A Physics-informed, Transfer Learning Approach to Structural

Health Monitoring
Trent Furlong1, Karl Reichard2

1,2
The Pennsylvania State University, University Park, PA, 16802, USA
tsf44@psu.edu
kmr5@psu.edu

ABSTRACT transfer learning to bridge the connections between the

untrained layers to the physics-informed layer using
One of the main challenges for structural health monitoring
experimental data from more complex structures. We will
(SHM) is a lack of failure data to make accurate health
verify our model using publically available SHM datasets
predictions. Obtaining desirable failure data is generally very
used in a variety of past literature experiments.
expensive, given the required testing needed to measure all
types of system failures, which may be unfeasible in many
1. PROBLEM STATEMENT
health monitoring applications. Machine learning has helped
to improve health monitoring performance but is still limited One of the main problems in any sort of prognostics-based
by the availability, relevance, and quality of the training data. study is having enough failure data [Bull, Worden, Manson,
This data dependence impedes data-driven models from and Dervilis (2018); Gardner, Lord, and Barthorpe (2018);
generalizing to unseen data, which is problematic for datasets Fuentes, Cross, Gardner, Bull, Rogers, et al (2020)]. The
lacking failure data. Physics-driven models, like finite- large imbalance between “healthy” and “failure” data
element models, are powerful tools for predicting structural increases the difficulty for models to classify and/or localize
responses when the governing physics are not clearly defined. damage, as well as the severity of damage. Desired damage
These models can generate simulated fault data to address the states often have to be manually added to the host structure
data limitation without having to physically damage a in order to accelerate the failure testing, since some structures
structure, but are computationally expensive and susceptible require significant wear and tear before actual damage may
to modeling errors that can prevent the data from being occur “naturally”. Testing using induced damage may not be
statistically comparable to experimental data. feasible if the structure is still intended for use.
A new trend has been to develop physics-guided machine Structural health monitoring (SHM) is “any automated
learning models (PGML), a hybridization of the two monitoring practice that seeks to assess the condition or
aforementioned models that have been shown to improve health of a structure” [Fuentes et al. (2020)]. Structural health
generalization of, and even outperform, pure data-driven monitoring differs from other non-destructive
models while using less training data. These PGML models testing/evaluation (NDT/NDE) techniques in that the sensors
can take many forms, but generally embed some form of are permanently installed and the monitoring system is on-
physics into a data-driven model as physically relevant line performing continuous analysis [Fuentes et al. (2020)].
constraints. Our research plan is to utilize PGML to improve The primary questions an SHM system attempts to answer
neural network capabilities to predict structural damage. The concern the existence, location, type, extent, and prognosis of
proposed PGML model will follow a neural network damage in the system [Rytter (1993); Farrar & Worden
architecture found in related literature consisting of feature (2012)]. Machine learning (ML) has become a popular
extraction, physics-informed, and label prediction layers. The method for answer these questions given its great ability of
physics-informed layer will consist of an aggregate of sub- finding patterns within a large amount of data.
networks trained from simplified structure models which
Two main types of SHM techniques include data-driven and
have known governing equations and can be used to generate
simulated training data. The full PGML model will use physics-driven methods. Data-driven methods, such as ML,
generally require a large amount of data in order to learn the
_____________________
statistically relevant relationship between input features and
Furlong, T., et al. This is an open-access article distributed under the terms
of the Creative Commons Attribution 3.0 United States License, which the structural health state. Generally speaking, the greater the
permits unrestricted use, distribution, and reproduction in any medium, representation of health states to structural responses within
provided the original author and source are credited. the data, the greater the reliability of the model. This is

International Journal of Prognostics and Health Management, ISSN 2153-2648, 2012 *** 1
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

