Early fault identification method for coal mill in thermal power units based on machine acoustic signature and deep neural network fusion

Abstract

As one of the most safety- and efficiency-critical auxiliary machines in thermal power units, the running condition of a coal mill has a direct bearing on the stability and economy of the whole unit. Conventional diagnostic schemes, however, tend to respond too late and to raise too many false alarms. To address these shortcomings, this study develops an early fault identification scheme in which machine acoustic-signature cues are jointly modeled with a deep neural network. Acoustic and vibration data covering both healthy and faulty states are gathered through sensors mounted at representative measuring points on the mill. Two complementary descriptors—Mel-frequency cepstral coefficients (MFCC) together with wavelet-packet energy features—are derived to characterise the acoustic signature, after which a hybrid CNN-LSTM network is built to learn and recognise the underlying fault patterns. For better adaptability in the field, an anti-noise front end and a load-adaptive normalisation stage are incorporated so that the method remains reliable under heavy industrial noise and fluctuating load. Field experiments show that the proposed scheme reaches 95.3% recognition accuracy for representative faults including bearing wear, pulverized-coal blockage and roller cracking, and is able to issue a warning 3-7 days ahead of failure, thereby giving solid support to condition-based preventive maintenance.