Fault diagnosis method for unbalanced data set of hydroelectric generating set generated by adaptive domain transformation
Through the adaptive domain transformation generation method, the data imbalance in the fault diagnosis of hydropower units is solved, the diagnostic reliability and accuracy of the model are improved, and higher quality data generation and fault identification are achieved.
Patent Information
- Application Number
- CN202410978237.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-05
AI Technical Summary
The deep learning model is limited by data imbalance in the fault diagnosis of hydropower units, which leads to low reliability in practical applications and cannot accurately predict the fault status of unknown monitoring signals of mechanical equipment.
Adaptive domain transformation generation method is adopted to build a cross-domain adversarial generation network to realize bidirectional sample conversion between the source domain and the target domain, establish forward generator, reverse generator and differentiator, and use Wasserstein-GP loss and minimum Two-multiple loss function optimization generator and differentiator to improve data quality and model diagnostic reliability.
It effectively solves the problem of data imbalance, improves the accuracy and generalization capabilities of the fault diagnosis model, and has higher data quality. The model provides more reliable fault diagnosis results in practical applications.
Smart Images

Figure CN118897987A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of hydropower unit fault diagnosis, and in particular to a method for diagnosing a hydropower unit unbalanced data set fault generated by adaptive domain transformation. Background Art
[0002] In the field of mechanical equipment fault diagnosis, with the rapid development of artificial intelligence, big data and computer technology, deep learning technology has become a model of the intelligent era, bringing revolutionary changes to fault diagnosis. However, the application of deep learning technology in this field also faces some challenges, especially in real industrial production environments.
[0003] Traditional fault diagnosis methods, such as theoretical analysis, physical simulation, and dynamic analysis, have revealed their limitations in the information and digital age. These methods are often based on manual experience and fixed patterns and cannot effectively cope with the complex and changeable working environment and fault types of modern mechanical equipment.
[0004] Deep learning technology, by simulating the operating mechanism of human brain neural network, can automatically learn and extract features from data, thereby realizing intelligent fault diagnosis of mechanical equipment. For example, various deep fault diagnosis network models, such as MSCNN-FA, AMDC-CNN, ADCAE, LEFE-Net, MRSFN and DTS-CNN, have demonstrated good feature extraction and fault diagnosis performance. As for data-driven deep learning, the performance of deep learning models is highly dependent on the quality and quantity of data. Sufficient and diverse fault category data is the key to training high-performance models.
[0005] However, in real industrial production, the abnormal and faulty conditions of large mechanical equipment such as hydropower units only account for a very small part of the overall equipment operation time, and only the monitoring data under normal conditions is easy to obtain and accurately marked. When there are slight abnormalities in mechanical equipment, the monitoring data is often ignored and not correctly collected and identified. This data imbalance will cause the deep learning model to tend to identify normal conditions, but the ability to identify fault conditions is insufficient. Therefore, in the field of diagnosis, the lack of fault samples under real operating conditions has made it difficult to apply deep learning methods. Because there are not enough real samples to train the deep diagnostic model, the model cannot accurately predict the fault state of the unknown monitoring signal of the machinery, and the model reliability is low, resulting in great obstacles to the actual application of the model.
[0006] Due to the lack of sufficient real fault samples, deep learning models find it difficult to fully learn various fault modes during training, resulting in low reliability of the model in practical applications. In the absence of sufficient real samples, the model may not be able to accurately predict the fault status of unknown monitoring signals of mechanical equipment, thus limiting the widespread application of deep learning in the field of fault diagnosis.
[0007] Therefore, solving the problem of data imbalance has become a major difficulty in deep learning. Summary of the invention
[0008] The present invention aims to address the technical defects of the prior art and provide a method for fault diagnosis of hydropower unit unbalanced data sets generated by adaptive domain transformation. By performing bidirectional sample conversion between the source domain and the target domain and aligning samples from different subdomains, the deep model training problem caused by data imbalance in the target domain is solved, data quality is improved, and the reliability of model diagnosis is improved.
