A refrigerating unit fault diagnosis method based on semi-supervised contrast learning

Abstract

A kind of refrigerating unit fault diagnosis method based on semi-supervised contrast learning, method contains three parts, first, design refrigerating unit data enhancement technique carries out data enhancement to original data, data is combined with pair constraint, create labeled data set with pair constraint relationship.Then, construct sample pair prediction network based on contrast learning, train model using labeled data set with pair constraint relationship, learn the association between a pair of data for creating full data set with pair constraint relationship.Finally, construct semi-supervised equipment fault diagnosis network based on memory enhancement, from full data set with pair constraint relationship, the class information implied in data is grabbed, and a memory module is introduced to make the features extracted by semi-supervised equipment fault diagnosis network more distinctive.The method proposed in the application can effectively mine the internal correlation of unlabeled data of the same fault type and improve the performance of refrigerating unit equipment diagnosis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of refrigeration unit fault diagnosis and data mining, and relates to a refrigeration unit fault diagnosis method based on semi-supervised contrast learning. BACKGROUND

[0002] Heating, ventilation, and air conditioning (HVAC) systems are the main energy-consuming equipment in industrial and commercial areas, which can provide refrigeration, heating, ventilation, and other services. However, with the rapid expansion of industrial and commercial areas, the HVAC system is also growing, and the energy consumption is also increasing dramatically. Once the system fails, the working efficiency of the system will be reduced, and the energy consumption and carbon dioxide emissions will increase. Therefore, many studies have begun to focus on designing effective equipment fault diagnosis methods to timely detect equipment failures and reduce energy consumption. In the HVAC system, the refrigeration unit is the most energy-consuming and complex subsystem in the HVAC system, and accordingly, research on refrigeration unit fault diagnosis has become one of the main aspects of refrigeration equipment fault diagnosis.

[0003] Traditional refrigeration unit fault diagnosis methods use linear or probabilistic models designed by humans to analyze and compare the differences between fault parameters and normal operating parameters, such as linear discriminant analysis models and Bayesian models, to identify faults by capturing the differences between the two and some existing expert knowledge. This method requires a lot of human intervention, and the effectiveness of the model depends on the existing fault diagnosis knowledge reserve, and is not suitable for the diagnosis of complex systems. With the development of artificial intelligence technology, some equipment diagnosis research work based on neural networks and SVM and other nonlinear models has been proposed. These works can automatically perform fault diagnosis, such as a neural network-based equipment diagnosis model that automatically adjusts model parameters through the residual error feedback between labels and model predictions, so that the model eventually outputs accurate diagnosis results.

[0004] However, current deep learning-based refrigeration unit diagnosis methods require a large number of labeled data samples for model training. However, in practical applications, fault data is difficult to collect on a large scale. Existing deep algorithms are difficult to apply to actual refrigeration unit fault diagnosis, therefore, a new semi-supervised learning-based fault diagnosis method needs to be designed to effectively explore and analyze the internal correlation of unlabeled data of the same fault type. SUMMARY

[0005] To solve the above problems, the application provides a refrigeration unit fault diagnosis method based on semi-supervised contrast learning, a labeled data set with paired constraint relationship is created through a refrigeration unit fault data enhancement method; a sample pair prediction network based on contrast learning is constructed to learn the correlation between data in the same fault type from the labeled data set with paired constraint relationship and predict whether the unlabeled data pair belongs to the same category, and a full data set with paired constraint relationship is created; a semi-supervised equipment fault diagnosis network based on memory enhancement is constructed to mine the internal correlation of the same fault type data from the full data set with paired constraint relationship, and accurate refrigeration unit fault diagnosis under semi-supervised learning is realized.

[0006] The method comprises three parts: refrigeration unit fault data enhancement technology design, sample pair prediction network construction based on contrast learning, and semi-supervised equipment fault diagnosis network creation based on memory enhancement. First, the refrigeration unit fault data enhancement technology is designed to enhance the original fault data, solve the problem of small amount of unit fault data, and create a labeled data set with paired constraint relationship to provide data support for the following semi-supervised contrast learning; the sample pair prediction network based on contrast learning learns the correlation between a pair of data from the labeled data with paired constraint relationship, and predicts the relationship of the unlabeled data pair; the semi-supervised equipment fault diagnosis network based on memory enhancement analyzes and mines the internal correlation of the same fault type data from the full data set with paired constraint relationship created by the sample pair prediction network.

