Robust and Adaptive Non Destructive Testing of Hybrids with Guided Waves and Learning Agents

Inhalt

Non-destructive Testing (NDT) of Structures And Structural Health Monitoring (SHM)

Big Data Challenges

• Monitoring of mechanical structures is a Big Data challenge concerning Structural Health Monitoring and Non-destructive Testing

• The sensor data produced by common measuring techniques, e.g., guided wave propagation analysis, is characterized by a

  • High dimensionality in the temporal domain, and moreover
  • High dimensionality in the spatial domain using 2D scanning.

The quality of the results gathered from guided wave analysis depends strongly on the pre-processing of the raw sensor data and the selection of appropriate region of interest windows (ROI) for further processing (Feature Selection).

Features

Some Definitions

Feature Selection

The task of feature selection separates relevant (information correlated) from irrelevant (information uncorrelated) data and performs the first significant data reduction of measuring sensor data.

• Feature selection is traditionally hand made by experts, sometimes using regression or curve fitting

Feature Extraction

The task of feature extraction derives meaningful information from the already pre-selected input data.

• Damages (class, localization, depth, ..)
• Fatigue
• Load changes
• Lifetime prediction

Measurement and Analysis Challenges

Noise and Models

• Noisy data makes feature selection difficult and unreliable

• Usually complex material compositions and unknown damage models increase the unreliability of the feature selection and extraction task

Automatic Feature Selection

• Adaptive and reliable input data reduction is required at the first layer of an automatic structural monitoring system.

• Image segmentation can be used to identify ROIs as one relevant feature selection technique

Non-destructive Testing

• NDT usually performs only a few point-to-point measurements to detect damages.

• The two-dimensional recording of the wave propagation and interaction of guided waves can be performed by using laser vibrometry or an airborne ultrasonic testing technique → Measuring time and big data are critical tasks

• By adjusting the geometry of the actuator or its electrode configuration, the amplitudes of individual modes can be amplified or attenuated to emphasize specific wave interaction → Parameter setting is critical task

• The identification of damages is made by wave interactions, such as reflection, scattering, mode conversion and wave number changes, in wave propagation → Feature Selection is critical task

• A locally resolved scan of the wave propagation is required, producing wave propagation images with only a few regions of interest → Segmentation is critical task

Automatic Monitoring System

• Automated NDT system is proposed featuring:
  • Adaptive Segmentation & Feature Selection,
  • Machine Learning, Adaptive Filtering, and
  • SHM algorithms

Hybrid materials: The Challenge

Hybrid Materials

Hybrid materials combine dissimilar materials of different material classes in a way that the individual material-specific advantages become effective in an optimal manner within lightweight structures.

• Based on their outstanding lightweight potential hybrid materials penetrate more and more into applications in transportation.

  • But transportation is characterized by varying load situations

• Such materials are characterized by complex, multiphase bonding zones

  • Damages can be complex and very difficult to be modeled

• For example, the failure mode of a failed structure is a result of the failure mechanisms leading to a propagating degradation of the structure.

Hybrid Materials

Damage Detection and Classification

In order to help identifying the failure mechanisms a non-destructive detection of the failure propagation in an early stage would significantly improve the understanding of the interrelation between failure mechanisms and failure modes.

• To enable the identification of a specific failure the correlation of sensor response and the failure propagation must be determined.

• This could be arranged by a systematic classification of the failure modes by means of materialographical analysis combined with clustering techniques of AI methods

• Clustering techniques require a proper, stable, and robust feture selction and sensor pre-processing!

The Multi-Agent System with Self-organization

Image Segmentation

Image segmentation is a method to divide an image in different regions (clusters) to identify regions of interest, i.e., isolating regions for further processing (feature extraction).

• Image segmentation

  • is feature selection;
  • is suitable for divide and conquer approaches;
  • but is sensitive to intensity variations and noise!

• In this work, one-dimensional vectors retrieved from time-resolved ultrasonic wave measurement are used for segmentation tasks.

Hybrid approach

• An adaptive multi-agent system is used to implement a self-organizing image segmentation

• Combined with Machine Learning to configure the agents

Image Segmentation

• The segmentation task splits irrelevant signal parts from relevant (ROI)

• The ROI extraction depends on the signal record, geometrical signal sender and receiver positions, signal quality (noise), and the probe geometry.

