Non-stationary semi-supervised domain generalization network-based multi-working-condition chiller fault diagnosis method

Abstract

The present application relates to a water chiller multi-working condition fault diagnosis method based on a non-stationary semi-supervised domain generalization network, first, the variables of the water chiller are subjected to a dickey-fuller test, and the variables of the water chiller are divided into stationary variables and non-stationary variables; the non-stationary variables are subjected to a cointegration analysis, the non-stationary variables with a cointegration relationship are transformed by using a cointegration matrix, and the transformed variables and the stationary variables are fused and used as input data of the domain generalization network. The present application takes the feature extractor of the domain generalization network as an encoder, constructs a corresponding decoder, and improves the fault feature extraction capability of the domain generalization network; one labeled source domain and two unlabeled source domain data are used to train the fault diagnosis model, and the trained model is used to realize the multi-working condition fault diagnosis of the water chiller.

Description

Technical Field

[0001] This invention relates to the field of data-driven fault diagnosis methods, and in particular to a multi-condition fault diagnosis method based on a non-stationary semi-supervised domain generalization network for multi-condition chiller units with non-stationary operating data. Background Technology

[0002] Multi-condition operation refers to the different operating states of a chiller unit due to varying settings such as temperature and operating load. In some conditions, data may be unlabeled, or even unavailable, making it difficult to establish a fault diagnosis model. Furthermore, the non-stationarity of chiller unit data complicates fault diagnosis. Chiller unit malfunctions result in significant energy waste and can even lead to safety accidents. Therefore, effective fault diagnosis is crucial for energy conservation, emission reduction, and personnel safety. The complex structure of chiller units makes knowledge-based and model-based fault diagnosis methods ineffective. In recent years, with the storage of vast amounts of historical data, data-driven fault diagnosis methods have been extensively researched.

[0003] In data-driven multi-condition fault diagnosis methods, deep transfer learning has been widely applied due to its powerful feature extraction capabilities. Deep transfer learning can be divided into three categories: instance-based methods, model-based methods, and feature-based methods. Instance-based methods select instances from the source domain that are beneficial for training the target domain model and weight these instances. Instance-based methods are relatively simple and easy to implement, but weight selection and similarity metrics require prior knowledge. Model-based methods assume that the source and target domains share common knowledge in their models, using a pre-trained deep learning model in the source domain to improve the target domain model. Model-based transfer learning can fully utilize the similarity between the source and target domain models, but model convergence is difficult. Feature-based methods learn mapping functions, transforming the original data of the source and target domains into a common feature space, learning domain-invariant features of the source and target domains, and using these domain-invariant features to achieve fault diagnosis in the target domain. Feature-based methods have achieved good results in feature selection and transformation, and are therefore adopted by most deep transfer learning methods. Among feature-based methods, domain adversarial networks have been widely used due to their powerful domain-invariant feature extraction capabilities. In fault diagnosis based on domain adversarial networks (DANs), if the training set data includes samples from the target domain, the DAN is called a domain adaptive network. Some researchers have used domain adaptive methods to achieve fault diagnosis for multi-condition chiller units. However, training a domain adaptive fault diagnosis model requires a small number of labeled samples from the target domain. In fault diagnosis based on DANs, if the training set data only contains samples from the source domain, the DAN is called a domain generalization network. Some researchers have used domain generalization methods to achieve fault diagnosis for multi-condition chiller units. Domain generalization-based fault diagnosis models only need to be trained using labeled samples from each source domain, and by extracting domain-invariant fault features, they can achieve fault diagnosis in the target domain. However, labeling samples requires significant human and financial resources. Furthermore, the non-stationarity of the data makes it difficult for domain generalization networks to extract domain-invariant fault features, resulting in poor fault diagnosis performance. Summary of the Invention

[0004] To address the challenge of multi-condition chiller unit fault diagnosis due to the non-stationarity of chiller unit data, this invention proposes a multi-condition chiller unit fault diagnosis method based on a non-stationary semi-supervised domain generalization network (NSSDG). First, to address the difficulty in extracting domain-invariant fault features due to data non-stationarity, cointegration analysis is performed on the non-stationary variables, and cointegration matrices are used to stabilize non-stationary variables with cointegration relationships. Second, to address the problem of poor fault diagnosis performance due to the inability of the domain generalization network to extract fault category features from source domain samples, the feature extractor also functions as a convolutional encoder, constructing a corresponding convolutional decoder to reconstruct the input data, thereby fully extracting the fault category features of the source domain samples and improving the fault diagnosis effect.