because a data-driven method can only make predictions machine learning (PIML), or physics-guided machine
from what it has seen from training data, making these learning (PGML), is a hybrid modeling approach that has
models difficult to generalize to structures that do not respond become increasingly popular since its initial online
the same as the training data, which can occur even for publication dating back to 2016 (see Figure 1). Karniadakis,
nominally identical structures [Fuentes et al (2020)]. Kevrekidis, Lu, Perdikaris, Wang, and Yang (2021) define
Labeling the data is also difficult and/or costly to perform PIML as a method that “integrates seamlessly data and
[Bull et al (2018)] and the existence of labels (or lack thereof) mathematical physics models, even in partially understood,
will determine whether the ML algorithm uses a supervised, uncertain and high-dimensional contexts.” The purpose for
unsupervised, or semi-supervised learning approach. These using a physics-informed model is to better constrain how a
data-driven models are considered “black-box” models due data-driven model learns from the data by teaching it known
to the decision-making process being relatively unintelligible information about the system being analyzed [Zhang, Liu,
to humans [Bull et al (2018); Cross, Gibson, Jones, and Sun (2020a)]. Known advantages for using PIML include
Pitchforth, Zhang, and Rogers (2022)] and as such can hinder a lower training cost [Yu, Yao, and Liu (2020)], requires less
the confidence operators have with the model’s predictions. training data, [Zhang et al (2020a), Zhang, Liu, and Sun
(2020b)], has improved generalizability [Zhang et al (2020a);
Physics-driven methods for SHM usually rely upon finite-
Yu et al (2020)], and generally outperforms data-driven
element (FE) modeling and analysis to provide constrained
data that is relevant to the observed structure. Two different
analysis techniques using FE models include inverse model-
driven methods and forward model-driven methods. Inverse
model-driven methods update the FE model by comparing
expected results to the actual measured results. However, this
model updating may be ill conditioned since unique, stable
solutions are not always feasible, and an adequate
interpretation of these updated parameters may not be
feasible when making a decision about the state of the
structure [Fuentes et al (2020)]. An additional challenge to
physics-driven methods is that fault mechanisms must be
included in the FE model to obtain predicted fault response
data, which fault mechanisms may not always be clearly
defined and/or could result in predicted fault response data
that does not match measured data.
Forward model-driven SHM techniques use the FE model to Figure 1. Number of publications (title and abstract) on
generate a training dataset from operational conditions that “physics informed machine learning” or “physics
are used in a supervised learning approach in an attempt to informed neural networks” from 2016 to 2022 (Source:
build machine-learning models and address the problem of app.dimensions.ai)
not having sufficient fault data [Fuentes et al (2020); Gardner
models [Yu et al (2020); Zhang et al (2020b)]. The added
et al (2018); and Balthorpe (2010)]. However, these methods
physics to data-driven models makes them gray-box models,
require that the generated dataset give statistically significant
which improves the human interpretability and confidence in
results that are consistent with measurements obtained from
the model’s predictions.
the actual structure, and require calibration and verification
from real world data, which may not be feasible [Fuentes et
2. NOVEL CONTRIBUTIONS
al (2020)].
Many of the PGML methods within SHM make use of FE
Additional challenges with FE models include inherent
models to either generate simulated data for different damage
modeling discrepancies due to structural complexity or lack
states or perform model updating to compare predicted
of knowledge about certain material properties that may be
results with measured results for anomaly detection.
simplified to perform the desired analysis [Farrar & Worden
However, FE models are generally computationally
(2012); Ozdagli & Koutsoukos (2021)]. Another challenge is
expensive that increases with the complexity of the model.
computational expense [Gardner et al (2018)] as well as
Additionally, FE models can have intrinsic modeling
efficiency from high fidelity models due to having a large
discrepancies (say from not knowing the exact material
number of parameters required to build the model [Fuentes et
properties for the model) which can similarly result in
al (2020)].
differences between the simulated and experimental data
A new on-going research strategy is developing hybrid-based [Ozdagli & Koutsoukos (2021)]. While PGML has been
models that use both data-driven and physics-driven methods shown to improve generalization even when a FE model has
to overcome their respective challenges. Physics-informed known modeling errors [see Ozdagli & Koutsoukos (2021)],

