Water-turbine generator set partial migration fault diagnosis method based on unsupervised domain adversarial network under clustering guidance

Abstract

The invention relates to a partial migration fault diagnosis method for a water-turbine generator set based on an unsupervised domain adversarial network under clustering guidance. An unsupervised domain adversarial network based on a batch standardization method of a specific domain is constructed, specific domain information is removed, and alignment between source domain features and target domain features is enhanced; an embedded clustering strategy is designed, target domain features of close clustering are learned in the partial domain self-adaption process, and the negative migration phenomenon in the partial domain self-adaption process is restrained; a joint optimization objective function is defined to balance different loss functions. The method effectively solves the problem that negative migration is easily caused due to the fact that target domain features extracted by an existing unsupervised partial domain adaptive model in iterative training are relatively dispersed, and can accurately extract the migratable features between the source domain and the target domain, suppress the negative migration phenomenon in the partial domain adaptive process, and improve the adaptive performance of the partial domain. And relatively high diagnosis precision is realized.

Description

Technical Field

[0001] The invention relates to the technical field of hydro-generator fault detection, and in particular to a method for diagnosing partial migration faults of hydro-generator sets based on an unsupervised domain adversarial network under clustering guidance. Background Art

[0002] In recent years, with the rapid advancement of modern industry, the reliability and safety requirements for the integrated machine systems of hydropower stations have been greatly improved. After long-term operation, the key rotating parts of the hydropower generator set, such as bearings and runners, face complex and changeable fault challenges. These faults will not only erode the health of the equipment, but may also cause serious safety accidents, posing a major threat to the operation of the power station. Therefore, the condition monitoring and accurate fault diagnosis of the hydropower generator set have become an indispensable key link in the power station operation and maintenance management system. Relying on advanced real-time online monitoring systems and swing monitoring technologies, it is possible to capture and analyze massive operating data in real time, develop data-driven and model-driven fault diagnosis technologies, provide solid data support and scientific basis for the formulation of condition maintenance strategies, and effectively curb the deterioration of potential faults through early warning and immediate intervention, greatly reducing the risk of major operating accidents, and building a solid line of defense for the safe, stable and efficient operation of hydropower stations.

[0003] Existing research methods mainly focus on fault diagnosis of hydro-turbine generator sets based on feature extraction methods. The accuracy of fault prediction is improved by extracting feature information from monitoring signals that can fully characterize the operating status of the unit, or by constructing intelligent fault diagnosis models based on methods such as convolutional neural networks, which automatically extract high-dimensional features directly from the original monitoring data and reduce the reliance on manual feature design. These networks can learn complex patterns in data and improve the accuracy of fault identification, but there are still some problems that need to be improved: (1) In actual industrial scenarios, since hydro-turbine generator sets are in a healthy state for a long time, the number of samples of fault states is very limited, which makes it difficult to meet the assumption that deep learning requires a large number of labeled identically distributed fault samples; (2) Data-driven intelligent fault diagnosis models need to collect complete data on various fault types of target mechanical equipment, but in some cases, it is impossible to collect data sets for all fault states, making the category space of the target domain only a subset of the category space of the source domain. The domain adaptation effect of the model will be greatly reduced, which is prone to negative transfer problems; Therefore, new technologies need to be introduced to accurately diagnose the fault status of the unit, realize precise monitoring of the unit's operating status and fault warning, and provide strong guarantees for the safe and stable operation of the hydropower station. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a partial migration fault diagnosis method for a hydro-turbine generator set based on an unsupervised domain adversarial network under clustering guidance, so as to realize accurate monitoring of the operating status of the unit and fault warning, and provide strong guarantee for the safe and stable operation of the hydropower station.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for diagnosing partial migration faults of a hydro-generator set based on an unsupervised domain adversarial network under clustering guidance, comprising the following steps: Step S1, collecting vibration data of a hydro-generator set with similar structure and parameters to the actual equipment in the project to construct a source domain data set, and dividing the source domain into training samples and test samples; Step S2, collecting the vibration data of the hydro-generator set collected from the hydro-generator set of a power plant to construct a target domain data set, and dividing the target domain into training samples and test samples; Step S3: linearly transform the source domain and target domain data to the range of [0, 1] using the minimum-maximum normalization method; Step S4, construct an unsupervised domain adversarial network based on domain-specific batch normalization, apply the DSBN layer to remove the interference information of the specific domain in the network model, enhance the alignment of the feature distribution of the source domain and the target domain to extract transferable features; Step S5: adopting an embedded clustering strategy based on KL divergence to cluster each target domain feature into one of the source domain categories during the partial domain adaptation process, and training and optimizing the model; Step S6, defining a joint optimization objective function to balance different losses to improve training and diagnosis performance, including embedded clustering loss, weighted domain classification loss, and weighted domain adaptation adversarial loss; Step S7, input the training samples of the source domain and the target domain, use the Adam algorithm to iteratively train and optimize the proposed network, train to obtain a partial migration fault diagnosis model for the hydro-generator set, and input the target domain test samples into the trained network for verification; Step S8: using the trained partial migration fault diagnosis model of the hydro-generator set to identify the operating status of the hydro-generator set.

