Water chilling unit fault diagnosis method in data sample imbalance state

Abstract

The invention discloses a water chilling unit fault diagnosis method in a data sample imbalance state. Aiming at the problem of data imbalance caused by insufficient fault sample collection in actual engineering, the invention provides a data enhancement model for improving the generative adversarial network, and by introducing label information and adding an orthogonal gradient penalty mechanism, the loss functions of a generator and a discriminator in the model are optimized, so that the reliability of the model is improved. The generative adversarial network training process is more stable and quicker, and samples generated by the model are guided to be more similar to real sample distribution. And meanwhile, considering that the water chilling unit is a complex multivariable nonlinear system and all devices have a high coupling relationship, introducing a graph convolutional network as a diagnostic model to mine associated information among samples, and constructing enhanced balance data into graph structure data as input, so that the model learns deeper feature information. The method effectively solves the problem of data imbalance, and achieves the accuracy of fault diagnosis of the water chilling unit when the data proportion is seriously unbalanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of water chiller fault detection, and particularly relates to a water chiller fault diagnosis method under a data sample unbalanced state. BACKGROUND

[0002] Exploring the design of clean, low-carbon, safe and efficient modern energy systems is a hot topic in today's building environment. In commercial and residential buildings, heating, ventilation and air conditioning systems are the main source of energy consumption, and water chillers, as one of the key components of heating, ventilation and air conditioning systems, have an operating energy consumption of up to 40% of the total system energy consumption. If a fault occurs, the energy waste will increase by nearly 30%. Therefore, achieving accurate diagnosis of water chiller faults and quickly restoring normal operation of the function is of great significance for energy saving and prolonging the life cycle of the device.

[0003] In recent years, the main method for water chiller fault diagnosis is deep learning algorithm, and convolutional neural network, recurrent neural network and other methods are widely used. However, existing methods all need a large amount of data as support, but in the actual industrial environment, the water chiller is usually in normal operation, and it is difficult for the sensor to collect all the fault type data, thereby causing a class imbalance state of sufficient normal samples and few fault samples, and further causing the model to judge more inclined to the state type with higher data proportion, making the diagnosis result distorted. At the same time, the water chiller is a complex multi-variable nonlinear system, and there is a high coupling relationship between each device. Most of these methods focus on the extraction of local features, and consider the data samples as independent and identically distributed for feature learning, ignoring the internal relationship between the samples, thereby failing to make accurate diagnosis. In fact, data samples of the same type often have shared similar features, and by mining the association between them, the typical feature information of the type can be better learned. As for data samples of different types, there may be some implicit relationships between the data, which may not be directly reflected in the original features, but can be mined through the interaction and information transmission between samples to learn deep features for distinguishing different types. For example, one-dimensional convolutional neural network (1DCNN) can effectively extract local features of data by relying on the convolution kernel structure, but its limited receptive field cannot capture global dependency information. Long short-term memory network (LSTM) can model sequence dependency within a certain time step through memory cells and gating mechanisms, but it is also limited by the time parameter and cannot extract global features. The unique message passing mechanism of graph convolution network (GCN) has the ability to aggregate features between data under the graph structure, each node obtains its feature information from adjacent nodes and fuses it with its own feature information to update the feature, thereby learning more complex and deep feature representation, providing a new idea for water chiller fault diagnosis.

[0004] Meanwhile, the generative adversarial network (GAN) is widely used to enhance data samples to solve the data imbalance problem due to its ability to extract data distribution characteristics. However, the traditional GAN has problems such as unstable training and mode collapse, especially when the number of minority class samples is limited, resulting in low quality, lack of diversity and authenticity of the generated minority class samples. The Wasserstein distance (Wasserstein distance), also known as the Earth Mover's Distance (EMD), is a distance measure used to compare the differences between two probability distributions in the theory of measurement, which can improve the stability of the GAN model. ADASYN (Adaptive Synthetic Sampling) is an oversampling method that generates new samples by calculating the proportion of majority class samples around each minority class sample.