ROI Identification by MAS

• The Multi-agent System (MAS) consists of simple agents with different behaviour:

Master Agent

The master agent controls the divide-and-conquer process and instantiates segment agents.

• The master agent transforms the input signal vector to a segment vector of fixed length. Each time a new data set is loaded, the segment agents are notified.

Segment Agent

Each segment agent is responsible for one data cell and listen for data events.

• If an event was detected, an initial explorer agent is created.

• An explorer agent is created with a specific set of parameters, which can be adapted by the master agent and the segment agent.

ROI Identification by MAS

Explorer Agent

The explorer agent has the goal to collect data from the current left and right side neighbourhood within a given radius to make a feature decision:

• The neighbourhood data values are compared with the current associated data value,i.e.:
  Difference δ=|s(i±d)-s(i)| with d={-r,..,-1,1,..,r})

• Differences lying within a given interval δ ∈ Δ are counted.

• If the counter lies within another given interval set {h_min,..,h_max}, the explorer marks the cell and reproduces itself → reproduction & amplification.

• If the counter values is outside of the interval, it migrates (virtually) to another neighbour cell, performing the exploration again → diffusion

• If random walk is enabled, the diffusion and reproduction direction are chosen randomly, otherwise one more agent is instantiated on diffusion (opposite direction) and two agents are reproduced (moving in opposite directions).

ROI Identification by MAS

Automatic Parameter Selection

Parameter Sets

• Signal records from acoustic measurements can differ significantly with respect to amplitudes, the frequency spectrum, and noise.

• The feature selection MAS relies on parameter sets.

• Different signal records require different parameter sets for optimal ROI extraction and minimal computational costs.

• Machine Learning is used to select optimal parameter sets!

The initial high-dimensional sensor data record is down sampled. Relevant features are extracted from the original and down-sampled record to provide a signal characterization: Constant offset s₀ (filtered mean value); Standard deviation s₁; Peak amplitude (positive & negative) s₂,s₃; Frequency distribution ranges (f₁,f₂,f₃,f₄); and the Histogram distribution (h₁,h₂,h₃,h₄).

Automatic Parameter Selection

Ultrasonic Measurement and Feature Selection

The experimental Setup

• Plate of Hybrid Material (Aluminum + Composite)
• Two sections with changeable gap between
• One Actuator, one sensor (placed on section A or B)
• Ultrasonic round piezoelectric wafer active sensors (PWAS)

Measurement and Analysis

• Transmission and reflection signals were recorded and analyzed

• The quality of the automatic ROI extraction was evaluated with a quality parameter:

• with #roi: Number of ROIs detected,
• roiw: weight of ROI, (width, i.e., roi₁-roi₀), roi₀ and roi₁ are the start and end time of the detected ROI (w₀: expected weight),
• k: error weight, c₀: expected center position

Analysis Results

• The ROI extraction of reflected signal records achieves mostly a high accuracy and quality Q > 0.5

• The ROI extraction of transmitted signal records is more difficult due to a much lower signal-to-noise ratio and reflections at the boundaries, but still most ROIs can be identified correctly with Q > 0.5.

Conclusion

• Monitoring of mechanical structures is a Big Data challenge concerning Structural Health Monitoring and Non-destructive Testing

• Information mining (e.g. detection of damages) is two folded: Feature Selection and Feature Extraction

• Hybrid materials poses:

  • Complex models
  • Complicating damage detection and identification

• Automatic and adptive Feature Selection is required

• A hybrid approach using self-organizing agents and machine learning enables robust feature selection and damage detection

References

1. S. Bosse, D. Schmidt, M. Koerdt, Robust and Adaptive Signal Segmentation for Structural Monitoring Using Autonomous Agents, In Proceedings of the 4th Int. Electron. Conf. Sens. Appl., 15–30 November 2017; doi:10.3390/ecsa-4-04917

2. Stefan Bosse, Armin Lechleiter, A hybrid approach for Structural Monitoring with self-organizing multi-agent systems and inverse numerical methods in material-embedded sensor networks, Mechatronics, (2016), DOI:10.1016/j.mechatronics.2015.08.005