[0006] Preferably, step S4 comprises the following steps: Step S41, applying a domain-specific batch normalization method DSBN (Domain Specific Batch Normalization, DSBN) instead of a batch normalization method BN (Batch Normalization, BN) to construct an unsupervised domain adversarial network, including two independent BN branches specifically for source domain and target domain samples. During the training process, each sample selects its corresponding BN branch in DSBN according to its corresponding domain; represents the input sample of the domain adversarial network, where , represents a specific set of parameters for the source domain, Representing the specific parameter set of the target domain, the DSBN layer is represented as; ; in: ; in, and represent In the model training phase, DSBN first divides the source samples and target samples into different small batches, and then calculates the mean and variance of the source domain and the target domain respectively. In the model testing phase, the mean and variance of each domain are used for the samples in the corresponding domain to perform batch normalization operations.

[0007] Preferably, step S5 comprises the following steps: Step S51, a target domain enhancement strategy is designed, using the k-means method to determine the target domain features, and some source domain samples are expanded into the target domain category space to fix the value of the number of clusters; for example, during the training process, for each batch, 60% are randomly selected as target domain samples and 40% are source domain samples (without using labels) to ensure that the category space of the augmented target domain is the same as the category space of the source domain, so that some domain adaptation problems are naturally transferred to pseudo domain adaptation problems, and the number of clusters is equal to the number of source classes. In addition, after the model uses the embedded clustering strategy, it will make the learned source domain sample features and target domain sample features more difficult to distinguish, so that the boundary of the target set obtained by the domain discriminator training is softer, which helps to reduce the impact of negative transfer.

[0008] Step S52: Select the KL divergence between the soft distribution and the auxiliary distribution as the clustering objective function to train and optimize the proposed model. The clustering objective function is expressed as: ; in, is the clustering objective function, and represent soft allocation and auxiliary target distribution respectively; The target domain sample Divided into categories The probability of , that is, soft allocation; Distribute probabilities for predefined high-confidence target domains; Step S53, in calculating When using −Distribution is measured as an embedded kernel function and The closer to the cluster centroid , probability The higher: ; ; in, for − degrees of freedom of the distribution, is the Euclidean distance metric, calculating each target domain feature and similarity; Step S54: In order to standardize the loss contribution of each centroid and prevent large clusters from distorting the hidden feature space, first Raised to the power of 2, and then normalized according to the frequency of each cluster, to calculate : ; in, Indicates the frequency of soft clusters.

[0009] Preferably, the step S6 is specifically as follows: The joint optimization objective function is expressed as: ; ; ; ; ; in, To jointly optimize the objective function, is the embedded clustering loss function, represents the weighted classification loss, represents the weighted domain adaptive adversarial loss, and It is a hyperparameter used to balance different losses in the joint optimization objective function. Indicates the number of source domain clusters, is the indicator function, and represents the feature generator and feature classifier, represents the partial adversarial domain discriminator, The target sample is processed by the Softmax function The predicted label value of yes The dimension weight vector represents the contribution weight of each source domain category; is a source domain sample The category weights, is a constant. As the number of iterations increases, From 0 to 1, It increases from 0 to 1. It is worth noting that in the initial training stage, the clustering loss Fixed to 0. When the number of training cycles reaches the preset value, is activated and set to a non-negative constant.

