Efficient predictive maintenance systems can significantly decrease expenses in modern industrial plants. An illustrative example is a predictive maintenance technique based on estimation of fan mill impact plates state in thermal power plants. This can be performed using acoustic signals which are a natural choice but they are very sensitive to ambient noise. The article considers procedures for discarding the segments which contain different types of contamination. The algorithm is based on patterns derived from spectrogram representation of recorded signals and support vector machines as a classifier. The results are verified on real acoustic signals recorded in a vicinity of a fan mill in thermal power plant Kostolac A1 in Serbia.