[0005] A multi-condition chiller unit fault diagnosis method based on nonstationary semi-supervised domain generalization network is characterized by the following steps:

[0006] A. Offline modeling stage:

[0007] 1) Obtain the training dataset. The training dataset consists of data from three operating conditions of the chiller unit. Seven types of fault data were collected for each of the three operating conditions. Data from one operating condition is labeled, while data from the other two operating conditions is unlabeled. The number of data points for each type of fault collected is the same. The training dataset can be represented as X = [x1, x2, ..., x...]. n ] T ∈R n×m Where n represents the number of samples and m represents the sample dimension, that is, the number of variables contained in each sample;

[0008] 2) Perform the Dickie-Fowler test on the m variables in the training dataset and divide them into stationary variables. Non-stationary variables Where m1 + m2 = m, the Dickey-Fowler test principle is as follows: time series y t The first-order autoregressive process is shown in equation (1):

[0009] y t =ρy t-1 +ε t (1)

[0010] Where ρ is the correlation coefficient, {ε t} represents the error, E* represents the mean of *, and Var(*) represents the variance of *. If ρ=1, {ε t} Independent and identically distributed, and Eε t =0Var(ε t )=σ 2<∞, then y t For a non-stationary time series, if |ρ| < 1, then y t It is a stationary time series;

[0011] 3) For non-stationary variables The Johansen test is performed, and the non-stationary variables that are linear combinations of stationary variables are identified based on the maximum characteristic statistic in the test results. The Johansen test uses a vector autoregression model to determine whether there is a long-term cointegration relationship among multiple variables. The vector autoregression model is shown in formula (2).

[0012] y t =A1y t-1 +A2y t-2 +…+A n y t-n +Bx t +e t (2)

[0013] Among them, y t It is an endogenous variable column vector, x t It is a vector of exogenous variables, A1, A2, ..., A n B and E are the coefficient matrices of equal estimation. t It is an error vector, and the error variables within the error vector are correlated. The error variables do not exhibit autocorrelation. The error vector is related to y. t ,y t-1 ,…,y t-n ,x t If they are uncorrelated, then nonstationary variables can be categorized into nonstationary variables with cointegration relationships. and nonstationary variables that do not have cointegration relationships

[0014] 4) Convert non-stationary variables with cointegration relationships into stationary variables, as shown in formula (3);

[0015]

[0016] in, Let B be the non-stationary variable after being stabilized by cointegration transformation, and let B be the coefficient matrix estimated by the vector autoregressive model. It is a cointegrated vector;

[0017] 5) Change the variables and The input variables are obtained by fusion. The fusion method used is concatenate, where m = m1 + m 21 +m 22 ;

[0018] 6) Input data Perform maximum and minimum standardization, as shown in formula (4);

[0019]

[0020] Where, x min Each value in the vector is the minimum value of each variable, x. max Each value in the vector is the maximum value of each variable.

[0021] 7) Input the data after the maximum and minimum normalization into the domain generalization network. The domain generalization network in this invention consists of four parts: feature extractor, domain classifier, classifier and convolutional decoder. The feature extractor extracts the domain invariant fault features as shown in formula (5).

[0022] h=f(W*x+b)(5)

[0023] Where x represents the input data, W represents the weight matrix of the convolution kernel during feature extraction, b represents the bias during feature extraction, * represents the convolution operation, f(·) represents the encoding process, and h represents the hidden layer features extracted by the feature extractor.

[0024] 8) The domain classifier consists of fully connected layers, and the calculation process is shown in formula (6);

[0025] y = f d (∑(w d h+b d ))(6)

[0026] Where y is the output of the current layer, f d Represents the ReLU activation function, Σ represents the weighted summation over all inputs, and w d b represents the weights of the fully connected layer. d The loss function of the domain classifier, which represents the bias term, is shown in Equation (7).