[0010] The beneficial effects of the partial migration fault diagnosis method of a hydro-turbine generator set based on an unsupervised domain adversarial network under clustering guidance provided by the present invention are as follows: Step S4, construct an unsupervised domain adversarial network based on specific domain batch normalization, apply the DSBN layer to remove the interference information of the specific domain in the network model, and enhance the alignment of the source domain and target domain feature distributions to extract transferable features; Step S5, adopt an embedded clustering strategy based on KL divergence to learn tightly clustered target domain features in the partial domain adaptation process, and suppress the negative migration phenomenon in the partial domain adaptation process; Step S6, define a joint optimization objective function to balance different loss functions to improve the fault diagnosis accuracy of key components of rotating machinery. The present invention solves the problem that the target domain features extracted by the existing unsupervised partial domain adaptation model in iterative training are relatively scattered, which is easy to cause negative migration. The method of the present invention enables the model to learn tightly clustered target domain features in the partial domain adaptation process, suppress the negative migration phenomenon, and achieve higher diagnostic accuracy in partial migration faults of hydro-turbine generator sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a flow chart of the method of the present invention; Figure 2 Schematic diagram of the construction of the unsupervised domain adversarial network of the present invention. DETAILED DESCRIPTION

[0012] A cluster-guided partial migration fault diagnosis method for a hydro-generator set based on an unsupervised domain adversarial network comprises the following steps: Step S1, collecting vibration data of a hydro-generator set with similar structure and parameters to the actual equipment in the project to construct a source domain data set, and dividing the source domain into training samples and test samples; Step S2, collecting the vibration data of the hydro-generator set collected from the hydro-generator set of a power plant to construct a target domain data set, and dividing the target domain into training samples and test samples; Step S3: linearly transform the source domain and target domain data to the range of [0, 1] using the minimum-maximum normalization method; Step S4, construct an unsupervised domain adversarial network based on domain-specific batch normalization, apply the DSBN layer to remove the interference information of the specific domain in the network model, enhance the alignment of the feature distribution of the source domain and the target domain to extract transferable features; Step S5: adopting an embedded clustering strategy based on KL divergence to cluster each target domain feature into one of the source domain categories during the partial domain adaptation process, and training and optimizing the model; Step S6, defining a joint optimization objective function to balance different losses to improve training and diagnosis performance, including embedded clustering loss, weighted domain classification loss, and weighted domain adaptation adversarial loss; Step S7, input the training samples of the source domain and the target domain, use the Adam algorithm to iteratively train and optimize the proposed network, train to obtain a partial migration fault diagnosis model for the hydro-generator set, and input the target domain test samples into the trained network for verification; Step S8: using the trained partial migration fault diagnosis model of the hydro-generator set to identify the operating status of the hydro-generator set.

[0013] Preferably, step S4 comprises the following steps: Step S41, applying a domain-specific batch normalization method DSBN (Domain Specific Batch Normalization, DSBN) instead of a batch normalization method BN (Batch Normalization, BN) to construct an unsupervised domain adversarial network, including two independent BN branches specifically for source domain and target domain samples. During the training process, each sample selects its corresponding BN branch in DSBN according to its corresponding domain; represents the input sample of the domain adversarial network, where , represents a specific set of parameters for the source domain, Representing the specific parameter set of the target domain, the DSBN layer is represented as; ; in: ; in, and represent In the model training phase, DSBN first divides the source samples and target samples into different small batches, and then calculates the mean and variance of the source domain and the target domain respectively. In the model testing phase, the mean and variance of each domain are used for the samples in the corresponding domain to perform batch normalization operations.

