Abstract
Aiming at the problem of abnormal data generated by a power transformer on-line monitoring system due to the influences of transformer operation state change, external environmental interference, communication interruption, and other factors, a method of anomaly recognition and differentiation for monitoring data was proposed. Firstly, the empirical wavelet transform (EWT) and the autoregressive integrated moving average (ARIMA) model were used for time series modelling of monitoring data to obtain the residual sequence reflecting the anomaly monitoring data value, and then the isolation forest algorithm was used to identify the abnormal information, and the monitoring sequence was segmented according to the recognition results. Secondly, the segmented sequence was symbolised by the improved multi-dimensional SAX vector representation method, and the assessment of the anomaly pattern was made by calculating the similarity score of the adjacent symbol vectors, and the monitoring sequence correlation was further used to verify the assessment. Finally, the case study result shows that the proposed method can reliably recognise abnormal data and accurately distinguish between invalid and valid anomaly patterns.
Introduction
Power transformers are the heart of the electrical grid. To prevent catastrophic failures, utilities use online monitoring systems to track parameters like oil temperature and dissolved gas concentrations. However, this data is notoriously "noisy" due to sensor malfunctions, communication drops, or external interference.
A critical gap in current data cleaning is that most algorithms treat all outliers the same. In reality, anomalies fall into two categories:
- Invalid Anomalies: Caused by technical errors. These should be filtered out to ensure accurate state predictions.
- Valid Anomalies: Caused by actual equipment faults. These contain vital information and must be preserved for health assessments.
Methodology
The proposed framework utilizes a two-stage technical approach to ensure data integrity.
1. Anomaly Recognition
The process begins with Empirical Wavelet Transform (EWT) to decompose non-stationary monitoring signals into different frequency components. Each component is then modeled using an ARIMA (Autoregressive Integrated Moving Average) model to predict the next value.
By subtracting predicted values from actual measurements, a "residual sequence" is created that strips away normal trends. The Isolation Forest (iForest) algorithm is then applied to these residuals to pinpoint exactly when an anomaly occurs, segmenting the time series for further analysis.
2. Pattern Differentiation
To distinguish between valid and invalid anomalies, the authors use an Improved Multi-dimensional SAX (Symbolic Aggregate Approximation). This version creates a 3D vector representing the mean, slope, and sample entropy of a segment. By calculating similarity scores between adjacent symbol vectors, the system determines if the signal's "character" has fundamentally changed (valid fault) or if it is a momentary technical blip (invalid noise).
Experimental Setup and Results
The study validated the method using real-world online monitoring data from a 500-kV transformer substation. The dataset included seven indicators sampled at 4-hour intervals over several months.
Key results include:
- Accuracy: The model correctly identified synthetic noise as "invalid" and real-world gas concentration spikes as "valid" anomalies, the latter being confirmed as internal hardware faults.
- Efficiency: The symbolic approach was significantly faster than direct numerical calculation, requiring only 5.27% to 10.55% of the time taken by traditional methods.
- Correlation Verification: Using Grey Correlation Analysis to cross-reference related sensors (e.g., checking if a spike in ethylene is reflected in methane levels) provided a vital "sanity check" against sensor noise.
Conclusion
By focusing on data morphology (shape and entropy) rather than just raw values, this framework effectively distinguishes between sensor errors and looming equipment failures. Future work will focus on automating threshold settings and testing the framework on high-frequency transient data.
Summary updated