[0027]

[0028] Among them, types d Represents the number of domain categories, i_tag d The domain label value representing a certain category, i_pre d This represents the probability value of the output.

[0029] 9) The classifier consists of fully connected layers, and the calculation process is shown in formula (8);

[0030] y c =f c (∑(w c h+b c ))(8)

[0031] Among them, y c It is the output of the current layer, f c Represents the ReLU activation function, Σ represents the weighted summation over all inputs, and w c b represents the weights of the fully connected layer. c The loss function of the classifier, which represents the bias term, is shown in formula (9).

[0032]

[0033] Among them, types c i_tag represents the number of fault categories. c The label value representing a certain category, i_pre c This represents the probability value of the output.

[0034] 10) The convolutional decoder consists of convolutional layers, and the calculation process is shown in formula (10);

[0035]

[0036] in, This represents the input data for reconstruction. This represents the weight matrix of the convolution kernel during the input data reconstruction process. The input data reconstruction process is represented by the bias, * represents the convolution operation, g(·) represents the decoding process, and the loss function for input data reconstruction is shown in formula (11).

[0037]

[0038] Where n represents the number of samples in the input data;

[0039] 11) Utilizing the loss function Loss = L + loss of a domain generalization network c +loss d Train the model;

[0040] B. Testing Phase:

[0041] 1) Obtain the test dataset. The test dataset consists of data from the fourth operating condition of the chiller unit. This condition collected data on seven different faults of the chiller unit, with the same number of data points for each fault type. The test dataset can be represented as follows: Where, n t This indicates the number of samples, and m represents the sample dimension, which is the number of variables contained in each sample.

[0042] 2) Based on the Dickey-Fowler test results in A1, X t Divide into stationary variables Non-stationary variables Where m1 + m2 = m;

[0043] 3) According to the Johansen test in A2, non-stationary variables... Non-stationary variables can be categorized into those with cointegration relationships. and nonstationary variables that do not have cointegration relationships

[0044] 4) Based on the VAR model in A2, non-stationary variables with cointegration relationships are converted into stationary variables, as shown in formula (12);

[0045]

[0046] in, For the non-stationary variable B after cointegration transformation to a stationary state, t The coefficient matrix is ​​estimated by the vector autoregression model.

[0047] It is a cointegrated vector;

[0048] 5) Change the variables and Variables obtained by fusion The fusion method used is concatenate, where m = m1 + m 21 +m 22 ;

[0049] 6) Apply formula (4) to the test data Perform maximum and minimum standardization;

[0050] 7) Standardize the test data by maximum and minimum values. The input is fed into the trained domain generalization network to determine the type of fault.

[0051] Beneficial effects

[0052] This invention utilizes cointegration analysis to stabilize non-stationary variables, ensuring that the domain generalization network effectively extracts domain-invariant fault features, thereby improving fault diagnosis performance. It also employs a convolutional autoencoder to reconstruct the input data, effectively extracting fault category features from each source domain, thus enhancing fault diagnosis accuracy. Attached Figure Description

[0053] Figure 1 The diagram shown is a flowchart of the fault diagnosis process for the nonstationary semi-supervised domain generalization network method of the present invention.

[0054] Figure 2 The figure shows the confusion matrix of the nonstationary semi-supervised domain generalization network method for diagnosing seven types of faults.

[0055] Figure 3The results show the comparative experimental results of the nonstationary semi-supervised domain generalization network method with other methods;

[0056] Figure 4 The results shown are ablation experiments of the nonstationary semi-supervised domain generalization network method.

