Learning to better see the unseen: Broad-Deep Mixed Anti-Forgetting Framework for Incremental Zero-Shot Fault Diagnosis

TL;DR

This work tackles the problem of diagnosing unseen faults in dynamic industrial settings by introducing Incremental Zero-Shot Fault Diagnosis (IZSFD) and the Broad-Deep Mixed Anti-Forgetting Framework (BDMAFF). BDMAFF combines a deep generative feature memory with an attribute-prototype memory and memory-driven updates to a diagnosis model, enabling category and attribute increments without storing historical data. The approach introduces anti-forgetting losses for the generator and a memory-based update rule for the attribute prototypes, achieving superior performance on real hydraulic systems and the Tennessee-Eastman benchmark, especially in balancing seen and unseen fault recognition. The work advances fault diagnosis in evolving industrial environments by bridging zero-shot learning with incremental learning under data/storage constraints, offering practical impact for robust, scalable maintenance systems.

Abstract

Zero-shot fault diagnosis (ZSFD) is capable of identifying unseen faults via predicting fault attributes labeled by human experts. We first recognize the demand of ZSFD to deal with continuous changes in industrial processes, i.e., the model's ability to adapt to new fault categories and attributes while avoiding forgetting the diagnosis ability learned previously. To overcome the issue that the existing ZSFD paradigm cannot learn from evolving streams of training data in industrial scenarios, the incremental ZSFD (IZSFD) paradigm is proposed for the first time, which incorporates category increment and attribute increment for both traditional ZSFD and generalized ZSFD paradigms. To achieve IZSFD, we present a broad-deep mixed anti-forgetting framework (BDMAFF) that aims to learn from new fault categories and attributes. To tackle the issue of forgetting, BDMAFF effectively accumulates previously acquired knowledge from two perspectives: features and attribute prototypes. The feature memory is established through a deep generative model that employs anti-forgetting training strategies, ensuring the generation quality of historical categories is supervised and maintained. The diagnosis model SEEs the UNSEEN faults with the help of generated samples from the generative model. The attribute prototype memory is established through a diagnosis model inspired by the broad learning system. Unlike traditional incremental learning algorithms, BDMAFF introduces a memory-driven iterative update strategy for the diagnosis model, which allows the model to learn new faults and attributes without requiring the storage of all historical training samples. The effectiveness of the proposed method is verified by a real hydraulic system and the Tennessee-Eastman benchmark process.

Figures (6)

• Figure 1: Training set for category and attribute increment tasks. (a) For ZSFD, the training set only contains samples of seen faults (Fault #1, #2), along with a fault category-attribute matrix that describes attributes of seen faults (Fault #1, #2) and unseen faults (Fault #3, #4). (b) For category increment, only samples of new seen faults (Fault #5) are available for training. Additionally, the category-attribute matrix is expanded to include attributes of new seen faults (Fault #5) and new unseen faults (Fault #6). (c) For attribute increment, no training samples are available for model training, while the category-attribute matrix is expanded to include new fault attributes for existing faults.
• Figure 2: Overview of the broad-deep mixed anti-forgetting framework. (a) describe the training process of the generator. In each learning stage, the generator is trained with real features of fault categories corresponding to now stage and generated features of fault categories corresponding to historical stages. Discriminator, the attribute prototype memory and the feature prototypes are used to supervise the generation quality and mitigate the problem of forgetting. (b) describe the update of the attribute prototype matrix used in the diagnosis model. The attribute prototype matrix is extended when new attributes are added to describe faults, and it is updated using a memory-driven iterative update strategy with the help of the memory matrix to keep memory.
• Figure 3: The accuracy of seen faults from Stage 1 in different learning stages.
• Figure 4: The accuracy of unseen faults of Stage 1 in different stages
• Figure 5: The centers of generated features for the hydraulic system. (a) and (b) depict the generated features of Stage 1 and Stage 4, respectively, using the generative model trained with $\mathcal{L}_{\mathrm{wgan}}$. (c) and (d) represent the generated features of Stage 1 and Stage 4, respectively, using the generative model trained with $\mathcal{L}_{\mathrm{wgan}}$ and $\mathcal{L}_{\mathrm{\mathrm{anti-att}}}$. (e) and (f) display the generated features of Stage 1 and Stage 4, respectively, using the generative model trained with $\mathcal{L}_{\mathrm{wgan}}$, $\mathcal{L}_{\mathrm{\mathrm{anti-att}}}$, and $\mathcal{L}_{\mathrm{\mathrm{anti-fe}}}$.