[0009] The present invention provides the following technical solution: a method for diagnosing a fault of a hydropower unit unbalanced data set generated by adaptive domain transformation, the method comprising the following steps:
[0010] Step 1: Construct basic sample sets of source domain and target domain;
[0011] The source domain includes several different types of fault samples and well-labeled data sets, and the target domain is a data set with an unbalanced number of samples;
[0012] Step 2: Data preprocessing;
[0013] Perform data normalization on the basic sample set in step 1 to obtain data that is distributed in [-1,1] as a whole, and then use a deep convolutional network to extract signal features and construct a two-dimensional data input model;
[0014] Step 3: Establish a domain transformation sample generation model based on a cross-domain adversarial generative network;
[0015] The domain transformation sample generation model includes a forward generator G, a target domain distinguisher Dt, a reverse generator F, and a source domain distinguisher Ds, which can realize the forward conversion of source domain samples into target domain samples and the reverse conversion of target domain samples into source domain samples;
[0016] Step 4: Construct the loss functions of the generator and the discriminator, and establish the loss functions of the composite Wasserstein-GP loss and the least squares loss;
[0017] Used to evaluate the difference between generated samples and real samples;
[0018] Step 5: Train the model.
[0019] A gradient optimization algorithm is used to optimize the bidirectional paired generator and discriminator, and the training model realizes cross-domain bidirectional sub-class sample conversion and cross-domain sub-domain corresponding alignment generation.
[0020] Preferably, the step 1 comprises the following steps:
[0021] Step 1-1: Get the original data, build a data set, divide the vibration signal into several sample segments, the data length is 1024, each sample signal must contain several vibration cycles to fully express the fault information;
[0022] Step 1-2: The source domain and the target domain have the same number of sample categories, wherein the source domain has a large number of samples in each category, and the target domain has a large number of normal category samples.
[0023] Preferably, the step 2 comprises the following steps:
[0024] Step 2-1: Map the monitoring data in step 1 to the specified range. The data normalization process adopts the Min-Max normalization method to obtain data with an overall distribution in [-1,1]. The formula is as follows:
[0025]
[0026] Among them, x max is the maximum value in the data set, x min is the minimum value in the data set, x new is the normalized data, and x is the original data;
[0027] Step 2-2: Convert each sample segment of length 1024 into 32×32 graph structure data.
[0028] Preferably, the step 3 comprises the following steps:
[0029] A bidirectional generative network deep model and a domain sample distinguishing deep model are established. First, a bidirectional generative network deep model is constructed, which is built with different structural layers, including a downsampling layer, a residual structure layer, and an upsampling layer; secondly, a domain sample distinguishing deep model is constructed, and the domain sample distinguishing deep model includes a sample generator model and a distinguisher model.
[0030] Preferably, the downsampling layer is composed of a 2D convolution structure, the input is the constructed two-dimensional data matrix x, and the output is x', and the formula is as follows:
[0031]
[0032] Where conv represents convolution operation, bn represents regularization, relu represents nonlinear processing using relu activation function, w, b represent convolution kernel and bias, and the main parameters of convolution are: kernel_size = (3,3), strides = (2,2), paading = 'same' during convolution;
[0033] The residual structure layer consists of the main convolution channel and the residual connection, and its formula is as follows:
[0034]
[0036] o=x+x"
[0037] Where x is the module input feature, o is the residual feature output, and the main convolution parameters are: kernel_size = (3, 3), strides = (1, 1);
[0038] The upsampling layer performs feature upsampling on the input feature map, and its formula is as follows:
[0039]
[0040] Where x is the module input feature, x' is the residual feature output, and the main convolution parameters are: kernel_size = (3, 3), strides = (2, 2).