[0005] The present application aims at the data imbalance problem caused by insufficient collection of fault samples in actual engineering. By fusing the ADASYN oversampling algorithm, the improved generative adversarial network (ACWGAN-OGP minority class sample expansion model) and the graph convolution network, the data imbalance and the poor diagnosis performance of high coupling data are solved. The ACWGAN-OGP minority class sample expansion model introduces label information and adds an orthogonal gradient penalty mechanism, optimizes the loss function of the generator and the discriminator in the model, makes the training process of the generative adversarial network more stable and faster, and guides the distribution of the generated samples to be more similar to that of the real samples. At the same time, considering that the water chiller unit is a complex multivariate nonlinear system with high coupling relationship between devices, in order to mine the association information between samples, the graph convolution network is introduced as a diagnostic model, and the enhanced balanced data is constructed as graph structure data as input, so that the model learns deeper feature information. The data imbalance problem is effectively solved, and when the data proportion is seriously unbalanced, the diagnostic accuracy is still high, realizing the effective diagnosis of the water chiller unit fault. SUMMARY

[0006] To address the challenges of high coupling and class imbalance in chiller unit data samples, which limit the diagnostic performance of models, this invention proposes a fault diagnosis method for chiller units under imbalanced data conditions. The method utilizes the ADASYN oversampling algorithm to presample minority class samples, overcoming the limitation of GAN model performance on sample size. It replaces the JS divergence in traditional GANs with Wasserstein distance as the loss function to measure the difference between two distributions, improving model stability. Label information is introduced to guide the generation of samples of specific classes, and an orthogonal gradient penalty mechanism optimizes the model's loss function, constraining its magnitude and direction during updates. This allows for the explicit calculation of the gradient magnitude and its direction towards the target data distribution. Analyzing the coupling point line connecting the generated and real data distributions yields the optimal discriminator, leading to faster convergence and improved sample quality. This improved model expands minority class samples to achieve data balance. Finally, it leverages the feature aggregation capability of GANs to extract feature information and uses focus loss to enhance the model's attention to difficult-to-classify minority class samples, improving its ability to identify minority class samples and thus enhancing fault diagnosis performance.

[0007] The present invention specifically includes the following steps:

[0008] Step 1: Collect chiller unit data under normal operating conditions and different fault conditions to form a labeled sample set. ;

[0009] Step 2: Process the sample set Preprocessing is performed to remove outliers and normalize the data, and finally, the data is divided into training sets. and test set ;

[0010] Step 3: Use the ADASYN oversampling algorithm on the training set. Oversampling makes the training set The number of minority class samples is equal to the number of majority class samples, and the generated minority class samples constitute the dataset. The dataset contains category feature labels that guide the generation of the data.

[0011] Step 4: Build the ACWGAN-OGP minority class sample augmentation model. The ACWGAN-OGP minority class sample augmentation model includes a generator and a discriminator, and uses the dataset... The generator takes the class feature labels as input to generate simulated data. The simulated data generated by the generator and the real data in the training set are then input into the discriminator to expand the minority class samples. The generated samples are then mixed with the real samples to form a new balanced training set, so that the number of samples in all classes is equal. The resulting balanced dataset is used as the input for the subsequent graph convolutional network diagnostic model.

[0012] Step five, according to the knowledge of graph theory, the balanced data set obtained in step four is used to construct a data graph structure, the graph structure includes nodes and edges connecting the nodes, the nodes represent samples and contain feature information; the distance matrix and the adjacency matrix of the graph structure are obtained through the nodes.

[0013] Step six, a GCN fault diagnosis model is constructed, the GCN fault diagnosis model includes two graph convolution layers and two fully connected layers. After normalizing the adjacency matrix in step five, the graph data in step five and the normalized adjacency matrix are input into the graph convolution layer, the feature aggregation of the node information and the connection relationship in the graph data is performed, the feature representation between the nodes is obtained, and the final node feature matrix is obtained after two layers of graph convolution. After the graph convolution layer, the node feature matrix is input into the fully connected layer, and the node feature matrix is converted into a probability form.

[0014] The classification loss is calculated using the focal loss function, and the GCN model weight is updated through back propagation and gradient descent algorithm, the Adam optimizer is used to optimize the loss function value, and the training is stopped when the training number reaches the set maximum training number.

[0015] Step seven, using the test set The fault diagnosis model obtained in step six is tested; the accuracy is used as an evaluation index, the fault diagnosis result of the test sample is output, and the effectiveness verification of the GCN fault diagnosis model trained by the balanced data set output by the ACWGAN-OGP minority class sample expansion model is completed.

[0016] Step eight, input the parameters of the cold water unit to be tested into the fault diagnosis model, judge whether the cold water unit to be tested has a fault and output the type of the fault.

