An Interpretable Fault Detection Approach for Industrial Processes Based on Improved Autoencoder

Deep learning has recently emerged as a promising method for data-driven fault detection in industrial processes, especially autoencoders (AEs), which have achieved great detection performance. However, the AE models are essentially black boxes, which makes it difficult to interpret and trust the de...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on instrumentation and measurement Jg. 74; S. 1 - 13
Hauptverfasser: Ma, Zhen-Lei, Li, Xiao-Jian, Nian, Fu-Qiang
Format: Journal Article
Sprache:Englisch
Veröffentlicht: New York IEEE 2025
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Schlagworte:
Algorithms
Autoencoders
Decoding
Deep support vector data description (deep SVDD)
Encoding
Error detection
Explainable artificial intelligence
Fault detection
Fault diagnosis
Hyperspheres
improved autoencoder (AE)
Indexes
Industries
Machine learning
Mathematical models
memory module
Memory modules
Modules
Reconstruction
Representations
Vectors
ISSN:0018-9456, 1557-9662
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Deep learning has recently emerged as a promising method for data-driven fault detection in industrial processes, especially autoencoders (AEs), which have achieved great detection performance. However, the AE models are essentially black boxes, which makes it difficult to interpret and trust the detection results. In addition, most existing AE-based methods only take reconstruction error as the detection index. Due to the overgeneralization ability of AEs, fault data may also be reconstructed well, resulting in a high missed detection rate. To tackle these problems, an improved AE with a memory module and deep support vector data description (Deep SVDD) module is proposed. First, the AE is designed using the algorithm unrolling technique, which can be regarded as having a clear theoretical basis and an interpretable network architecture. Then, the latent vectors extracted by the encoder are used as queries to retrieve the most relevant items and obtain the estimated latent representations in the memory module, which can reduce the missed detection rate. Furthermore, the estimated latent representations are used for reconstruction in the decoder and mapped into a hypersphere in the Deep SVDD module, respectively. Thus, a combined detection index is established considering both the reconstruction error and the distance to the center of the hypersphere such that can further improve the detection performance. Moreover, an explainable artificial intelligence technique is introduced to measure the feature contributions for the detection results and further ensure the transparent decision-making process of fault detection. Finally, two examples are given to verify the effectiveness of the developed method.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0018-9456
1557-9662
DOI:10.1109/TIM.2025.3550600