Engineers often cannot reach the parts of a structure that worry them most. Deep inside buildings, underground, or behind safety barriers, critical components may be impossible to instrument, leaving blind spots in structural health monitoring. A new study now shows that an artificial neural network can synthesize the experimental damage signals that such inaccessible locations would have produced, using numerical simulations as the raw material. The work, published in AI in Civil Engineering, reports prediction accuracy of around 99 percent on a laboratory aluminum plate, offering a baseline method for generating data where physical testing is impractical.
The research, led by Anoop Sharma, Neetika Saha, and Pijush Topdar, addresses a persistent gap in acoustic emission testing. Acoustic emission, or AE, is the release of stress waves caused by sudden movements or cracks in a material under stress, and it is widely used to detect and monitor the initiation of damage such as cracks, corrosion, and fatigue. In laboratory practice, these signals are frequently generated with the pencil lead break, or PLB, method, in which a mechanical pencil lead is broken against the surface of a specimen to simulate a damage source without actually harming the structure. The technique, also known as the Hsu-Nielsen source, is a standard way of producing synthetic damage signals under controlled conditions. But it requires physical access to the point being tested, which is precisely what inaccessible locations lack.
The team’s solution was to build a bridge between two kinds of data. They modeled the same structure numerically using the finite element method in the commercial software ABAQUS, obtaining mechanical displacement-versus-time responses, and they measured real electrical voltage-versus-time signals experimentally with an acoustic emission sensor. The numerical route is considerably simpler and more computationally efficient than producing the voltage output, and it can be performed anywhere in a model, including points that could never be tested physically. The purpose of the study was to train a neural network to convert one into the other: to synthesize the experimental voltage-type signal parameters from the numerically simulated displacement-type signal parameters.
The specimen was deliberately simple. The researchers used an isotropic aluminum plate measuring 1000 millimeters by 600 millimeters with a thickness of 1 millimeter. The plate was divided into four quadrants, each subdivided into square grids of 10 millimeters by 10 millimeters, giving 299 nodes per quadrant. Because the material is isotropic, the team exploited quadra symmetry and worked with one quadrant, Quadrant I, for the analysis. In the experiment, pencil leads of 0.5 millimeter diameter were broken against the plate surface at an inclination of 45 degrees at each of these nodes, while a wideband WD sensor mounted at the center of the plate captured the resulting AE signals. The sensor, known to operate effectively over a frequency range of 100 to 900 kilohertz, was coupled to the surface with vacuum grease, and a 2/4/6 preamplifier with a gain of 40 dB filtered out low-frequency noise and amplified the signals before data acquisition.
In parallel, the finite element model reproduced the same scenario. The plate was modeled as a three-dimensional deformable body using shell elements with simply supported edges, a mesh size of 5 millimeters by 5 millimeters, and linear 8-noded thick shell elements. Damage initiation was simulated with a mathematical forcing function corresponding to the PLB, applied at all nodes of Quadrant I, with the PLB simulation conducted for a duration of 100 microseconds. Displacement-time responses were collected at the center of the plate for every forcing application, mirroring the position of the physical sensor. The result was two parallel datasets, numerical and experimental, covering the same 299 locations.
Before any training could begin, the team had to decide which features of the signals mattered. They extracted a set of conventional and statistical acoustic emission parameters from both datasets: energy, amplitude, root mean square, or RMS, shape factor, and kurtosis. Energy represents the total strength of a signal over time, amplitude is the highest peak voltage attained by the waveform, RMS reflects the average intensity of the signal, shape factor is the dimensionless ratio between RMS value and average amplitude that characterizes smoothness or sharpness, and kurtosis describes the peakedness of the signal and how much of its energy is concentrated in the peak compared with the tails.
Two filtering steps narrowed this list down. First, parameters that remained constant across all data points were removed, because zero variance provides no discriminative information for a learning model and only adds computational cost. Second, the researchers applied Pearson correlation analysis, a statistical method that measures the strength and direction of the linear relationship between two continuous variables through a coefficient ranging from minus 1 to plus 1. When two parameters showed a high correlation, with an absolute coefficient above 0.9, they were considered to carry redundant information and one was dropped. In the numerical dataset, energy and RMS were perfectly correlated with a coefficient of 1, and shape factor and kurtosis showed a coefficient of approximately 0.97, so energy and kurtosis were identified as redundant. In the experimental dataset, energy and RMS correlated at 0.95, again triggering removal. The survivors in both cases were RMS, shape factor, and amplitude, which became the inputs and outputs of the network.
The model itself was a Levenberg-Marquardt artificial neural network built in MATLAB, a choice suited to datasets of moderate size. Its architecture was compact: an input layer of three neurons for the three selected numerical parameters, two hidden layers of 10 neurons each, and an output layer of three neurons producing the corresponding experimental parameters, making it a multiple-input, multiple-output system. A sigmoidal activation function was used, and training employed the feedforward backpropagation approach with the Levenberg-Marquardt algorithm. All 299 sample data points were used, divided into 80 percent for training, 10 percent for testing, and 10 percent for validation. The regression coefficient for the training phase came out at 0.99836, validation reached 0.99895, and testing achieved 0.99869, with an overall value of 0.99845 across the entire dataset, indicating minimal error and no evidence of overfitting.
Individual parameters told a more varied story. When the model’s predictions were examined parameter by parameter, amplitude achieved a regression value of 0.9008, while shape factor and RMS reached 75.34 percent and 81.14 percent respectively. Mean absolute errors were 0.0799 for RMS, 0.1012 for shape factor, and 2.3648 for amplitude, with amplitude showing the highest error percentage at 2.36 percent and RMS the lowest at 0.0799 percent. The training performance plot showed the mean square error decreasing and stabilizing around 110 iterations over 160 epochs, with the best validation performance, marked at the early stopping point, reaching an MSE of 0.038942. An error histogram built from 20 bins showed a tall central bar near zero, meaning most prediction errors were very small, with only a few outliers on either side.
The decisive test came with locations the model had never seen. Five random points, designated a through e, were selected from the test sample as stand-ins for inaccessible locations where voltage responses had not been obtained. The model predicted the RMS, shape factor, and amplitude values for each, and the errors ranged between 0.14 percent and 5.23 percent, which the authors describe as satisfactory performance. The best performance was achieved for location b. The model was also tested with an independent set of input data drawn from a different quadrant, whose data had not been used for training or validation, and its performance was again found to be satisfactory. The authors emphasize that this is a baseline methodology demonstrated on a structure with simple geometry and material properties under controlled laboratory conditions, and that future work could extend the approach to more complex structures and diverse environmental conditions, potentially integrating it with IoT-based platforms for structural health monitoring.
Subject of Research: Machine learning synthesis of acoustic emission damage signals at inaccessible structural locations
Article Title: Prediction of damage signals at inaccessible locations using a machine learning approach
Article References: Sharma, A., Saha, N., & Topdar, P. (2026). Prediction of damage signals at inaccessible locations using a machine learning approach. AI in Civil Engineering, 5(1), Article 8. https://doi.org/10.1007/s43503-026-00090-0
Image Credits: AI Generated
DOI: 10.1007/s43503-026-00090-0
Keywords: acoustic emission, artificial neural network, structural health monitoring, finite element method, pencil lead break, damage detection, machine learning, aluminum plate, signal prediction, non-destructive testing, Pearson correlation, Levenberg-Marquardt
News Source: Blake Davidson. (November 4, 2026). Neural network predicts structural damage signals where sensors cannot reach. Scienmag.