[0057] Figure 5 The diagram shown illustrates the operating principle of the chiller unit. Detailed Implementation

[0058] The ASHRAE RP-1043 public dataset is widely used in chiller unit fault diagnosis research. This dataset collects data on chiller units under seven typical fault operating conditions, as shown in Table 1. Table 1 lists severe faults, specifically: cooling water flow rate reduced to 40% or more; chilled water flow rate reduced to 40% or more; refrigerant levels 40% or more higher or lower; lubricating oil levels 68% or more higher than normal; 74 or more blocked pipes (representing 45% of the total pipe count); and 0.54 pounds or more of nitrogen in the refrigerant (representing 5% of the total capacity). These conditions indicate a severe chiller unit malfunction requiring shutdown and maintenance. This dataset collected data on 64 variables of the chiller unit, as shown in Table 2. The operating principle of the chiller unit is as follows: Figure 5 As shown, the hot water loop simulates the indoor environment, and the tap water loop simulates the cooling tower. The main heat exchange processes are as follows: the chilled water loop transfers heat from the indoor environment to the chilled water through interaction with the indoor environment; the refrigerant loop transfers heat from the chilled water to the refrigerant through interaction with the refrigerant; and the cooling water loop transfers heat from the refrigerant to the cooling water through interaction with the cooling water. Finally, the cooling water transfers heat to the outdoor environment through the cooling tower. In addition, the chilled water loop and the cooling water loop can also exchange heat directly through a chilled water-cooled water heat exchanger. The rate of heat exchange between the chilled water loop and the cooling water loop can be controlled by controlling the bypass loop of the cooling water.

[0059] Table 1. Introduction to Seven Types of Chiller Unit Faults

[0060]

[0061] Table 2. Introduction to Chiller Unit Variables

[0062]

[0063]

[0064] Table 3. Introduction to Four Operating Conditions of Chiller Units

[0065]

[0066] The ASHRAE RP-1043 public dataset collected data on nine operating conditions of chiller units based on different evaporator outlet water temperatures and condenser inlet water temperatures. This invention selected four of these operating conditions for experimentation. Table 3 describes the four operating conditions of the chiller unit. Each operating condition includes seven fault data points, and data for each fault was collected from 64 variables. 400 samples were selected from each fault for each of the four operating conditions, resulting in 2800 samples for each condition. In the experiment, the data for operating condition 1 was the labeled source domain, the data for operating conditions 2 and 3 were the unlabeled source domain, and the data for operating condition 4 was the target domain. The source domain was used to train the domain generalization network, and the target domain was used to test the domain generalization network.

[0067] Based on the above description, and according to the invention, the specific process is as follows:

[0068] A. Offline modeling stage:

[0069] 1) Obtain the training dataset. The training dataset consists of data from three operating conditions of the chiller unit. Seven types of fault data were collected for each of the three operating conditions. Data from one operating condition is labeled, while data from the other two operating conditions is unlabeled. The number of data points for each type of fault collected is the same. The training dataset can be represented as X = [x1, x2, ..., x...]. 8400 ] T ∈R 8400×64 8400 training samples were selected, with a sample dimension of 64;

[0070] 2) The Dickie-Fowler test was performed on the 64 variables in the training dataset, and they were divided into 26 stationary variables. and 38 non-stationary variables The Dickie-Fowler test principle is as follows: Time series y t The first-order autoregressive process is shown in equation (1). If ρ = 1, {ε t} Independent and identically distributed, and Eε t =0, Var(ε) t )=σ 2 <∞, then y t For a non-stationary time series, if |ρ| < 1, then y t It is a stationary time series;

[0071] y t =ρy t-1 +ε t (1)

[0072] 3) For non-stationary variables The Johansen test is performed, and the non-stationary variables that are linear combinations of stationary variables are identified based on the maximum characteristic statistic in the test results. The Johansen test uses a vector autoregression model to determine whether there is a long-term cointegration relationship among multiple variables. The vector autoregression model is shown in formula (2).

[0073] y t =A1y t-1 +A2y t-2 +…+A n y t-n +Bx t +e t (2)

[0074] Among them, y t It is an endogenous variable column vector, x t It is a vector of exogenous variables, A1, A2, ..., A n B and E are the coefficient matrices of equal estimation. t It is an error vector, and the error variables within the error vector are correlated. The error variables do not exhibit autocorrelation. The error vector is related to y. t ,y t-1 ,…,y t-n ,x t If they are uncorrelated, then nonstationary variables can be categorized into nonstationary variables with cointegration relationships. and nonstationary variables that do not have cointegration relationships

[0075] 4) Convert non-stationary variables with cointegration relationships into stationary variables, as shown in formula (3);

[0076]

[0077] 5) Change variable X s ∈R 8400×26 , and The input variables are obtained by fusion. The fusion method used is concatenate;

[0078] 6) Input data Perform maximum and minimum standardization, as shown in formula (4);

[0079]

[0080] Where, x min Each value in the vector is the minimum value of each variable, x. max Each value in the vector is the maximum value of each variable.

