Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Learning Damage Event Discriminator Functions with Distributed Multi-instance RNN/LSTM Machine Learning - Mastering the Challenge

Stefan Bosse

University of Bremen, Dept. Mathematics & Computer Science, Bremen, Germany

sbosse@uni-bremen.de

1 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Introduction

Motivation

This work addresses a novel distributed machine learning multi-instance approach to overcome limitations and flaws in decentralised Structural Health Monitoring using decentralised single-instance sensor processing

2 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Objectives

  1. Robust prediction of hidden damage events in mechanical structures using raw sensor time series and time-series prediction;

  2. Typical environmental vibrations of the structure are used for the measuting stimulus (no actuators are required);

  3. Scalability: The sensor processing and learning is performed locally on sensor node level with a global fusion of prediction results to estimate the damage location and the time of the damage creation.

3 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Structural Monitoring

There are at least four different levels of information that can be delivered by a Structural Health Monitoring (SHM) systems:

  1. Detection of damages and material changes;

  2. Localization of damage;

  3. Assessment of damages and impact on operational safety;

  4. Prediction of mechanical and operation behaviour.

4 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Sensor Networks

Integration Levels

  1. Traditional Sensor Networks used for SHM are applied separately to the structure
  2. Ongoing progress in miniatursisation enables Material-intergrated Sensor Networks → Sensorial Materials!

Material-integrated Sensor Networks

5 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Machine Learning

  • Two architectures:
    • Centralised Single-instance Learner
    • Decentralised Multi-instance Learner with Fusion

6 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Simulation and Modelling

Get the Data!

The dynamic behaviour of mechanical structures are simulated with a simple Multi-body Physics engine and model to compute virtual sensors for the experiments!

7 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Multi-body Physics Model

(a) Mass-spring model of structure (b) Sensor Node Network (c) Strain Sensors (d) Virtual defects and disturbant loads

8 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Concept

The sensor processing and damage prediction concept: Local sensor processing, learning, inference ⇒ Global fusion and prediction of position and time of damage event

9 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Training Data Variance

  • One major issue in training of machine learned models is specialisation of the model!

  • Broad variance of training data samples are required for generalised models!

  • Experimental collection of large sensor data bases with high variance is difficult to achieve!

  • Simulation can overcome this limitation:

    • Using Monte Carlo methods applied to sensor signals and experimental parameters create a broad variance of data samples!
10 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Different device and measuring setups together with MC randomising sensor and parameter data create a diverse data base used for ML

11 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Learning Predictor Functions

Using Associated Time-series Prediction with Recurrent state-based Artificial Neural Networks for Damage Prediction

12 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Time-Series Prediction

  • Recurrent ANN are state-based and remembering the history (e.g., of a series of data points)

  • A Long-short Term Memory cell (LSTM) architecture can be used for:

    • Time-series Prediction of a time-resolved sensor signal (e.g., strain gauge sensor) or data point series, i.e., there is a f(xi): xixi+δ

    • Associated data-series Prediction with a input-output variable mapping, i.e., there is a f(xi): xiyi+δ

13 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

(a) Simple Recuccrent Neural Network; data activates network sequentially (b) LSTM cell architecture (c) Time- and data series prediction xixi+δ (d) Associated data series prediction xiyi+δ

14 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Damage Location Prediction

Get the location by global fusion of local predictions!

Distributed Multi-Instance Prediction by global mass-of-center computation (D: Damage, S: Sensors, F: Damage Discriminator Function, d: damage, cross: estimated position of damage)

15 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Results and Evaluation

Experiment.

Artificial Sensor Network with 7 × 4 nodes. Each node was equipped with two orthogonal strain-gauge sensors.

A simple environmental vibration was the stimulus.

After a specific simulation time step t=tn a defect was introduced (hole).

The time-resolved signal record was processed.

16 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Distributed Damage Prediction

Network Activation (Local Prediction)

Examples of the response of the individual discriminator functions of the distributed sensor network on time records (around t=[100,150]) with different damage cases (H1..H9) occurred at t=100.

17 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Prediction Accuracy

Damage t100:nfound t250:nfound t100 error t250 error
None 0 0 - -
H1 10 9 1.6 ± 0.3 (12%) 2.0 ± 0.02 (15%)
H2 7 0 1.7 ± 0.04 (13%) -
H3 10 7 1.5 ± 0.1 (11%) 1.7 ± 0.1 (13%)
H4 10 9 1.2 ± 0.04 (9%) 2.0 ± 0.7 (15%)
H5 10 3 1.2 ± 0.2 (9%) 1.4 ± 1.7 (11%)
H6 10 10 1.1 ± 0.07 (8%) 1.4 ± 0.2 (11%)
H7 10 3 1.9 ± 0.2 (15%) 1.2 ± 0 (9%)
H8 10 0 1.1 ± 0.2 (9%) -
H9 10 7 1.6 ± 0.3 (12%) 1.3 ± 0.3 (10%)

Fusioned true-positive damage event predictions; position prediction error ε ± 2σ, percentage value is position error relative to plate size)

18 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Conclusions

19 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Lessons Learned

  1. Although the training of state-based damage discriminator functions mapping time-series of raw sensor signals on damage detectors is a challenge, some remarkable results could be achieved.

  2. Due to overlapping damage detection areas (redundancy) of single nodes, the fusion of the outputs of all sensor nodes lead to a significant improvement of the overall damage event detection probability.

  3. Multi-body physics and mass-spring models were used to simulate a vibration of the DUT and to compute virtual strain sensors accurately enough to test the damage prediction approach!

20 / 21

Stefan Bosse: Learning Damage Event Discriminator Functions with Distributed Multi-instance Machine Learning

Thank You!

Thank you for your attention. All questions are welcome!

#evol

#me

21 / 21