关键词: 设备故障, 故障诊断, 风险评估, LSTM 网络, 特征提取

Abstract: Duing to the large number of working parameters，complex fault kinds，and difficulties in evaluating faults，a Fault Diagnosis and Risk Evaluation （FDRE）for air-handling units was proposed based on Long Short-Term Memory （LSTM）network algorithm. Firstly，a new definition of fault scenario was given for feature extraction of air-handling units， which included multi-dimensional monitoring variables of air-handling units，and the changes of environmental，energy factors and three kinds of fault risks as well. Then，predict networks with various parameters were built and proved capable of analyzing the time series characteristics of monitor data，which could not only diagnose the specific causes of the fault，but also evaluate the risk of the fault before the fault occurs. Then the network was utilized in a large public hospital. Comparing with other series data-based neural network algorithms，the LSTM network with a time window of 35 was better in accuracy and stability for fault diagnosis. Practical engineering applications have validated that extracted features can reflect the mechanism of air-handling units，the warnings are reasonable for use，which can assist effectively the maintenance work of air-handling units on site.

Key words: equipment fault, fault diagnosis, risk evaluation, LSTM networks, feature extraction