[0007] In order to achieve the above purpose, the technical scheme adopted by the application is:

[0008] A refrigeration unit fault diagnosis method based on semi-supervised contrast learning, comprising the following steps:

[0009] Step 1, design refrigeration unit fault data enhancement technology, first generate fault data using a generative adversarial network, then filter low-quality data using a variational autoencoder, and create a data set with paired constraint relationship based on the high-quality data filtered;

[0010] Step 2, construct a sample pair prediction network based on contrast learning, learn the correlation between a pair of labeled data, and predict whether all sample pairs in the data set belong to the same category, and create a full data set with paired constraints;

[0011] Step 3, construct a semi-supervised equipment fault diagnosis network based on memory enhancement, mine the internal correlation of the same fault type unlabeled data from the full data set with paired constraints, and enhance the discriminative nature of the learned feature representation using memory enhancement.

[0012] The method can mine the internal correlation of the same fault type unlabeled data, realize accurate refrigeration unit fault diagnosis under semi-supervised learning, effectively extract the hidden correlation in the same fault type device data, and improve the performance of semi-supervised refrigeration unit equipment diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 It is a refrigeration unit fault diagnosis method framework based on semi-supervised contrast learning.

[0014] Figure 2 It is a basic flowchart of the refrigeration unit fault diagnosis method based on semi-supervised contrast learning. DETAILED DESCRIPTION

[0015] The embodiments of the present application will be further described below in combination with the drawings.

[0016] As shown in Figure 1 and Figure 2 , a refrigeration unit fault diagnosis method based on semi-supervised contrast learning of the present application contains three parts, design of refrigeration unit fault data enhancement technology, construction of sample pair prediction network based on contrast learning, and construction of semi-supervised equipment fault diagnosis network based on memory enhancement. First, the refrigeration unit data enhancement technology is designed to enhance the fault data, the generative adversarial network is constructed to generate data, and the variational automatic encoding is used to filter low-quality data, and then the data is combined in pairs to create a labeled data set with paired constraint relationship. Then, the sample pair prediction network based on contrast learning is constructed, the sample pair prediction network is trained using the labeled data set with paired constraint relationship created, the association between a pair of labeled data is learned, the relationship between unlabeled data pairs is predicted, and a full data set with paired constraint relationship is created. Finally, the semi-supervised equipment fault diagnosis network based on memory enhancement is constructed to extract the class information hidden in the full data set with paired constraint relationship, and the memory module is introduced to enhance the discriminativeness of the semi-supervised equipment fault diagnosis network to extract features, and finally the accurate prediction of the data fault type is realized.

[0017] Each step is implemented as follows:

[0018] Step 1, refrigeration unit fault data enhancement;

[0019] Considering that it is difficult to obtain large-scale labeled refrigeration unit fault data in reality, and that using only a small amount of fault data to train the diagnosis model can easily lead to model overfitting, the original refrigeration unit fault data needs to be enhanced and combined in pairs to meet the requirements of contrast learning so that it can be applied to the sample pair prediction network based on contrast learning. The specific implementation is as follows:

[0020] 1.1 Fault data generation

[0021] The present application constructs a generative adversarial network to realize the generation of fault data, and uses adversarial learning to ensure that the synthetic data is as similar as possible to the original data. (G,D) As shown in formula (1):

[0022]

[0023] Where P represents the data distribution, G represents the generator, and D represents the discriminator. data x represents a sample sampled from real data, and x represents a sample sampled from generated data. G The generator is used to generate data from random noise, and its goal is to minimize the adversarial loss, while the discriminator is used to determine whether the data is real, and hopes to maximize the adversarial loss. The generator and the discriminator are in a state of antagonism during the training process, and the generative adversarial network converges when the two reach a Nash equilibrium.

[0024] In specific implementation, the present application uses conditional WGAN as a basic model for adversarial learning. Conditional WGAN introduces label information as a condition y, and uses the condition and random noise as the input of the generator to generate data, further improving the quality of data generation. The corresponding adversarial learning formula L (G,D) is:

[0025]

[0026] Given is the data generated by the generator The specific generative adversarial formula L (G,D) As shown in formulas (3) and (4):

[0027]

[0028]

[0029] Wherein, represents the data generated by the generator, represents the linear interpolation between the generated data and the real data, and ε is a constant between 0 and 1. represents The partial derivative of , λ represents a penalty coefficient, which is used to add gradient penalty constraints in the discriminator.

