Water chilling unit fault diagnosis method based on generalized self-attention category-level feature fusion network in mixed fault level scene

Abstract

The invention discloses a water chilling unit fault diagnosis method based on a generalized self-attention category-level feature fusion network in a mixed fault level scene, and mutual interference of information between categories and levels is avoided. Firstly, a generalized self-attention global-local feature extraction network based on Lp-norm and multi-cluster cavity convolution is provided to obtain original features and category features of the water chilling unit, and secondly, grade features are obtained through a residual feature extraction network based on a convolution auto-encoder. And finally, fault diagnosis is carried out through fusion features. Experimental results on an ASHRAE Research Project 1043 data set show that representative features can be extracted from data mixed by various fault categories and fault levels, and the method is more in line with practical application scenarios.

Description

Technical Field

[0001] The present invention relates to the field of data-driven fault diagnosis technology, and in particular to a fault diagnosis technology for chillers. The data-driven method of the present invention is a specific application in the field of chiller fault diagnosis. Background Art

[0002] With the rapid development of the global economy and the advancement of science and technology, HVAC systems are increasingly used in buildings, and HVAC accounts for more than half of building energy loss. As the core component of HVAC, the chiller provides a cold source for the entire system. When a unit fails, it will cause a lot of energy waste and even cause serious safety problems. Therefore, timely and accurate fault diagnosis of the operating status of the chiller is of great research significance.

[0003] At present, research on data-driven fault diagnosis methods is widely used in chillers, such as support vector machine (SVM) based on machine learning, principal component analysis (PCA), convolutional neural network (CNN) based on deep learning, temporal convolutional network (TCN), etc.

[0004] However, the data of the chiller contains not only information about the fault category, but also information about the severity of the fault, which is called the fault level. In practice, data of different fault categories and fault levels coexist, and there is a big difference between the information of the fault category and the information of the fault level, and the two interfere with each other. The data of multiple fault categories and fault levels are mixed together, which greatly limits the fault diagnosis ability of previous models. In addition, it is equally important for chillers to diagnose the severity of the fault and the cause of the fault. Therefore, chiller fault diagnosis in mixed fault level scenarios is very necessary. Summary of the invention

[0005] In order to overcome the above problems, the present invention proposes a chiller fault diagnosis method in a mixed fault level scenario based on a generalized self-attention category-level feature fusion network (GACLFN). The category features of the chiller are obtained through a global local feature extraction network, and the level features of the chiller are obtained using a residual feature extraction network based on a convolutional autoencoder (CAE), and fused with the features of the data itself to avoid the problem of mutual interference between the two information. And a generalized self-attention mechanism based on the Lp-norm is proposed to better measure the importance of features and enhance the feature extraction capability of the network.

[0006] The present invention adopts the following technical solutions and implementation steps:

[0007] A. Offline modeling stage:

[0008] 1) Normalize the historical operating data of the chiller, which includes the operating data X under normal conditions and normal , also contains data of different fault types and fault levels X fault The normalized formula is:

[0009]

[0010] in, represents the normalized data, and X represents all the running data (X normal ,X fault ), X max With X min represent the maximum and minimum values ​​in X respectively.

[0011] 2) Build the GACLFN model and train it. The specific steps are as follows:

[0012] (2.1) Train the convolutional autoencoder (CAE) and obtain the hierarchical features: First, use the data of normal samples Train the CAE model and use the binary cross entropy loss function to calculate the reconstruction loss of the model. For the normal sample data input The reconstructed data is Then the reconstruction loss can be expressed as:

[0013]

[0014] Then all the data Input the trained CAE model to obtain the reconstructed data Subtract the reconstructed data from the original data to get the residual features:

[0015]

[0016] The absolute value of the residual feature is used as the final level feature:

[0017] F level =|X res |

[0018] (2.2) Train the generalized self-attention global-local feature extraction network (GAGLN) and obtain the original features and category features: Train two GAGLN models to extract the original features and category features respectively. During the training process, all label information is input into the network responsible for extracting the original features, while the network responsible for extracting the category features only inputs the fault category label, not the fault level label. In addition, an ArcFace loss function is used to calculate the classification loss of the network:

[0019]

