Abstract
Various types of low-level indexes and uncertainty of evaluation information due to quantitative index and qualitative index exist in converter transformer condition evaluation. To solve this problem, a novel converter transformer condition evaluation approach based on the belief rule base inference methodology and evidence reasoning (RIMER) is proposed. It firstly builds rule base with belief structure and transforms input information to a pre-defined form, then calculates the activation weight of input, after that using evidence reasoning for index aggregation to get the final evaluation distribution. At last, a converter transformer condition evaluation in the 500kv converter station is investigated to illustrate the usage of the methodology and its advantages.
Introduction
Converter transformers are critical nodes in High Voltage Direct Current (HVDC) systems, linking converter and inverter stations. Their failure can lead to significant system-wide outages. However, assessing their health is notoriously difficult because the data is heterogeneous: it includes precise quantitative measurements (e.g., insulation resistance) alongside subjective qualitative reports (e.g., "abnormal noise" or "history of defects"). Furthermore, sensor data can be noisy and expert opinions are often "fuzzy." Current methods often struggle to integrate these diverse information streams into a single, reliable evaluation.
Methodology
The proposed approach utilizes the RIMER (Rule-base Inference Methodology using Evidential Reasoning) framework, which consists of two main pillars:
- Belief Rule Base (BRB): Instead of traditional "If-Then" rules, BRB uses a belief structure. A rule might look like: "If insulation resistance is low, then the status is {Normal: 0.1, Warning: 0.3, Urgent: 0.6}." This allows the system to capture nuanced expert knowledge and uncertainty.
- Evidential Reasoning (ER) Algorithm: This is the inference engine that processes the rules. When new data arrives, the ER algorithm calculates how much each rule is "activated" and aggregates the results to produce a final probability distribution of the transformer's state.
The evaluation is structured hierarchically across three layers, as shown in Figure 1.
Expert knowledge is encoded into the system through a belief rule base. Table 1 illustrates how these rules are structured for the secondary index of electrical tests ($c_1$).
Case Study and Results
The methodology was validated using a real-world case study from a 500kV converter station. The data included historical maintenance records (2013–2015), preventive test data (e.g., insulation resistance >10,000 MΩ, oil breakdown voltage of 48 kV), and expert scoring. Membership functions were used to convert raw numbers into "degrees of confidence" (e.g., Good, Average, Bad).
The model produced the following evaluation for the overall transformer condition ($c_0$):
- Normal: 0.49
- Warning: 0.51
- Urgent: 0
The results indicated that while the transformer was generally safe, it was leaning toward a "Warning" state. This was driven primarily by high "Warning" scores in the electrical test ($c_1$) and operation condition ($c_4$) categories. This matched the actual physical reality: the transformer had a history of being overloaded twice, leading to partial aging of the internal insulation. The model correctly identified that while the unit could continue to run, it required close observation.
Conclusion
The RIMER-based approach effectively handles the inherent uncertainties and mixed data types in transformer health assessment. By providing a distributional output rather than a single score, it quantifies risk and uncertainty for maintenance teams. Future work will focus on automating rule base construction and optimizing premise weights using historical failure data.
Summary updated