IEEE Robotics & Automation Magazine - June 2023 - 12

processes, the production lines often
vary due to different production requirements
and it is inevitable that new types
of defects could appear. Thus, it is difficult
for the old model to adapt to all of
these new scenarios properly. To solve
this problem, we can either train several
models for each new independent defect
or train a generalized model for all defect
types. However, the former approach may
generate too many models and require a
huge deployment load in production lines.
Moreover, for some kinds of defects, the
number of data are scarce, which brings
difficulties to train an effective model.
For the latter approach, it requires to train
the model with all the data every time a
new type of defect arrives, which increases
not only the difficulty of the training
as the data increases, but also requires
extra memory resources for the data storage.
Therefore, it is necessary to develop
a lifelong learning defect recognition
framework to continuously update the old
model to adapt to new industrial scenarios. This also meets the
trend of moving toward the generalization of the production
line mentioned in Industry 4.0, where the adaptation ability of
the model is highly required [1].
While lifelong learning methods are helpful to handle
situations where different industrial scenarios arrive sequentially,
there is still a great challenge in lifelong learning
called the catastrophic forgetting problem [2], in which the
performance of the model in one scenario will drop drastically
after the model is trained for another scenario. As a
result, the updated model tends to be more suitable for new
scenarios rather than old ones (Figure 2). In this article, we
propose an industrial defect classification framework based
on lifelong learning, in which an RGO lifelong learning
method is adopted to train the defect classification model
"
IT UTILIZES THE
ANNOTATED DATA
COLLECTED FROM
PRODUCTION LINES
TO TRAIN A DEFECT
RECOGNITION MODEL,
AND IT HAS BEEN
GRADUALLY REPLACING
THE TRADITIONAL
MANUAL DEFECT
INSPECTION THAT
SUFFERS FROM HUGE
COST OF LABOR AND
TIME.
„
RELATED WORK
DEFECTION INSPECTION
The vision-based automatic defection inspection system has
been widely used in practical industrial manufacturing due to
its efficiency and accuracy. The system automatically extracts
features from the captured images and a following classifier is
used to classify defects. In the early days, handcrafted features
and traditional machine learning classifiers were usually combined
together for defect inspection. In [3], various color features
are extracted from the feature extraction region
optimized by genetic algorithm, and then the support vector
according to the characteristics of defect
data. The proposed framework is able to
obviously relieve the catastrophic forgetting
problem of lifelong learning. The
main contributions of this are are concluded
as follows:
■ A defect classification framework
based on lifelong learning is proposed,
which can recognize various
industrial defects in different scenarios
continuously. To the best of our
knowledge, it is the very first work
that introduces the lifelong learning
into the industrial defect inspection.
■ A novel RGO lifelong learning method
is used to train the defect classification
model, which only needs a
fixed network capacity and does not
need the defect data in previous scenarios
for training.
■ The proposed framework is evaluated
on an experimental setup of six defect
classification tasks, where the data sets
are developed by ourselves. Extensive
experiments are conducted demonstrating the effectiveness
of the proposed framework.
Industrial Production Line
Defect Inspection System
FIGURE 1. A representative machine defect inspection system in industrial production lines.
12 IEEE ROBOTICS & AUTOMATION MAGAZINE JUNE 2023

IEEE Robotics & Automation Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2023