[0081] 7) Input the data after the maximum and minimum normalization into the domain generalization network. The domain generalization network in this invention consists of four parts: feature extractor, domain classifier, classifier and convolutional decoder. The feature extractor extracts the domain invariant fault features as shown in formula (5).

[0082] h i =f(W*x i +b i (5)

[0083] Where, x i Let b represent the i-th sample in the input data, W represent the weight matrix of the convolution kernel during feature extraction, and b represent the weight matrix of the convolution kernel. i The bias in the feature extraction process is represented by *, the convolution operation is represented by f(·), and the encoding process is represented by h. i This represents the hidden layer features extracted by the feature extractor;

[0084] 8) The domain classifier consists of fully connected layers, and the calculation process is shown in formula (6);

[0085]

[0086] Where y is the output of the current layer, f d Represents the ReLU activation function, Σ represents the weighted summation over all inputs, and w d Represents the weights of the fully connected layer. The loss function of the domain classifier, which represents the bias term, is shown in Equation (7).

[0087]

[0088] Among them, the number of domain categories is 3, i_tag d The domain label value representing a certain category, i_pre d This represents the probability value of the output.

[0089] 9) The classifier consists of fully connected layers, and the calculation process is shown in formula (8);

[0090]

[0091] Among them, y c It is the output of the current layer, f c Represents the ReLU activation function, Σ represents the weighted summation over all inputs, and w c b represents the weights of the fully connected layer. c The loss function of the classifier, which represents the bias term, is shown in formula (9).

[0092]

[0093] Among them, the number of fault categories is 7, i_tag c The label value representing a certain category, i_pre c This represents the probability value of the output.

[0094] 10) The convolutional decoder consists of convolutional layers, and the calculation process is shown in formula (10);

[0095]

[0096] in, This represents the input data for reconstruction. This represents the weight matrix of the convolution kernel during the input data reconstruction process. The input data reconstruction process is represented by the bias, * represents the convolution operation, g(·) represents the decoding process, and the loss function for input data reconstruction is shown in formula (11).

[0097]

[0098] The number of input data samples is 8400;

[0099] 11) Utilizing the loss function Loss = L + loss of a domain generalization network c +loss d Train the model;

[0100] B. Testing Phase:

[0101] 1) Obtain the test dataset. The test dataset consists of data from the fourth operating condition of the chiller unit. This condition collected data on seven different faults of the chiller unit, with the same number of data points for each fault type. The test dataset can be represented as follows: The sample size is 2800, and the sample dimension is 64, meaning that each sample contains 64 variables.

[0102] 2) Based on the Dickey-Fowler test results in A1, X t Divide into stationary variables Non-stationary variables

[0103] 3) According to the Johansen test in A2, non-stationary variables... Non-stationary variables can be categorized into those with cointegration relationships. and nonstationary variables that do not have cointegration relationships

[0104] 4) Based on the VAR model in A2, non-stationary variables with cointegration relationships are converted into stationary variables, as shown in formula (12);

[0105]

[0106] in, For the non-stationary variable B after cointegration transformation to a stationary state, t The coefficient matrix is ​​estimated by the vector autoregression model.

[0107] It is a cointegrated vector;

[0108] 5) Change variable X ts ∈R 2800×26 , and Variables obtained by fusion The fusion method used is concatenate;

[0109] 6) According to formula (4), adjust the variables Perform maximum and minimum standardization;

[0110] 7) Input the standardized data after maximum and minimum values ​​into the domain generalization network to determine the fault category.