[0014] Preferably, step S5 comprises the following steps: Step S51, a target domain enhancement strategy is designed, using the k-means method to determine the target domain features, and some source domain samples are expanded into the target domain category space to fix the value of the number of clusters; for example, during the training process, for each batch, 60% are randomly selected as target domain samples and 40% are source domain samples (without using labels) to ensure that the category space of the augmented target domain is the same as the category space of the source domain, so that some domain adaptation problems are naturally transferred to pseudo domain adaptation problems, and the number of clusters is equal to the number of source classes. In addition, after the model uses the embedded clustering strategy, it will make the learned source domain sample features and target domain sample features more difficult to distinguish, so that the boundary of the target set obtained by the domain discriminator training is softer, which helps to reduce the impact of negative transfer.

[0015] Step S52: Select the KL divergence between the soft distribution and the auxiliary distribution as the clustering objective function to train and optimize the proposed model. The clustering objective function is expressed as: ; in, is the clustering objective function, and represent soft allocation and auxiliary target distribution respectively; The target domain sample Divided into categories The probability of , that is, soft allocation; Distribute probabilities for predefined high-confidence target domains; Step S53, in calculating When using −Distribution is measured as an embedded kernel function and The closer to the cluster centroid , probability The higher: ; ; in, for − degrees of freedom of the distribution, is the Euclidean distance metric, calculating each target domain feature and similarity; Step S54: In order to standardize the loss contribution of each centroid and prevent large clusters from distorting the hidden feature space, first Raised to the power of 2, and then normalized according to the frequency of each cluster, to calculate : ; in, Indicates the frequency of soft clusters.

[0016] Preferably, the step S6 is specifically as follows: The joint optimization objective function is expressed as: ; ; ; ; ; in, To jointly optimize the objective function, is the embedded clustering loss function, represents the weighted classification loss, represents the weighted domain adaptive adversarial loss, and It is a hyperparameter used to balance different losses in the joint optimization objective function. Indicates the number of source domain clusters, is the indicator function, and represents the feature generator and feature classifier, represents the partial adversarial domain discriminator, The target sample is processed by the Softmax function The predicted label value of yes The dimension weight vector represents the contribution weight of each source domain category; is a source domain sample The category weights, is a constant. As the number of iterations increases, From 0 to 1, It increases from 0 to 1. It is worth noting that in the initial training stage, the clustering loss Fixed to 0. When the number of training cycles reaches the preset value, is activated and set to a non-negative constant.

[0017] In this example, the source domain data set is constructed using the vibration data of a hydro-turbine generator set with similar structure and parameters to the actual equipment in the project, and the target domain data set is constructed using the vibration data of a hydro-turbine generator set collected from a hydro-turbine generator set in a power plant. The data collected includes vibration acceleration signals of normal state, bearing inner ring fault, bearing outer ring fault, bearing rolling element fault, bearing cage fault, and insulation fault, as shown in Table 1. 200 samples are extracted for each fault mode, of which 160 (80%) are used as training samples and the rest are used as test samples.

[0018] Table 1 Detailed information of the diagnostic dataset Fault Tags Failure Mode 1 Normal state 2 Bearing inner ring failure 3 Bearing outer ring failure 4 Bearing rolling element failure 5 Bearing cage failure 6 Insulation failure The network structure settings are shown in Table 2, and the training parameters are as follows: The value range of is [64,256], and the value range of the initial learning rate is {10 -4 , 10 -3 , 10 -2 , 10 -1}, the number of iterations is 300, the preset number of iterations is 20, and the Adam optimization algorithm is selected for training. In addition, it is found through experience that the preset number of iterations is set to 20 to provide a good initialization point for the embedded clustering loss.

[0019] Table 2 Domain adversarial network structure and parameters

[0020] In order to verify the effectiveness of the proposed method, four different groups of partial transfer learning tasks were set up, and the detailed setting information is shown in Table 3. In order to reduce random errors, each group of partial transfer learning tasks was repeated 10 times, and a variety of popular transfer learning methods were used for evaluation and comparison, including CNN, MMD domain adaptation method based on Gaussian kernel induction, partial adversarial domain adaptation method, and importance-weighted adversarial network.