[0041] Preferably, the sample generator model is composed of a basic convolution, a regularization layer, and independent modules in an input-output order. The input of the sample generator model is the data of the sample in the source domain processed by step 2, and its dimension is (32, 32). The output is the sample in the target domain, and its dimension is (32, 32). The structure of the sample generator model is as follows:
[0042]
[0043] Where LeakyReLU(x)=max(0,x)+a*min(0,x), a is used to control the slope of the negative part and change the output range;
[0044] The discriminator model is matched with the sample generator model, and the sample input in the data set is subjected to feature extraction to achieve class differentiation. The structure of the discriminator model is as follows:
[0045]
[0046]
[0047] The parameters of each branch of the discriminator model are set as
[0048] (1)1×1conv
[0049] (2) 3×3conv, 5×5conv
[0050] (3)7×7conv,11×11conv
[0051] (4) 3×3maxpool, 1×1conv
[0052] Among them, Conv means convolution and maxpool means maximum pooling.
[0053] Preferably, the generator (G, F) and the distinguisher (Dt, Ds) of the bidirectional generation have the same structure.
[0054] Preferably, step 4 comprises the following steps:
[0055] Step 4-1: Establish Wasserstein-GP loss, for the source domain sample x s , target domain sample x t , For the generated samples, the loss function with gradient penalty is as follows:
[0056]
[0057] where ∈~U[0,1],L D is the discriminator loss, L G is the generator loss;
[0058] Step 4-2: Establish the least squares loss based on the last dimension element y of the output vector of the discriminator D e , the formula is:
[0059] L GLS =(D(G(x s ))-1) 2
[0060] L DLS =(D(x t )-1) 2 ;
[0061] Step 4-3: Fusion of two sets of loss functions, the total loss target of the generator and the discriminator is:
[0062] L D =L DGP +L DLS
[0063] L G =L GGP +L GLS .
[0064] Preferably, the step 5 comprises the following steps:
[0065] Step 5-1: Model training starts with the input sample. The source domain sample is forward transformed by the generator to obtain the generated target domain sample.
[0066] Step 5-2: According to the real samples and generated samples of the target domain, calculate the loss function L of the generator and the discriminator according to step 4G , L D ;
[0067] Step 5-3: Use the deep learning gradient optimization algorithm Adam, set the learning rate to 0.0002, train the iterative generator and discriminator, and select 300-500 iterations;
[0068] Step 5-4: Swap the source domain and the target domain, change the data generation direction, repeat steps 5-1 to 5-3, train another pair of generators and discriminators, convert the generated samples in each data domain into samples in another domain, train the overall model again, cyclically convert between domain samples, supplement the minority samples in the unbalanced dataset, build a new balanced dataset, and train the model using the balanced sample set generated across domains.
[0069] The method for fault diagnosis of unbalanced data set of hydropower units generated by adaptive domain transformation in the present invention is a deep learning technology, which aims to solve the problem of data imbalance and improve the accuracy of fault diagnosis, and is achieved by the following steps:
[0070] 1. Construct basic sample sets of source and target domains:
[0071] Source domain: contains samples of various fault types and labeled datasets of normal states. These data are usually rich and can serve as the basis for training models.
[0072] Target domain: The number of samples in the dataset is unbalanced, that is, the number of normal state samples is much larger than the number of fault state samples. This situation is very common in practical applications because the probability of fault occurrence is low.