[0111] The above steps constitute the specific application of this method for fault diagnosis of chiller units. To verify the effectiveness of this invention in fault diagnosis of multi-condition chiller units, experiments were conducted using the publicly available dataset from the ASHRAE RP-1043 project. This experiment evaluated the performance of the method using fault diagnosis accuracy, recall, precision, and F1 score. True positives (TP) and true negatives (TN) indicate that the sample was correctly diagnosed; similarly, false positives (FP) and false negatives (FN) indicate that the sample was incorrectly diagnosed. The definitions of Accuracy, Recall, Precision, and F1 score are shown in formulas (13)-(16). Figure 2 Table 4 shows the confusion matrix for diagnosing seven types of faults using the method of this invention. The fault diagnosis accuracy of the method for these seven faults is above 94%, verifying the effectiveness of the method. Table 5 and... Figure 3 The table presents comparative experimental results between the method of this invention and three other methods. Analysis shows that the method of this invention achieved the best fault diagnosis results, verifying its superiority. (Table 6 and...) Figure 4To demonstrate the ablation experiment results of the method of this invention, the methods NSSDG, Domain Adversarial Neural Networks (DANN), and Domain Adversarial Network Incorporating Convolutional Autoencoder (DANICAE) were compared. The NSSDG method achieved the best result, further verifying the effectiveness of the method of this invention.

[0112]

[0113] Table 4. Fault diagnosis accuracy of NSSDG for seven types of faults.

[0114]

[0115] Table 5. Accuracy of DANN, IEDGNet, DACNN, and NSSDG for Faults

[0116]

[0117] Table 6. Accuracy of DANN, DANICAE, and NSSDG for Faults

[0118]

[0119]

Claims

1. A multi-condition chiller unit fault diagnosis method based on non-stationary semi-supervised domain generalization networks, characterized in that... Includes the following steps: A. The offline modeling phase consists of four parts. A1 uses the Dickie-Fowler test to classify the variables in the chiller unit training dataset into stationary and non-stationary variables; A2 performs cointegration analysis on non-stationary variables and calculates the cointegration matrix, thus stabilizing non-stationary variables with cointegration relationships. A3 integrates the non-stationary variables after stabilization, the stationary variables, and the non-stationary variables without cointegration to obtain the input data for the domain generalization network; A4 uses the input data obtained in A3 to train the domain generalization network; B. The testing phase consists of four parts. Based on the Dickie-Fowler test results in A1, B1 classifies the variables in the chiller unit test dataset into stationary and non-stationary variables; B2 uses the cointegration analysis matrix in A2 to stabilize non-stationary variables with cointegration relationships; B3 combines the stabilized non-stationary variables, stationary variables, and non-stationary variables without cointegration relationships to obtain the test data; B4 inputs the test data obtained in B3 into the model trained by A4 to obtain the test results; The A1 specifically includes the following steps: A11) Obtain the training dataset. The training dataset consists of data from three operating conditions of the chiller unit. Seven types of fault data were collected for each of the three operating conditions. Data from one operating condition is labeled, while data from the other two conditions is unlabeled. The number of data points for each type of fault collected is the same. The training dataset can be represented as follows: ,in, Indicates the number of samples. This represents the sample dimension, i.e., the number of variables contained in each sample; A12) For the training dataset The variables were subjected to the Dickie-Fowler test and divided into stationary variables. Non-stationary variables ,in, The Dickie-Fowler test principle is as follows: Time series The first-order autoregressive process is shown in equation (1): （1） in, The correlation coefficient, For error, Representative seeking The mean, Representative seeking The variance; if , Independent and identically distributed, and ,but For non-stationary time series, if ,but It is a stationary time series; The A4 specifically includes the following steps: A41) Input the data after the maximum and minimum normalization into the domain generalization network. The domain generalization network consists of four parts: feature extractor, domain classifier, classifier and convolutional decoder. The feature extractor extracts the domain invariant fault features as shown in formula (5). （5）； in, Indicates input data, This represents the weight matrix of the convolution kernel during feature extraction. This represents the bias in the feature extraction process. This represents the convolution operation. Represents the encoding process. This represents the hidden layer features extracted by the feature extractor; The A42) domain classifier consists of fully connected layers, and the calculation process is shown in formula (6); （6）； in, It is the output of the current layer. Represents the ReLU activation function. This represents a weighted summation of all inputs. Represents the weights of the fully connected layer. The loss function of the domain classifier, which represents the bias term, is shown in formula (7). （7）； in, Represents the number of domain categories, The domain label value representing a certain category. This represents the probability value of the output. The A43) classifier consists of fully connected layers, and the calculation process is shown in formula (8); （8）； in, It is the output of the current layer. Represents the ReLU activation function. This represents a weighted summation of all inputs. Represents the weights of the fully connected layer. The loss function of the classifier, which represents the bias term, is shown in formula (9). （9）； in, Represents the number of fault categories. The label value representing a certain category. This represents the probability value of the output. A44) The convolutional decoder consists of convolutional layers, and the calculation process is shown in formula (10); （10） in, This represents the input data for reconstruction. This represents the weight matrix of the convolution kernel during the input data reconstruction process. This represents the bias during the input data reconstruction process. This represents the convolution operation. The loss function for reconstructing the input data, representing the decoding process, is shown in Equation (11). （11）； in, This represents the number of samples in the input data; A45) Utilizing the loss function of a domain generalization network Train the model.

