Mid Sweden University

In this study, a convolutional neural network (CNN) model was developed for non-destructive damage classification of concrete materials based on acoustic emission techniques. The raw acoustic emission signal is used as the network model input, while the damage type is used as the output. In the study, 15,000 acoustic emission signals were used as the dataset, of which 12,000 signals were used for training, 1,500 signals for validation, and 1,500 signals for testing. Adaptive moment estimation (Adam) was used as the learning algorithm. Batch normalization and dropout layers were used to solve the overfitting problem generated in earlier versions of the model. The proposed model achieves an accuracy of 99.70% with 20,243 parameters, which provides a significant improvement over previous models. As a result, the classification of damages and decisions based upon them in non-destructive structural health monitoring applications can be improved. 

For the field of structural health monitoring (SHM), acoustic emission (AE) technology is important as a damage identification technique that does not cause secondary damage to concrete. Nowadays, applications of non-destructive concrete damage identification are mostly limited to commercial software or identification algorithms running on desktop computers. It has so far not been deployed in low-power embedded devices. In this study, a lightweight convolutional neural network (CNN) model for online non-destructive damage type recognition of concrete materials is presented and deployed on a resource-constrained microcontroller unit as a tiny machine learning (TinyML) application. The CNN model uses raw acoustic emission signals as input and damage recognition types as output. 15,000 acoustic emission signals are used as data sets divided into training, validation, and test sets in the ratio of 8:1:1. The experimental results show that an accuracy of 99.6% is achieved on the nRF52840 microcontroller (ARM Cortex M4) with only 166.822 ms and 0.555mJ for a single inference using only 20K parameters and 30.5KB model size. This work demonstrates the effectiveness and feasibility of the proposed model, which achieves a trade-off between high classification accuracy and deployability on resource-constrained MCUs. Consequently, it provides strong support for online continuous non-destructive structural health monitoring. 

Acoustic emission (AE) is a widely used non-destructive test method in structural health monitoring applications to identify the damage type in the material. Usually, the analysis of the AE signal is done by using traditional parameter-based methods. Recently, machine learning methods showed promising results for the analysis of AE signals. However, these machine learning models are complex, slow, and consume significant amounts of energy. To address these limitations and to explore the trade-off between model complexity and the classification accuracy, this paper presents a lightweight artificial neural network model to classify damage types in concrete material using raw acoustic emission signals. The model consists of one hidden layer with four neurons and is trained on a public acoustic emission signal dataset. The created model is deployed to several microcontrollers and the performance of the model is evaluated and compared with a state-of-the-art machine learning model. The model achieves 98.4% accuracy on the test data with only 4019 parameters. In terms of evaluation metrics, the proposed tiny machine learning model outperforms previously proposed models 10 to 1000 times. The proposed model thus enables machine learning in real-time structural health monitoring applications. 

This paper compares machine learning approaches with different input data formats for the classification of acoustic emission (AE) signals. AE signals are a promising monitoring technique in many structural health monitoring applications. Machine learning has been demonstrated as an effective data analysis method, classifying different AE signals according to the damage mechanism they represent. These classifications can be performed based on the entire AE waveform or specific features that have been extracted from it. However, it is currently unknown which of these approaches is preferred. With the goal of model deployment on resource-constrained embedded Internet of Things (IoT) systems, this work evaluates and compares both approaches in terms of classification accuracy, memory requirement, processing time, and energy consumption. To accomplish this, features are extracted and carefully selected, neural network models are designed and optimized for each input data scenario, and the models are deployed on a low-power IoT node. The comparative analysis reveals that all models can achieve high classification accuracies of over 99\%, but that embedded feature extraction is computationally expensive. Consequently, models utilizing the raw AE signal as input have the fastest processing speed and thus the lowest energy consumption, which comes at the cost of a larger memory requirement.