[0073] 2. Data preprocessing:
[0074] Normalize the basic sample sets of the source domain and the target domain to ensure that the data is distributed in the range of [-1,1] to facilitate subsequent deep learning model processing;
[0075] Use deep convolutional networks (such as CNN) to extract features of the signal and construct these features into a two-dimensional data input pattern in preparation for model training;
[0076] 3. Establish a domain transformation sample generation model based on a cross-domain adversarial generative network:
[0077] The domain transformation sample generation model consists of a forward generator G, a target domain discriminator Dt, a reverse generator F, and a source domain discriminator Ds;
[0078] The forward generator G is responsible for converting source domain samples into target domain samples, and the reverse generator F converts target domain samples back to source domain samples;
[0079] The target domain discriminator Dt and the source domain discriminator Ds are used to distinguish samples from the target domain and the source domain, respectively, to ensure that the distribution of the generated samples is consistent with that of the real samples;
[0080] 4. Construct the loss function of the generator and the discriminator:
[0081] Use the loss function of composite Wasserstein-GP loss and least square loss to evaluate the difference between generated samples and real samples;
[0082] Wasserstein-GP loss is a loss function based on Wasserstein distance, which can measure the difference between two probability distributions, while the GP (gradient penalty) term is used to enhance the stability of the model; the least squares loss is a commonly used regression loss function, which is used to minimize the squared difference between the predicted value and the true value;
[0083] 5. Training Model:
[0084] The parameters of the generator and discriminator are optimized using a gradient optimization algorithm. During the training process, the model will learn how to perform bidirectional sample conversion between the source domain and the target domain, and how to align samples from different subdomains to improve the model's ability to diagnose imbalanced data sets.
[0085] The present invention adopts a two-dimensional data structure, and uses two-dimensional convolution and transposed convolution to construct a generative model. Compared with the method using one-dimensional signal and one-dimensional convolution, the model training process has stronger stability, faster convergence speed, more obvious effect, higher quality of generated data, and the model can generate data with various characteristics, which has a wide range of applications.
[0086] In this scheme, a multi-scale sample discriminator is constructed. The model extracts the features of input data from multiple single scales and composite scales, which greatly improves the model's sample quality supervision ability. The powerful sample discrimination effect of the discriminator also promotes the improvement of the generator's generation quality. The proposed W-GP and LS joint loss effectively measures the model training process and can be used as a judgment indicator of the training effect. Compared with the training method using a single loss, it has obvious advantages and can obtain a higher generation effect. At the same time, the discriminator and loss have wide application capabilities and can improve the actual performance of various model methods.
[0087] Through this method, the data imbalance problem can be effectively solved and the generalization ability and accuracy of the fault diagnosis model in the target domain can be improved. At the same time, through domain transformation technology, the model can better learn and adapt to the feature distribution of the target domain, thereby providing more reliable fault diagnosis results in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0088] Figure 1A flow chart of a method for diagnosing a hydropower unit unbalanced data set generated by adaptive domain transformation according to the present invention;
[0089] Figure 2 The waveform analysis result 1 for generating different fault data, wherein (a) is the real data waveform, and (b) is the data waveform generated by the method of the present invention;
[0090] Figure 3 The waveform analysis result 2 for generating different fault data, wherein (a) is the real data waveform, and (b) is the data waveform generated by the method of the present invention;
[0091] Figure 4 The third waveform analysis result for generating different fault data, wherein (a) is the real data waveform, and (b) is the data waveform generated by the method of the present invention;
[0092] Figure 5 The waveform analysis result 4 for generating different fault data, wherein (a) is the real data waveform, and (b) is the data waveform generated by the method of the present invention;
[0093] Figure 6 It is the visualization result of the high-dimensional original data analyzed by T-SNE dimensionality reduction technology, which is used to illustrate the consistency and similarity between the generated data and the real data;
[0094] Figure 7 It is a schematic diagram of sample domain transformation;
[0095] Figure 8 is a diagram of the unbalanced data fault diagnosis method proposed by the present invention;
[0096] Fig. 9 It is a structural diagram of the domain transformation generator proposed by the present invention;
[0097] Fig.10 It is a distribution diagram of the data generated by each comparison method and the real data. DETAILED DESCRIPTION
[0098] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work based on the embodiments of the present invention shall fall within the scope of protection of the present invention.