[0020] Among them, j represents the jth category, and c represents the total number of sample categories. i represents the i-th sample, j i represents the category with the highest score for the i-th sample. Represents x i The feature vector of i The angle between the weight vectors on j Represents x i The angle between the feature vector of and the weight vector on category j. N represents the data batch size, s is the scaling factor, and m is the angular margin penalty term. The data loss L is obtained through the ArcFace loss function d and the class loss L c , the overall loss function of the network is defined as:

[0021] L=L d +αL c

[0022] Among them, α is the weight coefficient, which is used to adjust the importance of category loss in the total loss. Finally, the original feature F is obtained through the trained GAGLN model original and the category feature F categor y .

[0023] (2.3) The level feature F obtained from (2.1) level The original feature F obtained by (2.2) original and the category feature F category The fused features F are obtained by concatenation, and the CNN classifier is trained by the fused features. During the training process, the cross entropy loss function is used to optimize the parameters of the classifier.

[0024] B. Online diagnosis stage:

[0025] 3) The data monitored online by the chiller is normalized according to the formula in step 1).

[0026] 4) The normalized chiller data is input into the trained GACLFN model, three features are extracted and connected to obtain fusion features, and fault classification is performed through fusion features to obtain online fault diagnosis results.

[0027] Beneficial Effects

[0028] 1) In order to solve the problem that multiple fault level data of chillers are mixed in actual processes, and the fault category and fault level information interfere with each other, resulting in difficulty in feature extraction, the present invention proposes a feature extraction method of category-level feature fusion (CLFF). By separately extracting the category and level features in the data, the mutual influence of the two is avoided, and the information contained in the features is enriched.

[0029] 2) This paper proposes a generalized self-attention mechanism based on Lp-norm, which enhances the network's ability to focus on important features and improves the accuracy of chiller fault classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is the overall framework of the GACLFN model of the present invention;

[0031] Figure 2 This is the specific structure diagram of the GAGLN module;

[0032] Figure 3 This is the structural diagram of the MAC module in GAGLN;

[0033] Figure 4 This is the structural diagram of the GSA-L module in GAGLN;

[0034] Figure 5 This is the structural diagram of the GAS-G module in GAGLN;

[0035] Figure 6 It is the structural diagram of the CAE-based hierarchical feature extraction module;

[0036] Figure 7 This is the structure diagram of the CNN fault classifier;

[0037] Figure 8 It is the overall flow chart of the method of the present invention;

[0038] Fig. 9 The bar graph shows the average fault diagnosis accuracy of SVM, CNN, ResNet18, TCN and the method GACLFN proposed in the present invention under seven fault categories and four fault levels. ;

[0039] Fig.10Radar chart of the diagnostic accuracy of the five methods for seven fault categories and the normal category.

[0040] Fig.11 This is a line graph showing the diagnostic accuracy of the five methods for four fault levels. DETAILED DESCRIPTION

[0041] The ASHRAE Research Project 1043 (RP-1043) dataset is generated by a simulation experiment of a 90-ton centrifugal chiller using R134a as the refrigerant, and is often used to evaluate chiller fault diagnosis methods. The dataset contains parameters for the normal operation (No) of the chiller and seven fault conditions. Each fault category is divided into four fault levels according to the severity of the fault, with levels one to four gradually increasing. The specific content is shown in Table 1.

[0042] Table 1 Seven fault categories and four fault levels of chillers

[0043]

[0044] In order to verify the fault diagnosis ability of the model under mixed fault levels, 200 samples are selected from the data of each fault category as training data, and the samples come from four fault levels, each fault level contains 50 samples, totaling 1400 fault samples. Then 50 samples are selected from the normal data to form 1450 training samples. Using the same method, 100×4×7=2800 fault samples and 100 normal samples are obtained to form 2900 test samples.

[0045] Applying the method of the present invention to the above data set includes two major steps: offline modeling and online diagnosis, which are specifically stated as follows:

[0046] A. Offline modeling stage:

[0047] 1) Normalize 1450 training samples, which include 50 normal samples X normal =x1,x2,…,x 50 , and 1400 fault samples X from different fault categories and fault levels fault =x 51 ,x 52 ,…,x 1450 . Among them, each sample x i Contains 64 items of data measured and indirectly calculated between sensors. The normalized formula is:

[0048]

[0049] in, represents the normalized data, and X represents all the running data (X normal ,X fault ), X max With X min represent the maximum and minimum values ​​in X respectively.