[0030] 1.2 Generated data screening

[0031] To ensure the quality of generated data, the present application further filters low-quality data by using a variational autoencoder to screen the generated data. The variational autoencoder is trained by reconstructing the samples randomly sampled from the generated data, and after training, the variational autoencoder reconstructs and decodes all data. If the reconstruction loss is higher than the set threshold, it is considered that the data quality is too low and does not conform to the real data distribution. The objective function L of training the variational autoencoder is (E,F) :

[0032]

[0033] where E and F represent the encoder and the decoder, respectively. F (z|x) represents the posterior distribution approximated by the decoder. E (z) represents the prior distribution of the latent variable z. KL(q F (z|x)||p E (z)) represents the KL distance between the q F (z|x) posterior distribution and the p E (z) prior distribution.

[0034] 1.3 Data set modification

[0035] In order to meet the subsequent training of the sample pair prediction network based on contrast learning, the data set needs to be modified to construct a labeled data set with pair constraint characteristics. The pair constraint is guided by the category information, and two samples belonging to the same category are considered as having pair constraint, and samples not belonging to the same category do not have pair constraint. Given a data set The labeled data set can be represented as where

[0036]

[0037] where (a, b) is any pair of samples in the data set. Because the binary relationship of belonging to the same category or belonging to different categories is symmetric, the sample pair satisfies the following properties

[0038]

[0039] After obtaining the labeled data set α with pair constraint relationship, the corresponding unlabeled data set β with pair constraint relationship is represented as β = χ 2 \ α.

[0040] Step 2, construct a sample pair prediction network based on contrast learning;

[0041] Considering that it is difficult for a multi-classification model to predict that unlabeled data belongs to a certain class, the application simplifies the multi-classification problem into a binary classification problem. A sample pair prediction network based on contrastive learning is trained using the constructed labeled data set with pair constraint. A pair of samples is taken as input, and the similarity of the pair of samples is output. In the training process, the representations of the samples are ensured to be close to each other, and the corresponding contrastive loss As shown in formula (8):

[0042]

[0043] Where d(a,b) represents the Euclidean distance between the feature representations of a pair of samples, d(a,b) = ||h(a)-h(b)||, and h represents an encoding network composed of multiple fully connected layers. In the formula The maximum operator is used to ensure that the value is non-negative.

[0044] Since the goal of the sample pair prediction network based on contrastive learning is to generate a similarity score of a pair of samples, rather than directly predicting whether the two samples belong to the same class, the application further predicts whether the two samples belong to the same class according to the distance value between the two samples. A constant threshold T is designed when predicting a pair of unlabeled data. When the distance between the feature representations of the two samples is less than the threshold, it is considered that the two samples belong to the same class.

[0045] Because the contrastive loss is only effective for the labeled data set with pair constraint, the sample pair prediction network based on contrastive learning needs to extract features from unlabeled data through a reconstruction loss. The encoding network h is used to encode the input fault data, which is mapped to the latent space. Then the latent feature is decoded by the decoding network f to realize the reconstruction of the data. By constructing such an encoding-decoding framework, it is ensured that the encoding network h can learn the intrinsic features of the complete data while predicting the classification. The formula of the reconstruction loss function L(x) is as follows:

[0046]

[0047] The total loss L of the sample pair prediction network based on contrastive learning pair The formula is as follows:

[0048]

[0049] After the training is completed, the sample pair prediction network based on contrastive learning is used to construct pair constraints for all data.

[0050]

[0051] Where T represents a constant threshold for controlling whether a pair of samples belong to the same class.

[0052] Step 3, constructing a memory-enhanced semi-supervised equipment fault diagnosis network;

[0053] After the sample pair prediction network based on contrast learning is used to predict the pair constraint of the unlabeled data, the memory-enhanced fault diagnosis network is designed, the full data set with pair constraint relationship is used for training, and the fault type prediction is realized. The specific steps are as follows:

[0054] 3.1 training using full data set with pair constraint relationship

[0055] The sample pair prediction network based on contrast learning is used to predict the pair constraint of the unlabeled data, so as to obtain the full data set with pair constraint relationship. Then, the semi-supervised equipment fault diagnosis network p is constructed, and the basic module includes a classification network composed of multiple full connection layers and a classifier. The softmax function is used in the specific implementation of the classifier. After the semi-supervised equipment fault diagnosis network predicts the fault type of a pair of samples in the full data set with pair constraint relationship, the probability p s (a,b) represents

[0056]

[0057] Wherein, k represents a certain category, K represents the number of categories of the data set, p k (a) represents the prediction probability of sample a about category k, p k (b) represents the prediction probability of sample b about category k.

