ABSTRACT

This paper presents a new vibration-based hybrid anomaly detection model to prevent high-power generator failures in power plants. This idea was conceived as it causes large amounts of energy and economic losses once the generator remains in turn-off positions for a long time due to malfunctions. Furthermore, the failures in high-power generators, especially during peak load demanded time, can even cause systemic shortcomings. To overcome these limitations, multivariate linear regression, response surface methodology, and multi-layer perceptron-based hybrid anomaly detection model were designed to detect anomaly cases. The real-time vibration data collected by 20 distinctive sensors on powerful generators of each 360 MW from the Afşin-Elbistan B Thermal Power Plant were used. The hybrid model succeeded in determining the impact degrees of the sensors on the anomaly, and the number of effective sensors was reduced from 28 to 9. The most effective sensor for anomaly cases was determined as Pedastal 3 Vibration (x-axis), whereas the sensor with the most negligible impact was determined as Rotor 6 Vibration (y-axis). The model performance evaluation metric (R²) of the designed hybrid model was calculated over 0.92. The performance results were presented in graphs and tables.

摘要

本文提出了一种新的基于振动的混合异常检测模型，以防止发电厂中的大功率发电机故障。这个想法的构想是因为一旦发电机由于故障长时间保持在关闭位置，它会导致大量的能源和经济损失。此外，大功率发电机的故障，特别是在高峰负荷需求时间，甚至会导致系统缺陷。为了克服这些限制，设计了多元线性回归、响应面方法和基于多层感知器的混合异常检测模型来检测异常情况。使用了来自 Afşin-Elbistan B 热电厂的每台 360 MW 的强大发电机上的 20 个独特传感器收集的实时振动数据。混合模型成功确定了传感器对异常的影响程度，有效传感器的数量从28个减少到9个。异常情况下最有效的传感器确定为Pedastal 3 Vibration（x轴），而影响最可忽略不计的传感器被确定为转子 6 振动（y 轴）。模型性能评估指标设计的混合模型的(R ² )计算值超过 0.92。性能结果以图表形式呈现。