Elsevier

Computers & Chemical Engineering

Volume 126, 12 July 2019, Pages 189-203
Computers & Chemical Engineering

Modeling and fault diagnosis design for HVAC systems using recurrent neural networks

https://doi.org/10.1016/j.compchemeng.2019.04.011Get rights and content

Abstract

In this work, we develop models and a fault detection and isolation (FDI) methodology for heating, ventilation and air conditioning (HVAC) systems that utilizes recurrent neural networks (RNN). The FDI design does not require the existence of plant fault history, mechanistic models or a set of expert rules to isolate faults. The key is to first use plant data to build predictive models and input/output estimators, and then embed them within FDI filters. A distributed FDI framework is designed consisting of local FDI (LFDI) schemes that communicate with each other for improved FDI. The distributed FDI framework enables diagnosis of multiple faults in different components of the HVAC system when a fault in one of the control components directly affects the other subsystems. The effectiveness of the proposed FDI scheme is shown via simulation examples on a simulation test bed, as well as using real data. The simulations revealed superior performance of the proposed FDI methodology over FDI approaches using subspace based models for both simulation and real data cases.

Introduction

Modern heating, ventilation and air conditioning (HVAC) systems rely on automation to ensure occupant comfort and meet the energy regulations set by governments. However, as the level of automation increases, it becomes increasingly important to diagnose and repair failures in control components such as actuators and sensors, to minimize discomfort or wasteful energy expenditure. These considerations have motivated interest in fault diagnosis and fault tolerant control (FTC) frameworks and their applications to HVAC systems.

Model based fault diagnosis methods can be divided in two categories based on the type of model: mechanistic and empirical (see e.g., MacGregor and Cinar, 2012). Mechanistic/first principles models (hereon referred to as first principles models) have structures based on laws of physics with parameters are computed from plant data. Empirical or data-driven models have structures and associated parameters are determined based on the ability to model and predict plant data (see e.g., MacGregor and Cinar, 2012).

Mechanistic/first-principles models often contain information (analytical redundancies) that can be readily utilized by appropriately designed fault detection and isolation (FDI) structures to detect and isolate both simple and complex (multiple/simultaneous actuator/sensor) faults (see e.g., Du, 2012, Du, Scott, Mhaskar, 2013, Shahnazari, 2018). First principles models, however, in general remain difficult to develop and require special expertise to maintain, thus motivating interest in data driven models.

The structure/choice of the data driven model strongly influences the ability to achieve FDI. Principal component analysis (PCA) and/or partial least square (PLS) based models have been developed and used for FDI via multivariate statistical process monitoring. These FDI structures, however, fundamentally are only able to diagnose faults in a single variable (see e.g., Yoon and MacGregor, 2000 and Joe Qin (2003)), with or without the use of fault history (see e.g., Yoon and MacGregor, 2001) unless they are used in conjunction with other methods such as discriminant analysis (see e.g., Raich, Cinar, 1995, Raich, Cinar, 1996).

Other researchers have considered methods based on combinations of heuristics rules and machine learning methods such as artificial neural networks (ANN) and support vector machines (SVM) (see e.g., Lee, House, Park, Kelly, 1996, Liang, Du, 2007). In Du et al. (2014), a combination of multiple ANNs and subtractive clustering analysis are used for fault diagnosis in HVAC systems. The combined ANNs are used for fault detection while fault isolation is achieved by means of clustering on both normal and faulty data. In Khan et al. (2014), a fault detection methodology is proposed using pattern recognition techniques applied on residual neural networks for building energy management. In Guo et al. (2018), a fault diagnosis methodology is proposed using a deep belief network for variable refrigerant flow air-conditioning. However, all of the aforementioned results rely on either plant fault history, heuristic rules or pattern recognition techniques for fault isolation, thus requiring an exhaustive fault history to achieve FDI. One of the reasons for the inability of the above methods to achieve FDI is that they do not embed a model of the process that can capture the dynamic behavior of the process, but instead try to ‘model’ the faults directly.

Subspace identification-based dynamic models (see e.g., Ljung, 1998, Van Overschee, De Moor, 2012) have been more recently used within FDI based strategies and shown to achieve FDI for simple and complex faults (see e.g., Shahnazari, Mhaskar, House, Salsbury, 2018, Shahnazari, Mhaskar, House, Salsbury, 2018). The recent results do not require a fault history and are able to diagnose faults for which the effect of the fault is masked by the presence of feedback control. The existing results are however, limited by the ability of the identification technique to capture the process nonlinearity, and stand to gain from the strong modeling capabilities of recurrent neural networks (RNN). Kusiak and Xu (2012) used non-linear autoregressive with external input (NARX) recurrent neural networks to model and optimize the system energy consumption of an HVAC systems suggesting the viability of neural network based models for FDI in HVAC systems. Another important consideration in the FDI design is the complexity associated with the design. Thus, increasingly distributed FDI designs are sought, where local FDI designs are built, albeit with appropriate communication, thus reaping the benefits of a centralized FDI design, but without the possibility of the failure of the entire FDI structure should a component of the FDI design fail.