[0058] Similarly, p d (x,y)=1-p s (x,y) is the probability that a pair of samples belongs to different categories. In order to learn the category consistency of the same class data by using pair constraint, the error classification loss is designed to minimize the number of sample pairs that are misclassified.

[0059]

[0060] After the semi-supervised equipment fault diagnosis network is trained by using the full data set with pair constraint relationship, the category information hidden in the data is learned, and the prediction of the sample about the fault type is completed.

[0061] 3.2 create memory module

[0062] In order to make the features extracted by the semi-supervised equipment fault diagnosis network more discriminative and improve the prediction accuracy, a new memory module is designed to aggregate the discriminative feature representation and the fault classification results generated in the training process of the semi-supervised equipment fault diagnosis network, and then the learned class-related knowledge is transmitted to the next semi-supervised equipment fault diagnosis network training to enhance the difference between the features of different classes of unlabeled samples learned. The memory module takes the data feature representation and the fault classification results generated in the training process of the semi-supervised equipment fault diagnosis network as input, and constructs a key-value pair representation set. Among them, the key of the memory module represents the feature representation of the input sample, and the feature representation of each class is aggregated together to form the key representation of the current class. The value represents the fault classification result generated by the semi-supervised equipment fault diagnosis network, which is used to learn the class distribution of the data. The whole process is shown in formula (14):

[0063]

[0064] Where, k j and v j represent the key representation and value representation of the jth class. and are the gradient information used for memory module update each time. The specific calculation formula is shown in formula (15):

[0065]

[0066] In the starting stage, all key representations and value representations are set to 0 and

[0067] After obtaining the key representation and value representation, the probability prediction based on the memory module is represented as the weighted sum of all value representations

[0068]

[0069] Where, e represents the base of natural logarithm, and dist represents the Euclidean distance.

[0070] Then, the memory enhancement loss is designed based on the probability prediction based on the memory module for training the semi-supervised equipment fault diagnosis network, so that the features extracted by the network are more discriminative. The memory enhancement loss function includes two parts, the class distribution loss and the memory entropy loss. The class distribution loss is used to measure the gap between the prediction results of the semi-supervised equipment fault diagnosis network and the class probability distribution predicted by the memory module. The memory entropy loss is used to ensure that the probability distribution predicted by the memory is as accurate as possible. The memory enhancement loss function L m As shown in formula (17):

[0071]

[0072] The overall loss function L is as follows:

[0073]

[0074] Wherein, L s represents the supervised loss of labeled data, and omega is a hyperparameter of memory loss.

[0075] Verification results:

[0076] The present application adopts multiple groups of experiments to verify the effectiveness of the proposed method.

[0077] Dataset introduction:

[0078] The present application selects the centrifugal water chiller thermal fault experimental data in the public dataset ASHRAE RP-1043 project. ASHRAE RP-1043 is a research project launched by the American Society of Heat, Refrigerating and Air-Conditioning Engineers (ASHRAE) in the 1990s. The project artificially simulates various faults that may occur in the chiller under various conditions, and records the changes of various characteristic parameters. The chiller unit has seven types of faults, and each type is divided into four fault levels according to the severity of the fault. Among them, fault level 1 represents the initial stage of the fault stage, and fault level 4 indicates that the unit has a serious fault.

[0079] The evaluation criteria in the fault diagnosis experiment of the present application are classification accuracy (Accuracy) and F1 score (F1 Score). The training set is expanded to 1500 per class after a certain amount of real data is augmented. The number of real data selected in this experiment includes 10 and 100. In the experiment, 250 samples in each fault type are randomly selected as labeled data, and the others are unlabeled data to form a semi-supervised dataset. In the test set, the number of samples in each class of data is 500.

[0080] In order to verify the performance of the present application, seven commonly used machine learning-based diagnosis methods are selected as comparison methods: RF (Random Forest), SVM (Support Vector Machine), DT (Decision Tree), NB (Multinomial Naive Bayes), MLP (Multilayer Perceptron), KNN (Nearest Neighbor Method) and LR (Logistic Regression Method).