2
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

the question remains can you still get comparable results To model more complicated structures where the physics are
without the need for a computationally expensive FE model? not so clearly defined, we will add neural network layers
around this physics-informed layer, where the layers
The reason complex FE models are used in SHM is because
the physics for a complex structure may not be well-defined
or intrinsically deterministic (i.e., an ODE/PDE is not
defined). However, the individual components that make up
the structure have more defined physics that when put
together may form a closer approximation to the ground truth.
We hope to use PGML models to act as surrogate FE models
at the individual structural component level, where the
physics are more clearly defined. The model would initially
consist of neural network layers trained for simple structural
components using PGML. Then, through the principle of
transfer learning, we can model a more complex structure by
building a neural networked that aggregates these physics-
trained layers. We expect this will improve the
generalizability of SHM models while using less training data
Figure 2. Proposed physics-guided machine learning model
and without a complex FE model. This trained PGML is not
architecture for a) training and aggregating physics-informed
expected to take the place of a FE model, but should ideally
neural networks from simplified physical systems and b)
be compatible with complex FE models for determining the
performing transfer learning with the pre-trained physics-
health of a structure.
informed layer to train a model for a more complex physical
structure.
3. PROPOSED RESEARCH PLAN
Our proposed research plan is to iteratively build a physics- preceding the physics-informed layer act as feature extractors
informed neural network (see Figure 2) that can model and the layers following act as health label predictors (see
increasingly complex structures by: Figure 2b). This model architecture was implemented by
a) individually training a collection of independent Ozdagli & Koutsoukos (2021) and showed promising results,
physics-informed networks using simplified with the difference being their implementation used outputs
physics-based models to generate their respective from a FE model to act as the physics-informed layer instead
training data, of our proposed collection of sub-networks. The new model
will be trained using data collected from the complex
b) aggregating the trained physics-informed sub- structure in the form of transfer learning to bridge the
networks to form a single, physic-informed layer as connections between the added layers and the physics-
part of a larger neural network, and informed layer. The term “transfer learning” within the
c) using transfer learning to train and connect neural context of this study means the process of connecting new,
network layers around the physics-informed layer untrained model layers to pre-trained model layers (i.e.,
using data from the complex physical structure. connecting the known with the unknown).
The physics-informed sub-networks can be trained Freezing the pre-trained layers’ parameters during the
individually using simulated data derived from their training process reduces the model’s overall number of
respective physics-based model(s) and can be verified using parameters that need to be trained for the complex structure.
appropriate experimental data (when available). The This process provides the added benefit of reducing the total
individual performance of the sub-networks will be evaluated number of trainable parameters for a potentially deep
using both physics-guided neural networks and physics- network, preventing the need to train a blank network of
agnostic (i.e., purely data-driven, or vanilla) neural networks equal size and depth. This model type also allows for models
to test the validity of using PGML methods with their to be built incrementally for increasing complex structures,
expected benefits. Additionally, we will evaluate how much starting from an elemental level, to a structural sub-element
simulated data is required for accurate performance for each level, to then a full structural element. These tiers of models
sub-network, and gauge how that performance scales when may then serve as physics-informed building blocks for other
linked to more complex networks, where the physics is not so structures that may contain similar structural components.
clearly defined. It is expected that the physics-informed layer To get relevant experimental data, a number of datasets
(i.e., the aggregated trained physics-informed sub-networks) related to SHM are available online, each ranging in different
is sufficient to evaluate simple structures with clearly defined structural complexity and measurements. Some examples
physics (e.g., ODEs/PDEs, equations of motion, etc.). include datasets from the Los Alamos National Laboratory

3
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

SHM Data Sets and Software website containing vibration Identification. Handbook of Experimental Structural
data for vibration-based SHM approaches. We plan to also Dynamics, 1-72.
make extended measurements for a more complex structure Gardner, P., Lord, C., & Barthorpe, R. (2018). A probabilistic
(e.g., hydro-turbine blade) starting with a coupon test, then framework for forward model-driven SHM. In
graduating to a substructure testing, and eventually a full- Proceedings of the 9th European Workshop on
scale structure test. Structural Health Monitoring (EWSHM 2018), (Nov.
11), NDT.net.
For future work, a FE model could additionally be used to
Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P.,
generate training data from increasingly complex structures
Wang, S., & Yang, L. (2021). Physics-informed machine
as part of the transfer learning process. This would allow for
learning. Nature Reviews Physics, 3(6), 422-440.
a FE model to still be used with this neural network
Los Alamos National Laboratory. SHM Data Sets and
architecture that may potentially improve the transfer
Software, www.lanl.gov/projects/national-security-
learning training process. Additional benefits would need to
education-center/engineering/software/shm-data-sets-
be explored.
and-software.php.
Ozdagli, A. I., & Koutsoukos, X. (2021). Model-based
REFERENCES damage detection through physics guided learning.
Barthorpe, R. J., (2010). On model- and data-based In Annual Conference of the PHM Society (Vol. 13, No.
approaches to structural health monitoring. PhD thesis, 1).
University of Sheffield. Rytter, A., (1993). Vibrational based inspection of civil
Bull, L., Worden, K., Manson, G., & Dervilis, N. (2018). engineering structures. PhD thesis, Department of
Active learning for semi-supervised structural health Building Technology and Structural Engineering,
monitoring. Journal of Sound and Vibration, 437, 373- Aalborg University.
388. Yu, Y., Yao, H., & Liu, Y. (2020). Structural dynamics
Cross, E. J., Gibson, S. J., Jones, M. R., Pitchforth, D. J., simulation using a novel physics-guided machine
Zhang, S., & Rogers, T. J. (2022). Physics-informed learning method. Engineering Applications of Artificial
machine learning for structural health Intelligence, 96, 103947.
monitoring. Structural Health Monitoring Based on Zhang R., Liu, Y., & Sun, H. (2020a). Physics-guided
Data Science Techniques, 347-367. convolutional neural network (PhyCNN) for data-driven
Farrar, C. R., & Worden, K. (2012). Structural health seismic response modeling. Engineering
monitoring: a machine learning perspective. John Wiley Structures, 215, 110704.
& Sons. Zhang, R., Liu, Y., & Sun, H. (2020b). Physics-informed
Fuentes, R., Cross, E. J., Gardner, P. A., Bull, L. A., Rogers, multi-LSTM networks for metamodeling of nonlinear
T. J., Barthorpe, R. J., … & Worden, K. (2020). structures. Computer Methods in Applied Mechanics and
Structural Health Monitoring and Damage Engineering, 369, 113226.