[0017] Further, the step four is specifically: using a fully connected neural network to construct a generator and a discriminator, the fully connected neural network includes an input layer, a hidden layer and an output layer; using the Wasserstein distance instead of the JS divergence in the traditional GAN to measure the difference between two distributions, and using the orthogonal gradient penalty mechanism to optimize the loss function of the discriminator, by calculating the amplitude and direction of the gradient, analyzing the coupling point straight line between the connected generated data distribution and the real data distribution to obtain the optimal discriminator;

[0018] The data set The simulation data generated by the generator is input into the discriminator, and the discriminator outputs a probability value to determine whether the generated simulation data is real data, thereby guiding the generator to generate more realistic data.

[0019] Further, in the step five, the Euclidean distance measurement formula is used to calculate the distance value between each node and expressed in matrix form to obtain a distance matrix.

[0020] Further, in the step five, the k-nearest neighbor algorithm (KNN algorithm) is used to calculate the obtained distance matrix, and the N nodes closest to the current node are selected to determine whether there is an edge between the nodes, and if there is, the nodes are connected, and finally a complete graph data and a corresponding adjacency matrix are obtained.

[0021] Further, the activation function of each layer in the hidden layer uses ReLU or LeakyReLU, and the activation function of the output layer is usually Sigmoid or Tanh, and an optimizer is used to update the weight parameters through back propagation and gradient descent algorithm.

[0022] Further, the discriminator loss function is as follows:

[0023] ;

[0024] Wherein: , , ;

[0025] In the formula: represents the discriminator, represents the real sample, represents the generated sample, is a penalty coefficient, is a unit vector, is a balance coefficient between the real sample and the generated sample, is a linear interpolation point between the real data and the generated data, is the label information of , represents the interpolation distribution randomly selected interpolation samples , represents the prediction probability of the discriminator, represents the expectation of the output result.

[0026] Further, the generator loss function is as follows:

[0027]

[0028] In the formula: represents the generator, represents the oversampled minority class samples, represents the distribution of randomly sampling from the generated samples.

[0029] Further, the weight of the easily classified samples is reduced by the focal loss function, and two adjustable parameters are introduced, wherein the balance factor makes the model pay more attention to the minority class samples, and the focusing factor makes the model pay more attention to the difficult-to-classify samples with low prediction probability, which allocates greater weight to the difficult-to-distinguish and less sample data in the training data. The focal loss function formula is as follows:

[0030]

[0031] In the formula: represents the loss contribution of the sample, represents the probability value, represents the balance factor, is the focusing factor.

[0032] As preferred, the generator and the discriminator respectively use five and six fully connected layers.

[0033] Further, the method can identify various chiller fault types, including excessive lubricating oil, condenser fouling, refrigerant leakage, excessive refrigerant, non-condensable gas, insufficient cooling water, and insufficient chilled water.

[0034] The present application has the following beneficial effects:

[0035] 1. The present application aims at the data imbalance problem, improves the limitation of GAN model performance limited by the number of minority class samples through the ADASYN algorithm, and optimizes the loss function of the generative adversarial network using the Wasserstein distance and the orthogonal gradient penalty mechanism, enhances the stability of model training and improves the quality of generated data, makes the generated data distribution closer to the real data distribution, and effectively realizes the data expansion under the unbalanced samples;

[0036] 2.The application is aimed at the high coupling between sensors in the water chiller system and the characteristics that the data samples contain complex high-dimensional nonlinear relationships. Unlike traditional neural networks that treat data samples as independent distributed inputs, the method uses graph structure data as input, first constructs a graph structure through KNN algorithm, which can retain the global information and the relationship between each other, and then extracts features to realize fault diagnosis, providing a new method for water chiller fault diagnosis.

[0037] 3.The application combines data generation with fault diagnosis method, realizes accurate detection of each type of fault, can still realize high fault diagnosis accuracy when the imbalance rate of data samples is high, effectively realizes water chiller fault diagnosis, and has good application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The flowchart for the implementation of the application is shown in the figure.

[0039] Figure 2 The ACWGAN-OGP model structure and training process in the application are shown in the figure.

[0040] Figure 3 The GCN fault diagnosis model structure and processing process in the application are shown in the figure. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the embodiments and the accompanying drawings. The described embodiments are part of the embodiments of the application, not all.

[0042] A water chiller fault diagnosis method under the condition of data sample imbalance, the implementation process is shown in the figure. Figure 1 The specific steps are as follows:

[0043] Step one, collect water chiller data under normal operation state and different fault states to form a labeled sample set .

