1. Introduction
With the rapid development of computer and sensor technology, modern industrial systems present a tendency towards complexity and integration, and the data reflecting the operation mechanism and state of the system presents all the characteristics of “big data”. As the most vital equipment of the refrigeration system in an Internet Data Center (IDC), the chiller is primarily composed of the condenser, the evaporator, the compressor, the expansion valve, the cooling water circulation system, as well as the chilling water circulation system, as illustrated in
Figure 1. The primary function of the chiller is to provide cooling source for the IDC room and guarantee the normal operation of the data center. The chiller fault occurrence not only shortens the equipment service life, decreasing the system performance, but also results in the loss of the information stored in the servers, bringing severe and even irreversible economic losses [
1]. The accurate fault diagnosis, therefore, is of great significance to the safety of the IDC.
The commonly used methods for fault diagnosis can be divided into three categories: model-based fault diagnosis, knowledge-based fault diagnosis and data-driven fault diagnosis [
2]. Model-based fault diagnosis methods are based on the internal mechanism of system, which estimates the system by constructing the mathematical model that is sensitive to specific faults and achieves fault diagnosis through the deviation between estimates and measurements [
3]. However, this method is not scalable, and each model can only be used for each specific system.
Knowledge-based fault diagnosis methods do not depend on mathematical or physical models, but rather the diagnosis results are determined by the expert experience and the level of expert knowledge [
4]. The data-driven fault diagnosis methods mainly use various data mining techniques to extract historical data features during the operation of the equipment and realize fault diagnosis by judging the consistency of the current data and those historical data features. Among them, the data-driven fault diagnosis method has often been used for fault diagnosis of chillers in recent years, including multivariate statistical analysis methods, signal processing methods, and machine learning methods, as shown in
Figure 2.
Support Vector Machine (SVM) [
5], Back Propagation Neural Network (BPNN) [
6], multivariate statistical analysis methods comprising PCA [
7] or ICA [
8], are also called traditional intelligent fault diagnosis methods. References [
9,
10] studied the fault detection and diagnosis of chiller sensors and established a PCA-based diagnostic model. Compared with reference [
9], reference [
10] introduced a wavelet analysis method on the basis of PCA, effectively filtering the noise in the sensor fault information and improved the deficiencies of the principal component method in the fault detection and diagnosis of chiller sensors. As we all know, the main component of PCA is a linear combination of various variables. When the values of some variables are similar, this can lead to poor sensitivity. To solve the problem of insufficient sensitivity of the PCA method, reference [
11] proposed a fault diagnosis method for air-conditioning sensors based on sparse principal components. This method used an elastic net to sparse the load matrix, reduce the association between principal components and variables, enhance the interpretability of principal components, and thereby improved the sensitivity of fault detection. In addition, reference [
8] used the independent component analysis (ICA) method to extract the correlation of the chiller variables and reduce the dimensionality of the measurement data. This method checked whether the chiller fails by counting and calculating the threshold of the statistic. Experimental results showed that the method is very sensitive to early failures and can effectively reduce the rate of false negatives and false positives. This method improves the diagnostic performance of the PCA model to a certain extent. In order to solve the generalization ability of the fault diagnosis algorithm on the chiller multi-classification problem, reference [
12] used the support vector machine (SVM) method to classify seven common chiller faults. Reference [
13] established a BP-based neural network diagnostic model for typical local faults and system faults of centrifugal chiller units. By adjusting the network structure and parameters, and changing the training function to optimize the model, a good diagnosis effect was obtained in the local fault diagnosis of the chiller. However, due to the widespread impact of system failures (such as refrigerant leakage) on refrigeration systems, it is difficult to identify. Against the defects of error back-propagation neural network in chiller fault diagnosis, reference [
14] used particle swarm optimization (PSO) to apply the optimized weights and thresholds model to the fault diagnosis of centrifugal chiller. The experimental results showed that compared with the traditional BP neural network, the optimized BP by PSO has significantly improved fault diagnosis performance, and the false alarm rate of fault diagnosis is reduced, and system faults, especially refrigerant leakage faults, are significantly improved.
However, the chiller is a highly non-linear complex system. These traditional intelligent fault diagnosis methods face difficulties in representing complex functions and take a lot of time to extract effective features due to their poor performance and generalization ability. Moreover, they perform feature extraction and diagnosis separately, which will affect the final diagnosis performance.
Compared with the traditional intelligent fault diagnosis methods, deep learning methods contain a multi-layer hidden structure that may realize the feature matrix transformation layer by layer and guarantee effective feature extraction adaptively [
15,
16]. Deep learning, on the other hand, can approach complex functions better; thus, it may deal with high-dimensional and non-linear data efficiently and avoid the issue of insufficient diagnostic capability through multiple non-linear transformations and approximate complex non-linear functions [
17]. Although fault diagnosis based on deep learning has attracted extensive attention in industry and academia, there are relatively few studies on fault diagnosis of chillers. Reference [
18] offered a method on the basis of LSTM to diagnosis the fault of chillers, and obtained outstanding performance. Applying LSTM, reference [
19] offered a fault detection and diagnosis method for the sensors of an air conditioning system. The fixed and drifting biases of both liquid line and discharge temperature sensors were successfully identified by building the fault detection and diagnosis models, respectively.
In this study, the collected data set from a chiller contains time-series data of multi-sensors and presents the “big data” characteristic. In deep learning, RNN and 1D-CNN are more capable of capturing connections in the time dimension. During the RNN variations, GRU improves the training efficiency by simplifying connection and reducing trainable parameters on the premise of ensuring the memory ability of neurons compared to LSTM. Therefore, a new approach for feature extraction and fault diagnosis is constructed based on 1D-CNN and GRU for the chiller fault diagnosis in this paper. Experiences evaluate the feasibility of the proposed diagnosis model on datasets with four levels of severity. Besides, the advantages of the model are verified by comparing it with 1D-CNN, GRU, LSTM, BPNN, PCA_BPNN, as well as 1D-CNN_LSTM. The leading contributions of this study may be summarized as below:
- (1)
Applying 1D-CNN and GRU, an innovative approach for feature extraction and fault diagnosis of the chiller is introduced in this article. The proposed approach can implement automatically features extraction from raw sensor data and fault diagnosis, simultaneously [
3];
- (2)
The experiments are performed on 4 kinds of datasets with different fault severity; the experimental results reveal that the proposed fault diagnosis algorithm has a reasonable identification rate for minor faults.
The remainder of this paper is structured as follows: In
Section 2, the experimental platform is introduced and the experimental data are briefly analyzed. The preparations related to the proposed approach and detailed descriptions of the proposed approach are presented in
Section 3. The diagnosis process and evaluation are presented in
Section 4. The experimental results and corresponding analysis are offered in
Section 5, respectively. To end, the conclusions are provided in
Section 6.