A Deep Learning Approach for Fusing Sensor Data from Screw Compressors

Chillers are commonly used for thermal regulation to maintain indoor comfort in medium and large buildings. However, inefficiencies in this process produce significant losses, and optimization tasks are limited because of accessibility to the system. Data analysis techniques transform measurements coming from several sensors into useful information. Recent deep learning approaches have achieved excellent results in many applications. These techniques can be used for computing new data representations that provide comprehensive information from the device. This allows real-time monitoring, where information can be checked with current working operation to detect any type of anomaly in the process. In this work, a model based on a 1D convolutional neural network is proposed for fusing data in order to predict four different control stages of a screw compressor in a chiller. The evaluation of the method was performed using real data from a chiller in a hospital building. Results show a satisfactory performance and acceptable training time in comparison with other recent methods. In addition, the model is capable of predicting control states of other screw compressors different than the one used in the training. Furthermore, two failure cases are simulated, providing an early alarm detection when a continuous wrong classification is performed by the model.