[0044] This embodiment selects the water chiller data in the 1043-RP project of American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) as the sample set , the water chiller data collection interval is 10s, a total of 64 parameters, of which 48 parameters are directly collected by sensors, including 29 temperature parameters, 7 valve opening parameters, 5 pressure parameters, 5 flow parameters, etc., and the remaining 16 parameters are calculated in real time by VisSim software.

[0045] The embodiment adopts seven types of fault data with the lightest level of fault and normal state data as a sample set, seven typical fault types are shown in Table 1, 2500 data of each state are selected and one-hot coding is used, and a total of 20000 sample data sets are constructed .

[0046] Table 1 Fault type

[0047]

[0048] Step two, preprocessing the sample set , removing abnormal values in the data and normalizing, and finally dividing into a training set and a test set ;

[0049] After preprocessing the sample set , 500 samples of each type are randomly selected from the sample set, a total of 4000 data to form a test set , in order to simulate the unbalanced data, four different imbalance rates are set, 2000 normal state data and a number of fault type data are selected from the remaining data in the sample set , to form a training set with different imbalance rates , the specific training set sample composition is shown in Table 2:

[0050] Table 2 Training set sample composition

[0051]

[0052] Step three, using ADASYN algorithm to oversample the training set , so that the number of minority class samples is equal to the number of majority class samples, and the generated minority class samples form a data set ;

[0053] The minority class samples in the training set with different imbalance rates made in step two are sampled by ADASYN algorithm to generate new synthetic samples, so that the number of samples is equal to the number of majority class samples, and the generated minority class samples are combined to form a data set .

[0054] Step four, build ACWGAN-OGP minority class sample expansion model, the ACWGAN-OGP minority class sample expansion model includes a generator and a discriminator.

[0055] Specifically, a full connection neural network is used to construct the generator and the discriminator, the full connection neural network includes an input layer, a hidden layer and an output layer; the Wasserstein distance is used to replace the JS divergence in the traditional GAN to measure the difference between two distributions, and the orthogonal gradient penalty mechanism is used to optimize the loss function of the discriminator, and the optimal discriminator is obtained by calculating the amplitude and direction of the gradient, analyzing the straight line of the coupling point between the generated data distribution and the real data distribution, and analyzing the straight line of the coupling point between the generated data distribution and the real data distribution.

[0056] The data set and the data label are input into the generator to generate simulated data through a forward propagation process (the generator learns from the distribution of the data set and adds the label information of the data to better guide the data generation); the simulated data generated by the generator and the real data in the training set are input into the discriminator, and the discriminator outputs a probability value to determine whether the generated simulated data is real data, thereby guiding the generator to generate more realistic data, keeping the parameters of the generator unchanged, updating the parameters of the discriminator, and training the discriminator; keeping the parameters of the discriminator unchanged, updating the parameters of the generator, and training the generator; repeating the above operations, alternately training the generator and the discriminator, updating the parameters through the respective loss functions, and finally reaching Nash equilibrium, so that the model converges and completes the training; the generated samples and the real samples are mixed to form a new balanced training set, at this time all the class sample quantities are equal, and a balanced data set is obtained as the input of the subsequent graph convolution network diagnosis model.

[0057] In this embodiment, the number of full connection layers of the generator and the discriminator is five and six respectively, the number of neurons of each layer of the generator is 1024, 512, 256, 128 and 64, the number of neurons of each layer of the discriminator is 128, 256, 128, 64, 32 and 1, the activation function is ReLU, the batch size is 128, the maximum number of iterations is 10000, the learning rate is 0.0002, and the optimizer is Adam.

[0058] The activation function of each layer in the hidden layer can use ReLU or LeakyReLU, and the activation function of the output layer is usually Sigmoid or Tanh to ensure that the output value is within a reasonable range. At the same time, the weight parameter is updated through the back propagation and gradient descent algorithm, and the Adam optimizer, RMSProp optimizer, etc. can be used.

[0059] The loss function of the discriminator in this embodiment is as follows:

[0060] ;

[0061] Wherein: , , ;

[0062] In the formula: represents the discriminator, represents the real sample, represents the generated sample, is a penalty coefficient, is a linear interpolation point between the real data and the generated data, is the label information of represents that an interpolation sample is randomly selected in the interpolation distribution , represents the prediction probability of the discriminator, represents the expectation of the output result.