[0081] The method DSCD (Deep Semi-Supervised Contrast Refrigeration Unit Fault Diagnosis Method) proposed by the present application and the comparison methods are experimented under the four fault levels, and the diagnosis results are shown in Tables 1 and 2.

[0082] Table 1 Fault diagnosis results on the semi-supervised dataset constructed by selecting 10 real data

[0083]

[0084]

[0085] Table 2 Fault diagnosis results on semi-supervised dataset constructed by selecting 100 real data

[0086]

[0087] Two points can be observed from the table: (1) the method of the present application can achieve better results under different data conditions, which shows that it has good effect on extracting the class information implied in the unlabeled data; (2) the experimental results of the method proposed in the present application are better than those of the comparative method under all fault levels, which proves the effectiveness of the present application.

Claims

1. A method for refrigeration unit fault diagnosis based on semi-supervised contrastive learning, characterized in that, The steps are as follows: Step 1, data enhancement of refrigeration unit fault data; The original refrigeration unit fault data is enhanced and combined with pair constraints to meet the requirements of contrastive learning, so that it can be applied to sample pair prediction network training based on contrastive learning; The specific steps are as follows: 1.1 Fault data generation The GAN is constructed to generate fault data, and the adversarial learning is used to ensure that the synthetic data is as similar as possible to the original data. An adversarial learning function L (G,D) As shown in formula (1): Wherein, P represents data distribution, G represents generator, and D represents discriminator; x ~ P data represents a sample sampled from real data, x ~ P G represents a sample sampled from generated data; the generator and the discriminator are in confrontation in the training process, and the generative adversarial network converges when the two reach Nash equilibrium; 1.2 Generation data screening To ensure the quality of the generated data, a variational autoencoder is used to further filter the generated data; the variational autoencoder is trained for reconstruction using samples randomly sampled from the generated data, and after training, the variational autoencoder decodes all data for reconstruction; if the reconstruction loss is higher than a set threshold, it is considered that the data quality is too low and does not conform to the real data distribution; the objective function L of training the variational autoencoder is: (E,F) is: where E and F represent the encoder and the decoder, respectively; q F (z | x) represents the posterior distribution approximated by the decoder; p E (z) represents the prior distribution of the latent variable z; KL(q F (z | x) || p E (z)) represents the KL divergence between the q F (z | x) posterior distribution and the p E (z) prior distribution; 1.3 Data set modification In order to meet the sample pair prediction network training based on contrastive learning, the data set needs to be modified to construct a labeled data set with pair constraint characteristics; the pair constraint is guided by the category information, and two samples belonging to the same category are taken as the pair constraint, and the samples not belonging to the same category do not have the pair constraint; given a data set The labeled data set is re-expressed according to the pair constraint relationship Wherein, Where (a, b) is any pair of samples in the data set; Because this binary relationship of belonging to the same class or belonging to different classes is symmetric, the sample pair satisfies the following properties After obtaining the labeled data set α with pair-wise constraint relationship, the unlabeled data set β with pair-wise constraint relationship is represented as β = χ 2 \α; Step 2, construct sample pair prediction network based on contrastive learning; The constructed labeled data set with pair constraint relationship is used for sample pair prediction network training based on contrast learning, a pair of samples is taken as input, and the similarity of the pair of samples is output; during the training process, the representations of the samples are ensured to be close to each other, and the corresponding contrast loss As shown in formula (8): where d(a, b) represents the Euclidean distance between the feature representations of a pair of samples, d(a, b) = ||h(a) - h(b)||, h represents an encoding network that is a concatenation of multiple fully connected layers; in the formula guaranteed to be non-negative by the max operator; Since the goal of the sample pair prediction network based on contrastive learning is to produce a similarity score for a pair of samples, rather than directly predicting whether the two samples belong to the same class, further predict whether they belong to the same class according to the distance value between the two samples; In the actual prediction of unlabeled data pairs, a constant threshold T is designed, and when the distance between the two sample feature representations is less than the threshold, it is considered that the two samples belong to the same class; Because the contrastive loss is only effective for labeled data sets with pair constraints, the sample pair prediction network based on contrastive learning needs to extract features from unlabeled data through reconstruction loss, use the encoding network h to encode the input fault data, and map it to the latent space, and then use the decoding network f to decode the obtained latent features, and realize data reconstruction; By constructing such an encoding-decoding framework, the encoding network h learns the intrinsic features of the complete data while predicting the classification, and the reconstruction loss function L(χ) is as follows: The overall loss L of the sample pair prediction network based