The researchers from the University of KwaZulu-Natal in South Africa developed a smart system for diagnostics of the ground brushes of the shaft in turbo-generators. These brushes are designed for drainage of hazardous currents emerging on the rotating shaft and for protecting the bearings from destruction. The proposed method is based on analyzing the electric signals with the help of neural networks: the system processes the data from the generator in real-time mode and identifies signs of wear-and-tear, contamination or contact losses. This allows for identifying the defects prior to them leading to serious breaks and shutdowns of the equipment.
Quite often modern turbo-generators face the problem of stray voltage on the shaft. It may be caused by remnant magnetism, magnetic field asymmetry or power electronics impact. Such voltage may get as high as dozens of Volts and generate currents going through bearings resulting in lubricant deterioration, local overheating and wear-and-tear. To prevent it the earthing brushes are installed on the shaft sending the current to the ground. However, these brushes get worn-out with time or get contaminated, or lose contact with the shaft. If these defects are not timely identified, the protection system stops working, and the bearings are exposed to hazards again. This is especially dangerous in case of the variable-load operation of the generator, when the loads and operating conditions change often.
As a rule, the brushes are monitored with the help of threshold signals: e.g., if the voltage on the shaft exceeds 5 V, the system produces an alarm. However, such an approach often does not work on time: the brush may be worn-out already, but there are no signals so far. On the contrary, the alarm may come due to transient noize and glitches not connected with real defects. Visual inspection may be an alternative, but it requires shutdown of the machine and access to the equipment, which may not be always possible and may be costly. Hence, smart monitoring becomes more and more relevant — when the system can analyze the form of the signal and may identify not only the fact of deviation, but its cause as well.
Two brushes are installed on the opposite ends of the shaft: one is metering the voltage (the norm is about 0.9 V), the second brush drains the current (about 3.3 А on average). The signals are read with high frequency — 10,000 times per second, divided into short signals and converted into spectrograms — visual images reflecting the change of signal frequency with time. Images for current and voltage are formed separately and then are sent to convolution neural network. This neural network is trained to differentiate four states: normal, wear-and-tear of the brush, loss of contact and contamination.
Both real data from turbo-generators of up to 846 MW capacity and archive data from power plants were used in the training process. Contaminated brushes (oil and pulverized coal), samples with less than 30% of residual operation life were used to model the defects, and contact was deliberately violated. The trained model demonstrated high accuracy: up to 99% when identifying contamination and wear-and-tear, 98% — identifying loss of contact, 95% — in case of failure of the main current collecting brush. At the same time there were no false faulty actuations with sound brushes.
The scientists plan to finalize the model implementing the residual operating life assessment and creating the conditions for planning maintenance in advance. To improve the accuracy, they are also exploring the possibility of using generative neural networks for creating additional databases with rare types of failures.
