TL;DR
This paper introduces a novel three-level fuzzy comprehensive assessment (TLFCA) combined with a back-propagation neural network (BPNN) to assess the health degree (HD) of mine hoist braking systems. This approach offers a more nuanced and reliable method for intelligent maintenance and prognostics health management (PHM) compared to traditional threshold-based assessments.
What Problem Does This Paper Address?
Mine hoists are critical for underground operations, and their braking systems are the last line of defense against catastrophic failures. Accidents caused by braking system faults account for over 60% of all hoist accidents, highlighting the urgent need for improved safety and reliability. Existing health assessment methods often rely on simple threshold or limit assessments for individual parameters, lacking a comprehensive understanding of the system's overall health. This paper aims to bridge this gap by developing a method that integrates multiple sensor data points and considers the complex interactions within the braking system to provide a more accurate health degree.
Methods
The core of the proposed method is a hierarchical approach combining fuzzy logic and neural networks:
- Three-Level Fuzzy Comprehensive Assessment (TLFCA):
- Sensor Level: Raw sensor data from multiple time points are fused using multi-time fusion. Fuzzy comprehensive assessment (FCA) is then applied to these fused values to obtain fuzzy comprehensive assessment values (FCAVs) for each sensor's health condition.
- Subsystem Level: FCAVs from multiple sensors within a subsystem (e.g., brake, brake disc, hydraulic system) are fused using multi-sensor fusion to derive subsystem-level FCAVs.
- System Level: FCAVs from all subsystems are fused using multi-subsystem fusion to obtain the overall braking system's FCAVs.
- Back-Propagation Neural Network (BPNN): The FCAVs generated at all levels (sensor, subsystem, system) are fed into a pre-trained BPNN. This network acts as a defuzzification layer, transforming the fuzzy assessment values into a crisp Health Degree (HD) score, ranging from 0 to 1.
- Data Standardization: Monitored values are standardized into the [0,1] range using benefit, cost, and interval type formulas, accounting for different data characteristics.
- Fuzzy Membership Functions: Normal distribution functions are used to map standardized data to membership degrees for four health states: Health State (HS), Sub-Health (SH), Critical Fault (CF), and Fault State (FS).
- Weight Determination: A hybrid approach combining expert scoring, objective weight determination (based on deviation from expected values), and the Analytic Hierarchy Process (AHP) is used to assign weights to sensors and subsystems, ensuring objectivity and scientific rigor.
Datasets
The paper utilizes data from a test bed designed to simulate an ultra-deep well hoist braking system. This test bed is a scaled-down version (0.1 of actual hoist) equipped with sensors for brake clearance, brake disc temperature, deflection, motor current, oil pressure, and oil temperature. The paper presents a table of "on-line monitored values" (Table 3) from this test bed, which are used for standardization and subsequent calculations.
Key Insights and Ideas
- Hierarchical Fusion for Complexity: The TLFCA approach effectively breaks down the complex braking system into manageable levels (sensor, subsystem, system), allowing for a more systematic and comprehensive assessment.
- Fuzzy Logic for Ambiguity: Fuzzy logic is well-suited for handling the inherent uncertainty and vagueness in sensor data and system health, providing a more robust assessment than crisp logic.
- BPNN for Defuzzification and HD Calculation: Using a BPNN to translate FCAVs into a single HD score provides a powerful mechanism for quantifying overall system health and allows for learning complex relationships between fuzzy inputs and the desired output.
- Hybrid Weighting for Robustness: Combining expert knowledge with objective data analysis in weight determination leads to more reliable and impartial assessment results.
- HD as a Continuous Metric: The HD, a value between 0 and 1, offers a more granular and informative measure of health than discrete categories, enabling better decision-making for maintenance.
Key Results
The proposed method was validated using data from the test bed. Key findings include:
- Sensor-Level Assessment: The method accurately identified different health states for individual sensors. For example, BC 1-1 had an HD of 0.7219 (SH), while BC 4-1, showing signs of exceeding acceptable bounds, had an HD of 0.1371 (FS). This aligns with the expected behavior where lower HD indicates a higher likelihood of fault.
- Subsystem and System-Level Assessment: The method successfully assessed the health of subsystems and the overall braking system. In a scenario where BC 4-1 failed (FS), the brake subsystem and the entire braking system were correctly classified as FS with an HD of 0.1371.
- Sensitivity Analysis: The study demonstrated the method's sensitivity. As BC 4-1 data changed from near-normal to fault conditions, the HD of the sensor, subsystem, and system decreased accordingly, showing the method's ability to track degradation. For instance, as BC 4-1 transitioned from HS to FS, its HD dropped from 0.9803 to 0.1371.
Limitations and Future Work
- Data Completeness: The paper acknowledges that the monitored parameters must "reflect the performance of the equipment completely." Obtaining perfect and accurate monitored data is a prerequisite for the method's effectiveness.
- Standardization Parameter Selection: The selection of data limits for standardization requires relevant regulations and expert knowledge.
- Computational Demands: The method requires high computing speed and may necessitate parallel operation for complex systems.
- Future Work: The authors suggest that the calculated HD can be used to draw HD radar charts and trend charts for sensors, subsystems, and the entire system, providing a visual basis for operation, maintenance, and management. The method can also be extended for residual life prediction and serve as a reference for assessing the health of entire hoists or other complex equipment.
Summary updated