2. The multi-condition chiller unit fault diagnosis method based on non-stationary semi-supervised domain generalization network according to claim 1, characterized in that: The A2 specifically includes the following steps: A21) For nonstationary variables The Johansen test is performed, and the non-stationary variables that are linear combinations of stationary variables are identified based on the maximum characteristic statistic in the test results. The Johansen test uses a vector autoregression model to determine whether there is a long-term cointegration relationship among multiple variables. The vector autoregression model is shown in formula (2). （2）； in, It is an endogenous variable column vector. It is an exogenous variable vector. ,and It is an equal-estimated coefficient matrix. It is an error vector, and the error variables within the error vector are correlated. The error variables do not exhibit autocorrelation. The error vector and If they are uncorrelated, then nonstationary variables can be categorized into nonstationary variables with cointegration relationships. and nonstationary variables that do not have cointegration relationships ; A22) Convert non-stationary variables with cointegration relationships into stationary variables, as shown in formula (3); （3）； in, For non-stationary variables that have been stabilized through cointegration transformation, The coefficient matrix is ​​estimated by the vector autoregression model. It is a cointegrated vector.

3. The multi-condition chiller unit fault diagnosis method based on non-stationary semi-supervised domain generalization network according to claim 1, characterized in that: The A3 specifically includes the following steps: A31) Variables , and The input variables are obtained by fusion. The fusion method used is concatenate. ; A32) Input data Perform maximum and minimum standardization, as shown in formula (4); （4）； in, Each value in the vector is the minimum value of each variable. Each value in the vector is the maximum value of each variable.

4. The multi-condition chiller unit fault diagnosis method based on non-stationary semi-supervised domain generalization network according to claim 1, characterized in that: The B1 specifically includes the following steps: B11) Obtain the test dataset. The test dataset consists of data from the fourth operating condition of the chiller unit. This condition collected data on seven different faults of the chiller unit, with the same number of data points for each fault type. The test dataset can be represented as follows: ,in, Indicates the number of samples. This represents the sample dimension, i.e., the number of variables contained in each sample; B12) Based on the Dickie-Fowler test results in A1, Divide into stationary variables Non-stationary variables ,in, .

5. The multi-condition chiller unit fault diagnosis method based on non-stationary semi-supervised domain generalization network according to claim 1, characterized in that: The B2 specifically includes the following steps: B21) According to the Johansen test in A2, non-stationary variables... Non-stationary variables can be categorized into those with cointegration relationships. and nonstationary variables that do not have cointegration relationships ; B22) Based on the VAR model in A2, non-stationary variables with cointegration relationships are converted into stationary variables, as shown in formula (12); （12）； in, For non-stationary variables that have been stabilized through cointegration transformation, The coefficient matrix is ​​estimated by the vector autoregression model. It is a cointegrated vector.

6. The multi-condition chiller unit fault diagnosis method based on non-stationary semi-supervised domain generalization network according to claim 1, characterized in that: The B3 specifically includes the following steps: B31) Variables , and Variables obtained by fusion The fusion method used is concatenate. ; B32) According to formula (4), the variables Perform maximum and minimum standardization.

Images