[0099] like Figures 1 to 10 As shown, it shows a specific embodiment of the present invention: Figures 1 to 10 As shown, the present invention discloses a method for diagnosing a hydropower unit unbalanced data set generated by adaptive domain transformation, the method comprising the following steps:
[0100] Step 1: Construct basic sample sets of source domain and target domain;
[0101] The source domain includes several different types of fault samples and well-labeled data sets, and the target domain is a data set with an unbalanced number of samples;
[0102] Step 2: Data preprocessing;
[0103] Perform data normalization on the basic sample set in step 1 to obtain data that is distributed in [-1,1] as a whole, and then use a deep convolutional network to extract signal features and construct a two-dimensional data input model;
[0104] Step 3: Establish a domain transformation sample generation model based on a cross-domain adversarial generative network;
[0105] The domain transformation sample generation model includes a forward generator G, a target domain distinguisher Dt, a reverse generator F, and a source domain distinguisher Ds, which can realize the forward conversion of source domain samples into target domain samples and the reverse conversion of target domain samples into source domain samples;
[0106] Step 4: Construct the loss functions of the generator and the discriminator, and establish the loss functions of the composite Wasserstein-GP loss and the least squares loss;
[0107] Used to evaluate the difference between generated samples and real samples;
[0108] Step 5: Train the model.
[0109] A gradient optimization algorithm is used to optimize the bidirectional paired generator and discriminator, and the training model realizes cross-domain bidirectional sub-class sample conversion and cross-domain sub-domain corresponding alignment generation.
[0110] Preferably, the step 1 comprises the following steps:
[0111] Step 1-1: Get the original data, build a data set, divide the vibration signal into several sample segments, the data length is 1024, each sample signal must contain several vibration cycles to fully express the fault information;
[0112] Step 1-2: The source domain and the target domain have the same number of sample categories, wherein the source domain has a large number of samples in each category, and the target domain has a large number of normal category samples.
[0113] Step 1 describes how to construct basic sample sets in the source domain and the target domain, especially for fault diagnosis of the unbalanced data set of hydropower units, and obtain the original data: obtain the original vibration signal data from the vibration monitoring system of the hydropower unit, which usually contains various states of the equipment during operation, including normal state and various fault states; preprocess the original data to remove unnecessary information such as noise and outliers to ensure the accuracy and reliability of the data; divide the vibration signal into several sample segments in chronological order, and set the length of each sample segment to 1024 data points to fully capture the characteristic information in the vibration signal; ensure that each sample signal contains several complete vibration cycles, This is because the vibration period is one of the most important features in the vibration signal, which reflects the operating status and fault information of the equipment. The length of the vibration period can be determined by performing spectrum analysis or time domain analysis on the vibration signal, thereby ensuring that each sample segment contains enough vibration periods. When selecting sample segments, special attention should be paid to ensuring that the fault information can be fully expressed. The fault information is usually contained in a specific frequency band or specific feature of the vibration signal, so these factors need to be considered when dividing the sample segments. The feature extraction and analysis of the vibration signal under the fault state can be performed to determine which frequency bands or features contain the most important fault information, and focus on these frequency bands or features when dividing the sample segments;
[0114] The number of sample categories in the source domain and the target domain is the same: When constructing the source domain and the target domain, it is necessary to ensure that they contain the same number of fault categories. This is because deep learning models usually need to learn the differences and connections between different fault categories during the training process in order to accurately diagnose faults. In the source domain, the number of samples for each fault category should be relatively large. This is because the source domain is a dataset used to train deep learning models. A sufficient number of samples can ensure that the model can learn enough feature information to improve the recognition and generalization capabilities of the model. At the same time, due to the large number of samples in the source domain, more complex models and more advanced training strategies can be used during the training process to further improve the performance of the model. In the target domain, the number of normal class samples should be relatively large. This is because in actual applications, hydropower units are in normal operation most of the time, so the number of normal class samples is usually much larger than the number of fault class samples.
[0115] In order to solve the data imbalance problem, the number of normal class samples in the target domain can be increased to simulate the real application scenario. This can be achieved by collecting more normal state data from actual operation, using data enhancement technology to generate new normal class samples, etc.