[0021] Table 3 Experimental setting information of some transfer learning tasks Task Name Source domain category space Target domain category space A1 1,2,3,4,5,6 1,2,3,4,5,6 A2 1,2,3,4,5,6 1,2,3,4,5 A3 1,2,3,4,5,6 1,2,3,4 A4 1,2,3,4,5,6 1,2,3 Table 4 shows the diagnostic accuracy of different methods in partial transfer learning tasks, where L represents the diagnostic accuracy of the method in the last iteration. In addition, in order to quantitatively evaluate the impact of the negative transfer effect, Table 4 also lists the highest diagnostic accuracy of the method in the entire training process (denoted as M). It can be seen from the table that: compared with the other four popular transfer learning methods, the proposed method has achieved better diagnostic accuracy in both classic transfer learning tasks (A1) and partial transfer learning tasks (A2, A3, A4), indicating that the embedded clustering strategy can suppress the negative transfer phenomenon in the local domain adaptation process, and the model's specific domain batch normalization method can strengthen the alignment between the two domains to a certain extent; Table 4 Diagnostic accuracy of different methods in some transfer learning tasks

[0022] like Figure 1As shown, the content of the present invention can be mainly divided into four parts. The first part is to collect the vibration data of the hydro-turbine generator set with similar structure and parameters to the actual equipment of the project to construct the source domain data set, collect the vibration data of the hydro-turbine generator set from the measured hydro-turbine generator set of a certain power plant to construct the target domain data set and divide the target domain training samples and test samples, and divide them into the training samples and test samples of the source domain and the target domain respectively after removing the mean and normalization processing; the second part is to construct an unsupervised domain adversarial network based on specific domain batch standardization, apply the DSBN layer to remove the interference information of the specific domain in the network model, and enhance the alignment of the source domain and target domain feature distribution to extract transferable features; the third part is to adopt an embedded clustering strategy based on KL divergence, and promote each target domain feature to be clustered into one of the source domain categories in the partial domain adaptation process, and train and optimize the model; the fourth part is to define a joint optimization objective function to balance different losses to improve training and diagnostic performance; the fifth part is to use the Adam algorithm to iteratively train and optimize the proposed network, train to obtain a partial migration fault diagnosis model of the hydro-turbine generator set, and input the target domain test sample into the trained network for verification; like Figure 2 As shown in Figure 2, the basic structural modules of the proposed network model include feature generator, domain discriminator and classifier. DSBN contains two independent BN branches specifically for source domain and target domain samples. During the training process, each sample selects its corresponding BN branch in DSBN according to its corresponding domain.

[0023] The present invention is a method for diagnosing partial migration faults of hydro-turbine generator sets based on an unsupervised domain adversarial network under clustering guidance. An unsupervised domain adversarial network based on a batch normalization method for a specific domain is constructed to remove specific domain information and enhance the alignment between source domain and target domain features; an embedded clustering strategy is designed to learn tightly clustered target domain features in the partial domain adaptation process to suppress the negative migration phenomenon in the partial domain adaptation process; a joint optimization objective function is defined to balance different loss functions. This method effectively solves the problem that the target domain features extracted by the existing unsupervised partial domain adaptation model in iterative training are relatively scattered, which easily leads to negative migration. The proposed method can accurately extract transferable features between the source domain and the target domain, suppress the negative migration phenomenon in the partial domain adaptation process, and achieve a higher diagnostic accuracy, demonstrating its effectiveness and superiority in solving the problem of unsupervised partial migration fault diagnosis of hydro-turbine generator sets, which is better than the existing technology.