[0050] 2) Build the GACLFN model and train it. The specific steps are as follows:

[0051] (2.1) Train the convolutional autoencoder (CAE) and obtain the hierarchical features: First, use the data of normal samples Train the CAE model and use the binary cross entropy loss function to calculate the reconstruction loss of the model. For the normal sample data input The reconstructed data is Then the reconstruction loss can be expressed as:

[0052]

[0053] Then all the data Input the trained CAE model to obtain the reconstructed data Subtract the reconstructed data from the original data to get the residual features:

[0054]

[0055] The absolute value of the residual feature is used as the final level feature:

[0056] F level =|X res |

[0057] (2.2) Train the generalized self-attention global-local feature extraction network (GAGLN) and obtain the original features and category features: Train two GAGLN models to extract the original features and category features respectively. During the training process, all label information is input into the network responsible for extracting the original features, while the network responsible for extracting the category features only inputs the fault category label, not the fault level label. In addition, an ArcFace loss function is used to calculate the classification loss of the network:

[0058]

[0059] Where j represents the jth category, c represents the total number of sample categories. In this experiment, there are 29 sample categories, so c = 29. i represents the i-th sample, ji represents the category with the highest score for the i-th sample. Represents x i The feature vector of i The angle between the weight vectors on j Represents x i The angle between the feature vector of and the weight vector on category j. N represents the data batch size, s is the scaling factor, and m is the angular margin penalty term. In the experiment, the batch N is set to 128, s is set to 30, and m is set to 0.15. The data loss L is obtained by the ArcFace loss function. d and the class loss L c , the overall loss function of the network is defined as:

[0060] L=L d +αL c

[0061] Among them, α is the weight coefficient, which is used to adjust the importance of category loss in the total loss. In this experiment, α is set to 0.85. Finally, the original feature F is obtained through the trained GAGLN model. original and the category feature F category .

[0062] (2.3) The level feature F obtained from (2.1) level The original feature F obtained by (2.2) original and the category feature F category The fused features F are obtained by concatenation, and the CNN classifier is trained by the fused features. During the training process, the cross entropy loss function is used to optimize the parameters of the classifier.

[0063] B. Online diagnosis stage:

[0064] 3) Normalize the 2900 test samples according to the formula in step 1).

[0065] 4) The normalized chiller data is input into the trained GACLFN model, three features are extracted and connected to obtain fusion features, and fault classification is performed through fusion features to obtain fault diagnosis results.

[0066] The above steps are the specific application of the method of the present invention to fault diagnosis on the RP-1043 data set. In order to verify the effectiveness of the method of the present invention, the traditional machine learning method SVM and the deep learning methods CNN, ResNet18, TCN are compared with the GACLFN proposed in the present invention. The average fault diagnosis rate (FDR) is used a ) and standard deviation (SD) are used to measure the fault diagnosis ability of the model under mixed fault levels. The fault diagnosis rate FDR is defined as:

[0067]

[0068] Among them, N represents the total number of samples, N T Indicates the number of samples that are accurately classified. We conducted 10 experiments on each model and took the average value of the FDR of the 10 experiments to obtain the average fault diagnosis rate FDR a . Then SD can be defined as:

[0069]

[0070] SD measures the dispersion of data relative to the mean. The lower the SD, the more concentrated the data is, and the higher the stability and reliability.

[0071] According to the above settings, 10 experiments were performed on each method to obtain the FDR of the five methods. a and SD are shown in Table 2.

[0072] Table 2 FDR of five fault diagnosis models a and SD

[0073]

[0074] It can be seen from Table 2 that the FDR of the method proposed in the present invention is a The highest value reaches 95.8%. In addition, the SD value is also the smallest among the five methods, only 0.28%, which shows that the method proposed in the present invention has excellent fault diagnosis ability in the scenario of mixed fault levels and has high stability.

[0075] In order to demonstrate the fault diagnosis capability of the proposed method in mixed fault level scenarios, the fault diagnosis results in Table 2 are further analyzed, and the FDR of 29 categories is obtained by calculating the confusion matrix. a , which includes seven fault categories and one normal category. Each fault category is subdivided into four fault categories according to the fault level.

