Research On Sensor Fault Diagnosis Method Of The Chiller Based On Deep Learning

Chillers are the key devices of the HVAC system,the sensor fault detection and diagnosis of the chiller is of great significance for ensuring the normal operation of the HVAC system.In recent years,many scholars have paid a lot of attention to the problems of the chiller sensor fault detection and diagnosis,and have made some process.However,the detection effect of deviation fault and gradual fault of the chiller sensor is not satisfied.RNN has problems of gradient explosion and gradient disappearance in dealing with nonlinear time problems,to solve this problem,Schmidhuber & Sepp Hochreite proposed a time recurrent neural network,Long Short-Term Memory(LSTM).It is widely used to process and predict high-dimensional,high-coupling,high-time correlation data.In view of the above problems,taking into account the characteristics of LSTM,this paper considered two aspects from classification and prediction and proposed a deep learning method based on improved LSTM for the sensor deviation fault and gradual fault detection of the chiller.In terms of deviation fault,this method treats sensor fault detection as a multi-classification problem,and different fault sensors correspond to different categories.This method combined sensor fault detection and diagnosis into one and can directly locate the fault sensor.The paper collected sensor data of the water-cooled chiller on site to train the improved LSTM network.It could be know through simulation experiments that the detection efficiency of different sensors is different,and the result was compared with the result of three other methods: Auto encoder,Principal Component Analysis(PCA)and standard LSTM.Finally,the detection efficiency of the method proposed in this paper is significantly better than the other three methods in sensor deviation fault detection of the water-cooled chiller.Furthermore,the detection efficiency of the proposed method has better symmetry for positive and negative fault levels with same absolute magnitude.For the gradual fault,considering from the point of prediction,each sensor readings are correlated,and each sensor reading can be predicted based on other sensor readings.This paper described the correlation between different sensors according to the Person similarity and selected the appropriate sensor reading as the input to the predicted sensor model,and used the predicted sensor reading as an output,and established an improved LSTM network prediction model for each sensor reading.Using the prediction error threshold as a criterion to judge whether there is a fault sensor.Finally,the detection results were compared with Principal Component Analysis(PCA),it could be know that the proposed method can diagnose the gradual fault of the sensor earlier.