Motivated by the above considerations, we propose a new distributed FDI methodology for HVAC systems based on RNN, and demonstrate the application on simulated and real data. The rest of the manuscript is organized as follows: In Section 2, brief descriptions of the simulation/real testbed for the HVAC system under consideration are described and RNNs are reviewed. Then we briefly review our previous FDI design for HVAC systems presented in Shahnazari and Mhaskar (2018b) and Shahnazari et al. (2018a). In Section 3, RNN’s are used to build dynamic predictive models and input/output estimators for the various components of the HVAC system, and are validated against new data sets consisting of healthy plant data. This combination also allows us to build separate local models for components of HVAC systems that work successfully partly because of their small scale. In Section 4, the proposed FDI methodology is presented. A distributed FDI framework is designed for the HVAC system that consist of local FDI (LFDI) schemes communicating with each other for improved FDI using the methodology presented in Shahnazari and Mhaskar (2018b) and Shahnazari et al. (2018a). This enables diagnosis of single/multiple faults in different components of HVAC systems in the presence of faults that are originated in one component but affect the other components directly. Thresholds are selected in a way that accounts for modeling uncertainty and guarantee robustness of the FDI filter. In Section 5, the effectiveness of the FDI design is shown via simulation examples and real building data. The results illustrate the superior capabilities of the RNN based FDI design compared to previous results based on models obtained via subspace identification techniques. Finally, Section 6 presents some concluding remarks.

Section snippets

Simulation/real testbed

In this section, we describe the test bed HVAC system which is a five-zone HVAC system that consists of standard components used in HVAC systems in both air handling unit (AHU) and zone level equipment known as variable air volume (VAV) boxes (see e.g., Shahnazari et al., 2018b for further details), modeled using Modelica1 The

Model identification and validation using LRNs

The following algorithm briefly describes the methodology for selecting a neural network structure that is commonly used to avoid over-fitting (see e.g., Chollet, 2017):

  • 1.

    Change the following network parameters to get the desired fit on training data i.e., the maximum observed value for absolute error is less than d1 (where d1 is a user specified parameter) on the training data:

    • (a)

      Number of epochs

    • (b)

      Number of neurons in each hidden layers

    • (c)

      Number of hidden layers

    • (d)

      Number of delays

    To start the training,

Fault detection and isolation design for multiple sensor faults using LRNs

In the present contribution, we adapt the methodology presented in Du and Mhaskar (2014) by replacing the state predictor and state estimators with output predictors and output estimators designed using LRNs. Again a bank of residuals are designed to detect and isolate single, and multiple faults where each residual is insensitive to a subset of sensor faults. To begin with, the insensitive residuals to a single fault in yi is generated as follows:ri=yi˜yi^where yi˜ and yi^ are the predicted

Simulation results

We first consider the case where a positive bias fault occurs in the supply air temperature sensor at 550 min into the simulation of a summer day (June 16). This is followed by a fault in the VAV box damper of the south zone 750 min into the simulation. Fig. 13 shows the estimated and predicted trajectories for supply air temperature and outlet water temperature. As can be seen, after fault occurrence the estimated values of outlet water temperature (double dotted red line) deviate from its

Conclusion

In this work, we used layer recurrent network (LRN)s for modeling and fault diagnosis in HVAC systems. The FDI design does not require the existence of plant fault history, mechanistic models or a set of expert rules to isolate faults. The idea is to use neural networks to build predictive models and input/output estimators utilizing healthy plant data. Then, FDI filters are designed in a way that can diagnose multiple faults in sensors or actuators. The effectiveness of the proposed FDI scheme

Acknowledgment

Financial support from the Natural Sciences and Engineering Research Council of Canada (20004471) through the Collaborative Research and Development Program (in collaboration with Johnson Controls, Inc.) is gratefully acknowledged.

References (30)

  • S. Yoon et al.

    Fault diagnosis with multivariate statistical models part i: using steady state fault signatures

    J. Process Control

    (2001)
  • G. Zhang et al.

    Forecasting with artificial neural networks: the state of the art

    Int. J. Forecast.

    (1998)
  • F. Burden et al.

    Bayesian regularization of neural networks

    Artificial Neural Networks

    (2008)
  • R. Caruana et al.

    Overfitting in neural nets: backpropagation, conjugate gradient, and early stopping

    Advances in Neural Information Processing Systems

    (2001)
  • F. Chollet

    Deep Learning with Python

    (2017)
  • Cited by (84)

    View all citing articles on Scopus
    View full text