[0063] The updating mode of the discriminator parameter is as follows:

[0064] ;

[0065] ;

[0066] In the formula: represents the algorithm learning rate, represents the sample quantity, represents the real sample, represents the generated sample, represents the gradient of the discriminator.

[0067] The loss function of the generator in the embodiment is as follows:

[0068] ;

[0069] In the formula: represents the generator, represents the oversampled minority class sample, represents a distribution randomly sampled from the generated sample.

[0070] The updating mode of the generator parameter is as follows:

[0071] ;

[0072] ;

[0073] In the formula: represents the algorithm learning rate, represents the gradient of the generator.

[0074] ​The data samples generated after the model are added to the original training set to obtain a balanced data set, and each type of data in the balanced data set is 2000, and a new training set of 16000 is composed.

[0075] Step five, according to the knowledge of graph theory, the balanced data set obtained in step four is used to construct a data graph structure, the graph structure includes nodes and edges connecting the nodes, the nodes represent samples and contain feature information; The distance matrix and the adjacency matrix of the graph structure are obtained through the nodes.

[0076] Specifically:

[0077] Step 5-1, each data is a node in the graph structure, the node has sample features, the distance values between nodes are calculated using the Euclidean distance measurement formula, and are expressed in matrix form, as follows The distance matrix of the node in the graph structure is represented as follows:

[0078] ;

[0079] In the formula: Euclidean distance between the first data and the first data, and represent the codes of each data.

[0080] Step 5-2, use KNN algorithm to construct the correlation graph: calculate the distance matrix obtained by KNN algorithm, select the nearest N nodes to the current node, judge whether there is an edge between the nodes, if there is an edge, connect the nodes, and finally obtain the complete graph data and the corresponding adjacency matrix;

[0081] The adjacency matrix is an n x n square matrix, represented as .

[0082] Among them, represents the connection relationship between the first node and the first node, when there is an edge between the nodes, is 1, when there is no edge between the nodes, is 0.

[0083] Step six, construct a GCN fault diagnosis model, such as Figure 3As shown, the GCN fault diagnosis model includes two graph convolution layers and two fully connected layers. After the normalization operation is performed on the adjacency matrix in step five, the graph data in step five and the normalized adjacency matrix are input into the graph convolution layer, the feature of each node information and connection relationship in the graph data is aggregated, the feature representation between nodes is obtained, and the final node feature matrix is obtained after two layers of graph convolution. After the graph convolution layer, the node feature matrix is input into the fully connected layer, the first layer of the fully connected layer uses the ReLU activation function, the second layer of the fully connected layer uses the Softmax function, and the node feature matrix is converted into a probability form.

[0084] The classification loss is calculated using the focal loss function, and the model weight is updated through the back propagation and gradient descent algorithm, the Adam optimizer is used to optimize the loss function value, and the training is stopped when the number of training reaches the set maximum number of training.

[0085] Each row of the node feature matrix represents a feature vector of a node, the number of rows is the number of data samples, and the number of columns is the number of extracted features.

[0086] In this embodiment, the number of channels of the graph convolution layer is 64 and 128 respectively, the number of neurons of the fully connected layer is 256 and 8, the activation function is ReLU, the learning rate is 0.01, the weight decay rate is 0.001, and the maximum number of iterations is 300.

[0087] The feature extraction process of the graph convolution layer of the GCN fault diagnosis model is as follows: ; ;

[0088] In the formula: denotes the normalized adjacency matrix, denotes the adjacency matrix of the degree matrix, denotes the activation function, denotes the feature matrix output by the layer of graph convolution, denotes the weight matrix of the layer of graph convolution.

[0089] The conversion of the node feature matrix into a probability form is represented as , and the predicted category is obtained according to the probability value corresponding to each node; the calculation formula of the Softmax function is as follows: ;

[0090] In the formula: denotes the probability value, denotes the number of samples, denotes the feature of the sample.

[0091] Traditional classification models use cross-entropy as the loss function, and each class sample in the data is given the same attention. For unbalanced data sets, using focal loss instead of cross-entropy loss function can increase the misclassification weight of minority class samples and reduce the misclassification weight of majority class samples, thereby strengthening the model's attention to minority class samples to calculate the classification loss.

