Professional Documents
Culture Documents
13.lec 12 Transfer Learning 2
13.lec 12 Transfer Learning 2
Farzaneh Abdollahi
Winter 2024
DTL Solutions
Generalization Performance Improvement
Varying WCs
Different Machines
Partial-Doamin FD
Emerging FD
Compound Fault Decoupling
DTL Solutions
DTL Solutions
▶ DTL solutions are provided for four application scenarios:
1. Generalization performance improvement: In this scenario, the label
space of target domain is identical with the label space of the source
domain, that is, Y T ≡ Y S , which imposes strict restriction on the
fault types of domains and mainly focuses on improving the
generalization performance of DTL model under varying environments.
[1]
[1]
▶ Other Scenarios
▶ There are other challenges like types and sampling frequency of
sensors, different the number of training instances of each class, result
in imbalanced instances a huge distribution diversity between realistic
industrial data.
▶ Several approaches are introduced in different scanrios
▶ E.g. to deal with the problem where the instances of each fault class
are imbalanced during model training, by generating the minority
instances through the GAN, the instances can get balanced.[14]
DTL Solutions
▶ DTL solutions are provided for four application scenarios:
2. Partial-domain FD: the label space of target domain is a proper subset
of the label space of the source domain, Y T ⊂ Y S , which relaxes the
same label space requirement and mainly focuses on transferring
knowledge from a large- scale but redundant source domain to an
unknown small-scale target domain.
[1]
Partial-Domain FD
[1]
Partial-Domain FD
▶ Challenges:
▶ In target domain the information is not labeled.
▶ Annotating data is expensive and some times impassible
▶ Outlier source faults may lead to negative transfer.
▶ The large-scale but redundant source dataset is diverse enough to
subsume all fault classes of the small-scale target dataset.
▶ Directly transferring between the entire source and target domains as
the popular DTL methods is not an optimal and effective solution
▶ Solution: aligning the distribution of source and target domains to
positive transfer of the relevant data and to alleviate the negative
transfer of irrelevant data
▶ Model-based approaches inherently rely on the label information of
target instances, so they cannot help in these scenarios
Partial-Domain FD
Partial-Domain FD
DTL Solutions
[1]
Emerging FD
▶ Motivation: unpredicted faults are prone to occur since the machines
typically operate in complex and uncertain environments during
long-term service.
▶ Goal: Expanding diagnosis knowledge, by detecting the unknown
faults absent in the labeled source dataset and annotate them with
correct labels
[1]
Emerging FD
▶ Challenges:
▶ No knowledge about the new faults are available
▶ Separate the known and unknown fault classes in an unsupervised
manner is difficult.
▶ The emerging fault classes may also jeopardize the knowledge
alignment between the source and target domains due to the absence
of emerging faults in the source domain.
▶ negative transfer will happen if the distribution of the target domain is
directly matched with that of the whole source domain.
Emerging FD
Emerging FD
▶ E.g. Deep Adversarial Transfer Learning Network (DATLN) has two
components trained by adversarial training: [18]
▶ a feature extractor
▶ extracts features from input data, and the classifier outputs K + 1
dimension probability,
▶ K : the number of known faults in source domain and the (K + 1):
the probability of the unknown fault.
▶ a classifier
▶ build a decision boundary to recognize the unknown fault in the target
domain.
▶ limitation:they can only detect all unknown faults as one category
even if there exist multiple emerging faults
▶ More complex emerging FD when across machines and sensors are
added are still open to investigate
DTL Solutions
[1]
[1]
▶ Challenges:
▶ Sometimes the key parts are defected or even damaged.
▶ the fault characteristics of each component are coupled and exerted
influence reciprocally. Decoupling them is difficult.
▶ It is difficult and unrealistic to collect all types of compound fault data
in industrial cases.
▶ The traditional classifiers using the Softmax as the activation fcn. of
the last fully connected layer only outputs one label for a testing
instance, consider the compound fault as an independent pattern for
classification and the relationship between the compound fault and its
corresponding single faults is ignored.
References I
References II
References III
References IV
References V
References VI
References VII