly temperature changes. These fluctuations can induce significant distortions, leading to false alarms if not properly compensated. In addition, embedded sensors generate large volumes of data, while high-speed connectivity (e.g. Wifi, 4G, 5G) is not always available or may be costly, thereby requiring local processing. As a result, an SHM system intended for real-world deployment must be capable of performing on-board data analysis while accounting for physical constraints, such as environmental and operational conditions.

In this context, the CEA-List has been working over the past years on the development of a complete SHM processing chain, ranging from the hardware design of guided-wave acquisition system to advanced data processing and analysis methods. In this work, we present a fully embedded solution based on the Geronimo platform, integrating signal acquisition, environmental effect compensation, imaging techniques and decision-making algorithms to convert complex ultrasonic measurements into reliable alerts. Artificial intelligence is implemented at two key stages: first, an autoencoder-based model is used to address the limitations of conventional emperature compensation methods. Second, as the selected imaging technique remains qualitative, a convolutional neural network is used to estimate the position and the size of defects.

Preliminary laboratory experiments have been conducted independently for each stage of the processing chain. The autoencoder-based compensation model was evaluated on several experimental datasets, showing excellent performances compared to conventional approaches. A classic imaging method, namely delay and sum, has been successfully deployed on the Geronimo platform, achieving a computation speed compatible with real-time acquisition. Then, analysis models trained on simulation data generated by the CIVA software and adapted to the real conditons via transfer learning on experimental datasets were implemented on system, enabling accurate defect characterization. Finally, all processing stages were fully integrated and combined within the Geronimo platform to evaluate the performance of the complete SHM chain under laboratory conditions.

The safety and integrity of composite structures can be ensured by real-time detection of damages such as impact events or delamination. Real-time Structural Health Monitoring (SHM) systems help to achieve immediate damage detection, localization, and even quantification with a fast response time. At the same time, the acquisition of damage-sensitive indicators from multiple onboard sensors generates vast amounts of data, imposing high computational resources. This requires the development of low-power embedded systems capable of performing high-speed and efficient data processing. Field-Programmable Gate Array (FPGA)-based systems address these challenges by offering parallel processing capabilities, flexibility, high-speed multi-I/O interfaces, and low power consumption. Machine learning algorithms aid SHM by processing extensive datasets to automatically extract damage-sensitive features, enabling the analysis and detection of damage.

The proposed system utilizes Ultrasonic Guided Waves (UGWs) to detect damage-sensitive features in composite structures. Piezoelectric transducers (PZTs), attached permanently to the structures, are used to transmit and receive UGWs, that exhibit feature changes when damage is present. Pre-processing data using various signal processing techniques is carried out with integrated electronic embedded devices after data acquisition. As the primary objective of the system is real-time monitoring using FPGA, the machine learning model was trained with synthetic data, under simulated damage conditions, considering hardware resource requirements. The model, reconstructed using trained parameters in FPGA, was evaluated for latency performance.

This paper presents a real-time SHM system that combines the capabilities of FPGA and machine learning to evaluate the durability and integrity of composite structures. The developed ML model achieved over 95% accuracy in classifying damages and estimated damage sizes with a standard deviation less than 2 mm. Implementing an application on an FPGA using traditional design methods (Hardware Description Languages, HDLs) demands substantial learning and significant development time. To address these challenges, a framework known as High Level Synthesis (HLS) was employed to develop an FPGA system, enabling the hardware behavior to be described using high-level programming languages such as C or C++, thereby accelerating the learning curve. A range of optimization techniques using HLS was applied, resulting in parallel execution and reduced latency. Overall, this system offers a robust and rapid solution for real-time SHM, thereby contributing to condition-based maintenance and damage escalation prevention.

Structural Health Monitoring (SHM) increasingly relies on data-driven methods applied to dense vibration measurements; however, reliable damage detection from acceleration time series remains challenging due to environmental variability, limited labeled damage data, and severe class imbalance.

This work investigates self-supervised learning for anomaly detection in bridge acceleration signals. First, we adopt a contrastive framework that captures temporal and contextual dependencies to pretrain an encoder exclusively on fixed-length windows of data collected under healthy conditions. To account for real-world variability, physics-informed data augmentations are introduced to simulate measurement disturbances such as sensor noise and signal variations. Following pretraining, anomalies are identified by measuring deviations of streamed windows from the learned representation of healthy structural behavior, enabling damage detection without explicit labels or predefined damage categories. Experiments on the RT345 multi-scenario bridge benchmark show that the learned representations achieve superior separation between healthy and damaged windows compared to classical baselines and reconstruction-based deep autoencoders.

Finally, we evaluate an online post-processing strategy that aggregates consecutive anomaly scores to emulate streaming deployment. While this aggregation improves detection sensitivity (achieving a True Positive Rate above 96%), it can increase false alarms when using a fixed operating point. These findings indicate that the proposed contrastive-learning-based approach enables effective damage detection from vibration signals in an online setting.