Deep Semantic Canonical Correlation Embedding for Zero-Shot Industrial Process Fault Diagnosis

Fault diagnosis is crucial in ensuring the safety and efficient operation of industry processes. However, achieving excellent diagnosis performance in traditional fault diagnosis tasks becomes challenging when the training process lacks test fault information. To address this challenge, a novel model for zero-shot industrial process fault diagnosis based on deep semantic canonical correlation embedding (DSCCE) is proposed in this study. The DSCCE model utilizes an embedding reconstruction model to establish the relationship between fault and their semantic attributes, which is accomplished the task through the transfer of fault semantic attribute descriptions. Initially, DSCCE encodes fault samples and semantic attribute descriptions into an embedding space through two embedding networks. Then, the model employs canonical correlation analysis with two embedding features to assess the consistency between semantic attribute vectors and fault samples. Simultaneously, correlation constraints are applied between fault embedding feature subsets to enhance differential information. Subsequently, the classifier is trained using embedding features to complete the task of zero-shot fault diagnosis. Finally, the effectiveness and superiority of the proposed DSCCE model are verified using two case studies.