Machine health monitoring plays a key role in the reliable and efficient operation of industrial processes by the provision of timely fault detection and prognostic supports for maintenance decisions. Health indicator (HI) construction provides a simple way to realize machine health monitoring. In recent years, convex weight learning data fusion methods have been preferred for incipient fault detection. However, the existing weight learning approaches treat HI construction and fault threshold determination as two separate tasks in that a HI is firstly established through optimizing an objective function and a HI threshold is then determined statistically. Without the joint determination of a HI and its detection threshold, these methods may not be sensitive enough for incipient faults. To solve this problem, this study proposes a new convex weight learning data fusion model that realizes the joint assessment of a HI and its fault threshold, providing a more sensitive interpretable-HI for incipient faults. To form a sensitive yet stable HI, a novel objective function and its constraints are put forward, and the optimized model weights are used to construct a monitoring HI and its threshold simultaneously. Besides, to handle non-Gaussian distributed data and improve model explainability, kernel density estimation (KDE) was adopted for computing detection thresholds. The proposed fault detection method requires no prior knowledge about the time of fault initiation and can be implemented online to localize informative frequency bands. Benchmarking against state-of-the-art methods, we used the proposed weight learning model to analyze both process and vibration data sets. We show that the proposed method greatly improves fault detection in terms of sensitivity and achieves a satisfactory rate of false alarms. We also show that the introduced model possesses interpretability when it is employed for fault detection based on vibration signals.Machine health monitoring plays a key role in the reliable and efficient operation of industrial processes by the provision of timely fault detection and prognostic supports for maintenance decisions. Health indicator (HI) construction provides a simple way to realize machine health monitoring. In recent years, convex weight learning data fusion methods have been preferred for incipient fault detection. However, the existing weight learning approaches treat HI construction and fault threshold determination as two separate tasks in that a HI is firstly established through optimizing an objective function and a HI threshold is then determined statistically. Without the joint determination of a HI and its detection threshold, these methods may not be sensitive enough for incipient faults. To solve this problem, this study proposes a new convex weight learning data fusion model that realizes the joint assessment of a HI and its fault threshold, providing a more sensitive interpretable-HI for incipient faults. To form a sensitive yet stable HI, a novel objective function and its constraints are put forward, and the optimized model weights are used to construct a monitoring HI and its threshold simultaneously. Besides, to handle non-Gaussian distributed data and improve model explainability, kernel density estimation (KDE) was adopted for computing detection thresholds. The proposed fault detection method requires no prior knowledge about the time of fault initiation and can be implemented online to localize informative frequency bands. Benchmarking against state-of-the-art methods, we used the proposed weight learning model to analyze both process and vibration data sets. We show that the proposed method greatly improves fault detection in terms of sensitivity and achieves a satisfactory rate of false alarms. We also show that the introduced model possesses interpretability when it is employed for fault detection based on vibration signals.Machine health monitoring plays a key role in the reliable and efficient operation of industrial processes by the provision of timely fault detection and prognostic supports for maintenance decisions. Health indicator (HI) construction provides a simple way to realize machine health monitoring. In recent years, convex weight learning data fusion methods have been preferred for incipient fault detection. However, the existing weight learning approaches treat HI construction and fault threshold determination as two separate tasks in that a HI is firstly established through optimizing an objective function and a HI threshold is then determined statistically. Without the joint determination of a HI and its detection threshold, these methods may not be sensitive enough for incipient faults. To solve this problem, this study proposes a new convex weight learning data fusion model that realizes the joint assessment of a HI and its fault threshold, providing a more sensitive interpretable-HI for incipient faults. To form a sensitive yet stable HI, a novel objective function and its constraints are put forward, and the optimized model weights are used to construct a monitoring HI and its threshold simultaneously. Besides, to handle non-Gaussian distributed data and improve model explainability, kernel density estimation (KDE) was adopted for computing detection thresholds. The proposed fault detection method requires no prior knowledge about the time of fault initiation and can be implemented online to localize informative frequency bands. Benchmarking against state-of-the-art methods, we used the proposed weight learning model to analyze both process and vibration data sets. We show that the proposed method greatly improves fault detection in terms of sensitivity and achieves a satisfactory rate of false alarms. We also show that the introduced model possesses interpretability when it is employed for fault detection based on vibration signals.

Machine health monitoring; Incipient fault detection; Health indicators; Convex-Optimization; Condition monitoring.