[0076] The method proposed in this invention has a better diagnosis effect on most faults than other methods, and the FDR of each fault is a All of them are above 89%. This shows that the GACLFN model proposed in the present invention can accurately distinguish different types of faults and different levels of each fault in a scenario where multiple fault types and fault levels are mixed.

[0077] In order to further compare the fault diagnosis capabilities of these methods, we statistically analyzed the above experimental results from the perspectives of fault category and fault level. First, from the perspective of fault category, for the seven fault categories and normal category of the chiller, the FDR of the above five methods in ten experiments was calculated.a .

[0078] FDR of GACLFN for each fault category a All of them are above 90%. Except for FWE, which is similar to the other four methods in diagnostic effect, GACLFN is better than other methods in other categories. Among them, the FDR of NC, RL, RO, and CF a The classification effect of normal samples exceeds that of other methods by more than 10%, which shows that the GACLFN model proposed in the present invention can achieve good fault diagnosis effect on different fault categories.

[0079] Similarly, from the perspective of fault level, the FDR of the above five methods at four fault levels is statistically analyzed. a .

[0080] Depend on Fig.10 It can be seen that the FDR of the GACLFN model in the four fault levels a All of them reached more than 93%, which is higher than the other four methods. Among them, the diagnostic effect on fault level 1 exceeds that of other methods by more than 7%, indicating that the method proposed in the present invention can also achieve a good fault diagnosis effect for low-level faults.

Claims

1. A chiller fault diagnosis method based on a generalized self-attention category-level feature fusion network in a mixed fault level scenario, including two stages: "offline modeling" and "online diagnosis", characterized in that: The specific steps are as follows: A. Offline modeling stage: 1) Normalize the historical operating data of the chiller, which includes the operating data X under normal conditions and normal , also contains data of different fault types and fault levels X fault ; The normalized formula is: in, represents the normalized data, and X represents all the running data (X normal ,X fault ), X max With X min Represent the maximum and minimum values ​​in X respectively; 2) Build the GACLFN model and train it. The specific steps are as follows: (2.1) Train the convolutional autoencoder CAE and obtain hierarchical features: First, use the data of normal samples Train the CAE model and use the binary cross entropy loss function to calculate the reconstruction loss of the model; for the normal sample data input The reconstructed data is Then the reconstruction loss can be expressed as: Then all the data Input the trained CAE model to obtain the reconstructed data Subtract the reconstructed data from the original data to get the residual features: The absolute value of the residual feature is used as the final level feature: F level =|X res | (2.2) Train the generalized self-attention global-local feature extraction network GAGLN and obtain the original features and category features: Train two GAGLN models at the same time to extract the original features and category features respectively. During the training process, all label information is input into the network responsible for extracting the original features, while the network responsible for extracting the category features only inputs the fault category label, not the fault level label; use an ArcFace loss function to calculate the classification loss of the network: Among them, j represents the jth category, c represents the total number of sample categories; x i represents the i-th sample, j i Represents the category with the highest score of the i-th sample; Represents x i The feature vector of i The angle between the weight vectors on j Represents x i The angle between the feature vector of and the weight vector on category j; N represents the data batch size, s is the scaling factor, and m is the angular margin penalty term; the data loss L is obtained by the ArcFace loss function d and the class loss L c , the overall loss function of the network is defined as: L=L d +αL c Among them, α is the weight coefficient, which is used to adjust the importance of category loss in the total loss; finally, the original feature F is obtained through the trained GAGLN model original and the category feature F category ; (2.3) The level feature F obtained from (2.1) level The original feature F obtained by (2.2) original and the category feature F category The fusion feature F is obtained by concatenation, and the CNN classifier is trained by the fusion feature. The cross entropy loss function is used to optimize the parameters of the classifier during the training process. B. Online diagnosis stage: 3) Normalizing the data monitored online by the chiller according to the formula in step 1); 4) The normalized chiller data is input into the trained GACLFN model, three features are extracted and connected to obtain fusion features, and fault classification is performed through fusion features to obtain online fault diagnosis results.

Images