Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors

Spatial Damage Prediction in Composite Materials using Multipath Ultrasonic Monitoring, advanced Signal Feature Selection and combined Classifier-Regression Artificial Neural Network

Stefan Bosse¹, Christoph Polle²

¹ University of Bremen, Dept. Mathematics & Computer Science, Bremen, Germany
² Faserinstitut Bremen, 28359 Bremen, Germany

sbosse@uni-bremen.de

1 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Introduction

Introduction

Damage detection in hybrid and laminate materials is still a challange!

Guided ultrasonic waves (GUW) can be used to detect hidden damages (e.g., caused by impact events)

But: The damage feature information contained in a time-resolved GUW signal is typically less that 1% of the signal energy, the ROI on the time-axis is not known in advance (depends on spatial parameters), and the damage feature depends on environmental parameters (e.g., temperature)

2 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Objectives

Objectives

The time-resolved GUW signal data is high-dimensional (typically several thousands variables) and requires high computational time for processing

In this work, multi-path sensing is combined with feature selection using numerical algorithms reducing the dimensionality of the input data signficantly (1:1000)

The pre-processed features are passed to a damage predictor function predicting the damage event and the damage position

3 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Data Flow

Data Flow

The principle data and methodological flow from measurement to damage prediction

4 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Predictor Model Function

Predictor Model Function

The predictor function combines a classifier and a spatial regression model, reducing computational and memory complexity, a constraint for the implementation in embedded sensor node systems.

One two-dimensional output predictor function M combines a damage classifier and a damage position (p_x=y₁,p_y=y₂) regression function

5 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Multi-path Sensor Data

Multi-path Sensor Data

A plate (device under test) is equipped with 12 piezo-electric transducers that can act as an ultrasonic actuator and sensor, too. Six direct paths are measured simultaneously. The time-resolved sensor data is processed by an analytical signal feature selection.

6 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Feature Selection for ML

Feature Selection for ML

Features

$\\ \psi_{{{i},{j}}}:\vec{{s}}_{{{i},{j}}}\to\vec{{f}}_{{{i},{J}}}\\ \vec{{s}}={\left\langle{s}{\left({t}={0}\right)},{s}{\left({t}={1}\right)},..\right\rangle}\\ \vec{{F}}={\left(\matrix{{T}\\\omega\\\vec{{f}}_{{{i},{j}}}}\right)}\\$

The features are derived from the maximum peak of the analytical signal of the GUW signal
T: Temperature, ω: Signal base frequency, f: Peak features

Model

$\\ {M}{\left({F}\right)}:\vec{{{F}}}\to\vec{{p}}\\ \vec{{{F}}}={\left(\matrix{{T}\\\omega\\\text{max}_{{{i},{j}}}\\\text{tmax}_{{{i},{j}}}\\..}\right)},\\ \forall{\left({i},{j}\right)}\in\text{conf}{\left({i},{j}\right)}\\ \text{conf}{\left({i},{j}\right)}={\left\lbrace{\left\langle{0},{6}\right\rangle},{\left\langle{1},{7}\right\rangle},{\left\langle{2},{8}\right\rangle},\right.}\\ {\left\langle{3},{9}\right\rangle},{\left\langle{4},{10}\right\rangle},{\left\langle{5},{11}\right\rangle}{\rbrace}\\ \vec{{p}}={\left({x},{y}\right)}\\$

The feature vector F is composed of feature vectors f_i,j of different measuring paths

7 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Measuring data

Measuring data

http://openguidedwaves.de

The GUW signal data was recorded from a CFK plate (500x500 mm) and attached pseudo defects. The data was taken from the OpenGuidedWaves data base.

There are two data sets:

Dynamic: With only four different defect positions but with a temperature profile (T=20°C-50°C]

Static: With 28 defect positions at a static temperature (T=24°C)

8 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors The Model

The Model

The predictor function is a feed-forward Artificial Neural Network
- The input layer consists of |F| neurons, one for each feature variable
- The output layer consists of two neurons predicting the damage positions in normalized and shifted corrdiantes
- One or two hidden layers with less than 10 neurons
Differents experiments:
- Training with dynamic and static data sets, prediction on both
- Training with dynamic data set, prediction on static

After training this simple model can be finally compiled in a standalone C function that can be used for embedded sensor nodes

9 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Results

Results

Results for different parameter configurations (ω=40kHz, D: Dynamic temperature data set, S: Static temperature)

10 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Results

Results

Results for different parameter configurations (ω=80kHz, D: Dynamic temperature data set, S: Static temperature)

11 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Results

Results

Computation times for the central feature selection (Computation of analytical signal by Hilbert transform) and the application of the ANN predictor model for various processing platforms (gcc: C and native machine code, node.js/quickjs: JavaScript code)

12 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors Conclusion

Conclusion

Using multi-path sensing of guided ultrasonic waves, advanced feature selection, and a simple artificial neural network we were able to detect pseudo defects applied to a CFK plate with a high probability (typically nearly 100%) and position accuracy (typically below 20 mm or better) in a wide range of material temperature (20-50°C).

It could be shown that the proposed analysis method is suitable to be implemented in embedded systems including material-integrated nano computers providing damage detection within 10 seconds after signal measurement, which is sufficient for a broad range of applications in SHM.

13 / 14

Stefan Bosse: Spatial Damage Prediction using combined Classifier-Regression Artificial Neural Network Predictors End.

End.

Thank you for your attention. All questions are welcome!

Further information can be found here:
- http://edu-9.de
- http://sblab.de

#evol

#me

14 / 14