[0092] By reducing the weight of easily classified samples through the focal loss function, two adjustable parameters are introduced. The balance factor makes the model pay more attention to minority class samples, and the focus factor makes the model pay more attention to difficult-to-classify samples with low prediction probability, which allocates greater weight to difficult-to-distinguish and less sample data in the training data. The focal loss function formula is as follows:

[0093] ;

[0094] In the formula: represents the loss contribution of the sample, represents the probability value, represents the balance factor, is the focus factor.

[0095] Step seven, using the test set Test the fault diagnosis model obtained in step six; use accuracy as the evaluation index, output the fault diagnosis result of the test sample, and complete the effectiveness verification of the GCN fault diagnosis model trained by the balanced data set output by the ACWGAN-OGP minority class sample expansion model.

[0096] Step eight, input the parameters of the cold water unit to be tested into the fault diagnosis model, judge whether the cold water unit to be tested has a fault, and output the type of fault.

[0097] To verify the data augmentation ability of the ACWGAN-OGP model and the fault diagnosis ability of the GCN fault diagnosis model trained by the balanced data set output by the ACWGAN-OGP minority class sample expansion model (ACWGAN-OGP-GCN method) under unbalanced data in this embodiment, the final diagnosis results of the method and other methods obtained by using the test set are shown in Tables 3 and 4. As can be seen from Table 3, when using the same classifier for testing, the accuracy of the data set after data augmentation by ACWGAN-OGP is the highest, which is obviously better than other models, proving its strong data generation ability. As can be seen from Table 4, the diagnosis accuracy will decrease as the imbalance rate increases, and the accuracy of the comparative method decreases significantly under high imbalance rate, while the accuracy of the ACWGAN-OGP-GCN method of the present application is higher than that of the comparative model under all imbalance rates, and the accuracy is still 91.2% when the imbalance rate reaches 200:1, which shows good diagnosis performance in the data imbalance problem.

[0098] Table 3 Accuracy under different data augmentation methods

[0099]

[0100] Table 4 Diagnosis accuracy of each model

[0101]

[0102] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, those skilled in the art should understand that the discussion of any embodiment above is only exemplary and is not intended to imply that the scope of the present disclosure (including claims) is limited to these examples. Under the idea of the present application, the parameters (such as network layers) and hyperparameters (such as learning rate) in the above embodiments or different embodiments can be modified according to the actual situation, and there are many other changes of different aspects of the present application as described above. In order to be brief, they are not provided in details.

[0103] The embodiments of the present application are intended to cover all such alternatives, modifications and variations falling within the broad scope of the appended claims, therefore, any technical solutions obtained by logical analysis, reasoning or limited experiments according to the concept of the present application on the basis of the prior art should be within the protection scope determined by the claims.

Claims

1. A method for diagnosing faults in chiller units under unbalanced data sample conditions, characterized in that: Specifically, the following steps are included: Step 1: Collect chiller unit data under normal operating conditions and different fault conditions to form a labeled sample set. ; Step 2: Process the sample set Preprocessing is performed to remove outliers and normalize the data, and finally, the data is divided into training sets. and test set ; Step 3: Use the ADASYN oversampling algorithm on the training set. Oversampling makes the training set The number of minority class samples is equal to the number of majority class samples, and the generated minority class samples constitute the dataset. The dataset contains category feature labels that guide the generation of the data; Step 4: Build the ACWGAN-OGP minority class sample augmentation model. The ACWGAN-OGP minority class sample augmentation model includes a generator and a discriminator, and uses the dataset... The generator takes the class feature labels as input to generate simulated data. The simulated data generated by the generator and the real data in the training set are input into the discriminator to expand the minority class samples. The generated samples are mixed with the real samples to form a new balanced training set, so that the number of samples in all classes is equal. The resulting balanced dataset is used as the input for the subsequent graph convolutional network diagnostic model. Step 5: Based on graph theory, construct a data graph structure using the balanced dataset obtained in Step 4. The graph structure includes nodes and edges connecting the nodes. Nodes represent samples and contain feature information. Obtain the distance matrix and adjacency matrix of the graph structure through the nodes. Step 6: Construct the GCN fault diagnosis model. The GCN fault diagnosis model includes two graph convolutional layers and two fully connected layers. After normalizing the adjacency matrix in Step 5, input the graph data and the normalized adjacency matrix in Step 5 into the graph convolutional layer. Perform feature aggregation on the node information and connection relationships in the graph data to obtain the feature representation between nodes. After two layers of graph convolution, the final node feature matrix is ​​obtained. After passing through the graph convolutional layer, the node feature matrix is ​​input into the fully connected layer, which converts the node feature matrix into a probabilistic form. The focus loss function is used to calculate the classification loss, and the weights of the GCN model are updated through backpropagation and gradient descent. The Adam optimizer is used to optimize the loss function value, and the training stops when the maximum number of training iterations is reached. Step 7: Utilize the test set Test the fault diagnosis model obtained in step six; use accuracy as the evaluation index, output the fault diagnosis results of the test samples, and complete the validity verification of the GCN fault diagnosis model trained on the balanced dataset output by the ACWGAN-OGP minority class sample augmentation model. Step 8: Input the parameters of the chiller unit under test into the fault diagnosis model, determine whether the chiller unit under test has a fault, and output the type of fault.

2. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 1, characterized in that: Step four specifically involves: constructing a generator and a discriminator using a fully connected neural network, which includes an input layer, a hidden layer, and an output layer; using Wasserstein distance instead of the JS divergence in traditional GANs to measure the difference between the two distributions; optimizing the loss function of the discriminator using an orthogonal gradient penalty mechanism; and obtaining the optimal discriminator by calculating the magnitude of the gradient and its direction toward the target data distribution and analyzing the straight line connecting the coupling point between the generated data distribution and the real data distribution. Dataset The data labels are input into the generator to generate simulated data through a forward propagation process; the simulated data generated by the generator and the real data in the training set are respectively input into the discriminator, and the discriminator outputs a probability value to determine whether the generated simulated data is real data, thereby guiding the generator to generate more realistic data; keeping the parameters of the generator unchanged, the parameters of the discriminator are updated to train the discriminator; Keeping the discriminator's parameters unchanged, update the generator's parameters and train the generator; repeat the above operation to train the generator and discriminator alternately, continuously updating the parameters through their respective loss functions, and finally reaching Nash equilibrium, after which the model converges and completes training; mix the generated samples with real samples to form a new balanced training set, at which point the number of samples of all categories is equal, and the balanced dataset is used as the input for the subsequent graph convolutional network diagnostic model.

3. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 1, characterized in that: In step five, the distance between each node is calculated using the Euclidean distance metric formula and expressed in matrix form to obtain the distance matrix.

4. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 1, characterized in that: In step five, the distance matrix obtained is calculated using the k-nearest neighbor algorithm to select the N nodes that are closest to the current node, and it is determined whether there are edges between the nodes. If there are, the nodes are connected, and finally the complete graph data and the corresponding adjacency matrix are obtained.

5. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 2, characterized in that: The hidden layers use ReLU or LeakyReLU as the activation function, while the output layers typically use Sigmoid or Tanh as the activation function. An optimizer is used to update the weight parameters through backpropagation and gradient descent algorithms.

6. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 2, characterized in that: The discriminator loss function is as follows: ； in: , , ; In the formula: Indicates the discriminator, Represents a real sample. Indicates the generation of samples, It is the penalty coefficient. It is a unit vector. It is the balance coefficient between real samples and generated samples. It is a linear interpolation point between real data and generated data. yes Tag information, Indicates the interpolation distribution Randomly select interpolation samples , This represents the predicted probability of the discriminator. This indicates the expected output result.

7. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 2, characterized in that: The generator loss function is as follows: ； In the formula: Represents a generator. This represents the minority class samples after oversampling. This represents the distribution of a random sample drawn from the generated sample.

8. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 1, characterized in that: The focus loss function reduces the weight of easily classified samples by introducing two adjustable parameters: a balance factor that makes the model pay more attention to minority class samples, and a focus factor that makes the model pay more attention to difficult-to-classify samples with low prediction probabilities. This assigns greater weight to data that is difficult to distinguish and has few samples in the training data. The formula for the focus loss function is as follows: ； In the formula: Indicates the loss contribution of the sample. Represents the probability value. Represents the balance factor. It is a focusing factor.

9. The method for diagnosing chiller unit faults under unbalanced data sample conditions as described in claim 1, characterized in that: The generator and discriminator use five and six fully connected layers, respectively.

10. The method for diagnosing chiller unit faults under data sample imbalance as described in any one of claims 1 to 9, characterized in that: The method can identify various types of chiller unit malfunctions, including excessive lubricating oil, condenser scaling, refrigerant leakage, excessive refrigerant, non-condensable gases, insufficient cooling water, and insufficient chilled water.

Images