Research On HVAC Fault Diagnosis Method Based On Multidimensional Taylor Networ

The energy consumption of Heating,Ventilation and Air Conditioning(HVAC)systems is a major part of the energy consumption of commercial buildings.The Air Handling Unit(AHU)is the key component of HVAC,and once a fault occurs,it will affect indoor comfort and air quality,while also causing a significant increase in energy consumption.Therefore,investigating AHU fault diagnosis has important engineering application value.Currently,the fault diagnosis methods for HVAC systems are generally divided into three categories: modelbased methods,knowledge-based methods,and data-driven methods.Among them,the first two methods face difficulties in model development,poor portability,and time-consuming creation of process knowledge bases.In recent years,various neural networks based on datadriven methods have been widely used for AHU fault diagnosis.Although these methods have high diagnostic accuracy,their network structure is complex,leading to large computational requirements and poor real-time performance.By contrast,the Multi-dimensional Taylor Network(MTN)is a new type of nonlinear network that uses a polynomial network to approximate nonlinear functions.It has a strong nonlinear fitting ability,a simple network structure,and good real-time performance.Therefore,this paper proposes a fault diagnosis method based on the MTN for AHU fault diagnosis,and the work is introduced as follows:First,an analysis of the theoretical foundation of the MTN was conducted and its network structure was discussed.A detailed comparison was then made between its structure and that of traditional neural networks,and the research showed that MTNs have the advantages of fewer parameters,simpler structure,and better real-time performance.Secondly,in response to the common problem that traditional deep learning-based fault diagnosis methods lack a clear explanation of the fault mechanism and require high-quality labeled data,this paper proposes a novel fault diagnosis method for the AHU system based on a Back Propagation MTN(BP-MTN)fitting model and a new dynamic Statistical Process Control(SPC).The method includes the following features: 1)adding activation functions and fully connected layers after the output of the MTN to construct a BP-MTN fitting model for estimating the fault variables;2)proposing a new method of SPC and establishing a Dmatrix to achieve fault isolation,which improved fault diagnosis accuracy.The effectiveness of this method is verified through experiments.Finally,in response to the current problem that the MTN fitting model cannot be used alone for fault diagnosis,this paper proposes a BP-MTN classifier for fault diagnosis.Its features include: 1)adding a softmax layer after the fully connected layer of the aforementioned BP-MTN fitting model to achieve classification;2)the classifier does not require the extraction of fault features,thus enabling end-to-end fault diagnosis.In addition,this paper conducts in-depth research on the influence of network polynomial order and activation function on the accuracy of fault diagnosis and model complexity.Finally,experiments show that the BP-MTN classifier can efficiently achieve fault diagnosis for the AHU system.