A Deep Nonnegative Matrix Factorization Approach for Industrial Fault Diagnosis

Data-driven fault diagnosis represents a significant challenge in the realm of intelligent industry safety management. In practice, industrial data often encompasses a blend of various physical attributes, including the operating environment and fault location. Traditional models may struggle to enhance diagnostic performance due to their shallow architectures. To overcome this problem, we introduce Deep Nonnegative Matrix Factorization (DNMF), which employs a series of matrix factors to preserve the desirable structure of the process data. Specifically, DNMF systematically decomposes data into multiple hierarchical levels, allowing for the extraction of implicit features of industrial data. Our experimental results demonstrate that the DNMF model outperforms both shallow and deep methods while also requiring less computational time compared to deep structures.