[0024] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A cluster-guided partial migration fault diagnosis method for a hydro-generator unit based on an unsupervised domain adversarial network, characterized in that: The following steps are involved: Step S1, collecting vibration data of a hydro-turbine generator set with similar structure and parameters to the actual equipment in the project to construct a source domain data set, and dividing the source domain into training samples and test samples; Step S2, collecting the vibration data of the hydro-generator set collected from the hydro-generator set of a power plant to construct a target domain data set, and dividing the target domain into training samples and test samples; Step S3: linearly transform the source domain and target domain data to the range of [0, 1] using the minimum-maximum normalization method; Step S4, construct an unsupervised domain adversarial network based on domain-specific batch normalization, apply the DSBN layer to remove the interference information of the specific domain in the network model, enhance the alignment of the feature distribution of the source domain and the target domain to extract transferable features; Step S5: adopting an embedded clustering strategy based on KL divergence to cluster each target domain feature into one of the source domain categories during the partial domain adaptation process, and training and optimizing the model; Step S6, defining a joint optimization objective function to balance different losses to improve training and diagnosis performance, including embedded clustering loss, weighted domain classification loss, and weighted domain adaptation adversarial loss; Step S7, input the training samples of the source domain and the target domain, use the Adam algorithm to iteratively train and optimize the proposed network, train to obtain a partial migration fault diagnosis model for the hydro-generator set, and input the target domain test samples into the trained network for verification; Step S8: using the trained partial migration fault diagnosis model of the hydro-generator set to identify the operating status of the hydro-generator set.

2. According to claim 1, a cluster-guided hydro-generator partial migration fault diagnosis method based on an unsupervised domain adversarial network is characterized in that: The step S4 comprises the following steps: Step S41, applying a domain-specific batch normalization method DSBN to replace the batch normalization method BN to construct an unsupervised domain adversarial network, including two independent BN branches specifically for source domain and target domain samples. During the training process, each sample selects its corresponding BN branch in the DSBN according to its corresponding domain; represents the input sample of the domain adversarial network, where , represents a specific set of parameters for the source domain, Representing the specific parameter set of the target domain, the DSBN layer is represented as; ； in: ； in, and represent In the model training phase, DSBN first divides the source samples and target samples into different small batches, and then calculates the mean and variance of the source domain and the target domain respectively. In the model testing phase, the mean and variance of each domain are used for the samples in the corresponding domain to perform batch normalization operations.

3. According to claim 1, a cluster-guided hydro-generator partial migration fault diagnosis method based on an unsupervised domain adversarial network is characterized in that: The step S5 comprises the following steps: Step S51: Use the k-means method to determine the target domain features, and expand some source domain samples into the target domain category space to fix the value of the number of clusters; Step S52: Select the KL divergence between the soft distribution and the auxiliary distribution as the clustering objective function to train and optimize the proposed model. The clustering objective function is expressed as: ； in, is the clustering objective function, and represent soft allocation and auxiliary target distribution respectively; The target domain sample Divided into categories The probability of , that is, soft allocation; Distribute probabilities for predefined high-confidence target domains; Step S53, in calculating When using −Distribution is measured as an embedded kernel function and The closer to the cluster centroid , probability The higher: ； ； in, for − degrees of freedom of the distribution, is the Euclidean distance metric, calculating each target domain feature and similarity; Step S54: In order to standardize the loss contribution of each centroid and prevent large clusters from distorting the hidden feature space, first Raised to the power of 2, and then normalized according to the frequency of each cluster, to calculate : ； in, Indicates the frequency of soft clusters.

4. According to claim 1, a cluster-guided hydro-generator partial migration fault diagnosis method based on an unsupervised domain adversarial network is characterized in that: The step S6 is specifically as follows: The joint optimization objective function is expressed as: ； ； ； ； ； in, To jointly optimize the objective function, is the embedded clustering loss function, represents the weighted classification loss, represents the weighted domain adaptive adversarial loss, and It is a hyperparameter used to balance different losses in the joint optimization objective function. Indicates the number of source domain clusters, is the indicator function, and represents the feature generator and feature classifier, represents the partial adversarial domain discriminator, The target sample is processed by the Softmax function The predicted label value of yes The dimension weight vector represents the contribution weight of each source domain category; is a source domain sample The category weights, is a constant. As the number of iterations increases, From 0 to 1, Increases from 0 to 1.

Images