A novel multi-sensor fault diagnosis framework based on PPS feature selection and K-means initialized particle swarm classifier using MiniRocket feature extraction

Abstract

Acoustic sensors and microphone array technologies are widely used in industrial condition monitoring and fault diagnosis, and their reliability directly affects monitoring stability and equipment safety. However, existing acoustic sensor fault diagnosis methods under complex operating conditions still face limitations, including insufficient feature representation, reliance on handcrafted features, high computational cost, and difficulty in real-time deployment. To address these issues, this paper proposes a lightweight acoustic sensor fault identification method for complex operating conditions, enabling prior hard-fault removal, fine-grained soft-fault classification, and efficient feature selection. First, an anomaly identification mechanism integrating statistical screening and adaptive decision-making is developed to rapidly detect and remove explicit hard-fault samples using multidimensional statistical features, ensuring reliable inputs for subsequent diagnosis. Second, MiniRocket is used to map the screened acoustic time-series signals into high-dimensional features, enhancing fault representation. Then, the Predictive Power Score is introduced to evaluate and select discriminative core features, reducing feature redundancy, inference latency, and computational burden. Finally, K-means-based initial partitioning is combined with a granular-ball classifier, where local geometric coverage and a relaxation factor enable robust soft-fault classification. The proposed method balances accuracy, robustness, interpretability, and efficiency, providing an effective solution for lightweight intelligent diagnosis of acoustic sensor faults in complex industrial environments.