on contrastive learning pair The formula is as follows: After training is completed, the sample pair prediction network based on contrastive learning is used to construct pair constraints for all data; Where T represents a constant threshold for controlling whether a pair of samples belong to the same class. Step 3, construct semi-supervised device fault diagnosis network based on memory enhancement; After the sample pair prediction network based on contrastive learning is used to predict the pair constraints of unlabeled data, a fault diagnosis network based on memory enhancement is designed, which is trained using the full data set with pair constraints to realize fault type prediction; The specific steps are as follows: 3.1 Use the full data set with pair constraints to train The sample pair prediction network based on contrast learning is used to make pair constraint prediction on the unlabeled data, so as to obtain a full data set with pair constraint relationship; then, a semi-supervised equipment fault diagnosis network p is constructed, which includes a classification network with multiple full connection layers and a classifier; the classifier uses a softmax function in specific implementation; after the semi-supervised equipment fault diagnosis network makes fault type prediction on a pair of samples in the full data set with pair constraint relationship, the probability p that the pair of samples belongs to the same class is calculated s (a,b) represents where k denotes a certain class, K denotes the number of classes of the data set, p k (a) denotes the predicted probability of sample a with respect to class k, p k (b) denotes the predicted probability of sample b with respect to class k; Similarly, p d (x, y) = 1 - p s (x, y) is the probability that a pair of samples belongs to different classes; to exploit pair-wise constraints to learn class consistency of same-class data, the misclassification loss is designed to minimize the number of misclassified sample pairs; After the semi-supervised device fault diagnosis network is trained using the full data set with pair constraints, it learns the class information hidden in the data and completes the prediction of the sample with respect to the fault type; 3.2 Create a memory module In order to make the features extracted by the semi-supervised equipment fault diagnosis network more discriminative and improve the prediction accuracy, a new memory module is designed to aggregate the discriminative feature representation and the fault classification results generated during the training process of the semi-supervised equipment fault diagnosis network, and then the learned class-related knowledge is transferred to the next semi-supervised equipment fault diagnosis network training to enhance the differences between the features of different classes of unlabeled data samples. The memory module takes the data feature representation and the fault classification results generated during the training process of the semi-supervised equipment fault diagnosis network as input, and constructs a key-value pair representation set. Among them, the key of the memory module represents the feature representation of the input sample, and the feature representation of each class is gathered together to form the key representation of the current class. The value represents the fault classification result generated by the semi-supervised equipment fault diagnosis network, which is used to learn the class distribution of the data. The whole process is shown in formula (14): where k j and v j denote the key representation and value representation of the j-th class;▽k j and▽v j are the gradient information used for memory module update each time; the specific calculation formula is shown in formula (15): At the beginning stage, all key representations and value representations are set to 0 and After obtaining the key representation and the value representation, the probability prediction based on the memory module is represented as a weighted sum of all value representations Where e represents the base of the natural logarithm, and dist represents the Euclidean distance. Then, a memory-enhanced loss is designed based on the probability prediction of the memory module to train the semi-supervised equipment fault diagnosis network, so that the features extracted by the network are more distinctive. The memory-enhanced loss function includes two parts, a category distribution loss and a memory entropy loss. The category distribution loss is used to measure the gap between the prediction results of the semi-supervised equipment fault diagnosis network and the category probability distribution predicted by the memory module. The memory entropy loss is used to ensure that the probability distribution predicted by the memory is as accurate as possible. The memory-enhanced loss function L m As shown in formula (17): The overall loss function L is as follows: where L s denotes the supervised loss of labeled data, and ω is a hyperparameter of the memory loss.

2. The method of claim 1, wherein the method is a semi-supervised contrastive learning based refrigeration unit fault diagnosis method. In the fault data generation, the conditional WGAN is used as a basic model for adversarial learning, label information is introduced as a condition y, and data generation is performed by using the condition and random noise as the input of the generator to further improve the data generation quality; and the corresponding adversarial learning formula L (G,D) is: Given Generating data for the generator Specific generative adversarial formula L (G,D) As shown in formulas (3) and (4): wherein, denotes the generator generating data, denotes a linear interpolation between the generated data and the real data, and ε is a constant between 0 and 1; denotes with respect to the partial derivative of the loss function, and λ denotes a penalty coefficient for adding a gradient penalty constraint in the discriminator.

Images