[0116] Preferably, the step 2 comprises the following steps:
[0117] Step 2-1: Map the monitoring data in step 1 to the specified range. The data normalization process adopts the Min-Max normalization method to obtain data with an overall distribution in [-1,1]. The formula is as follows:
[0118]
[0119] Among them, x max is the maximum value in the data set, x min is the minimum value in the data set, x new is the normalized data, and x is the original data;
[0120] Step 2-2: Convert each sample segment of length 1024 into 32×32 graph structure data.
[0121] For the vibration signal data of the hydropower unit, the Min-Max normalization method can be used for data normalization processing, and the original data can be linearly transformed into a specified range. The data is required to be distributed in the range of [-1,1]. For sample segments with a length of 1024, they are converted into 32x32 graph structure data. Specifically, the one-dimensional array can be arranged into a 32x32 matrix in rows or columns, so that the convolutional neural network can be used.
[0122] Preferably, the step 3 comprises the following steps:
[0123] A bidirectional generative network deep model and a domain sample distinguishing deep model are established. First, a bidirectional generative network deep model is constructed, which is built with different structural layers, including a downsampling layer, a residual structure layer, and an upsampling layer; secondly, a domain sample distinguishing deep model is constructed, and the domain sample distinguishing deep model includes a sample generator model and a distinguisher model.
[0124] By constructing such a bidirectional generative network deep model and a domain sample distinguishing deep model, and combining them with appropriate loss functions and optimization algorithms for training, sample conversion and alignment between the source domain and the target domain can be achieved, thereby solving the problems of data imbalance and cross-domain learning.
[0125] Preferably, the downsampling layer is composed of a 2D convolution structure, the input is the constructed two-dimensional data matrix x, and the output is x', and the formula is as follows:
[0126]
[0127] Where conv represents convolution operation, bn represents regularization, relu represents nonlinear processing using relu activation function, w, b represent convolution kernel and bias, and the main parameters of convolution are: kernel_size = (3,3), strides = (2,2), padding = 'same' during convolution;
[0128] The residual structure layer consists of the main convolution channel and the residual connection, and its formula is as follows:
[0129]
[0131] o=x+x"
[0132] Where x is the module input feature, o is the residual feature output, and the main convolution parameters are: kernel_size = (3, 3), strides = (1, 1);
[0133] The upsampling layer performs feature upsampling on the input feature map, and its formula is as follows:
[0134]
[0135] Where x is the module input feature, x' is the residual feature output, and the main convolution parameters are: kernel_size = (3, 3), strides = (2, 2).
[0136] The downsampling layer is mainly used to reduce the spatial dimension of the data and increase the number of feature channels in order to extract features at a more abstract level. This is usually achieved through convolutional layers and pooling layers. The convolutional layer is used to learn local features, while the pooling layer is used to reduce the spatial dimension, speed up the calculation, and enhance the robustness of the model.
[0137] The residual structure layer is a commonly used structure in deep neural networks. It helps solve the problems of gradient disappearance and model degradation during network training. It adds the input directly to the output, making it easier for the network to learn and thus easier to optimize.
[0138] The upsampling layer is the opposite of the downsampling layer, which is used to increase the spatial dimension of the data. The upsampling layer is used to restore the original size (or larger) image from a smaller feature map.
[0139] Preferably, the sample generator model is composed of a basic convolution, a regularization layer, and independent modules in an input-output order. The input of the sample generator model is the data of the sample in the source domain processed by step 2, and its dimension is (32, 32). The output is the sample in the target domain, and its dimension is (32, 32). The structure of the sample generator model is as follows:
[0140]
[0141] Where LeakyReLU(x)=max(0,x)+a*min(0,x), a is used to control the slope of the negative part and change the output range;
[0142] The discriminator model is matched with the sample generator model, and the sample input in the data set is subjected to feature extraction to achieve class differentiation. The structure of the discriminator model is as follows:
[0143]
[0144] The parameters of each branch of the discriminator model are set as
[0145] (1)1×1conv
[0146] (2) 3×3conv, 5×5conv
[0147] (3)7×7conv,11×11conv
[0148] (4) 3×3maxpool, 1×1conv
[0149] Among them, Conv means convolution and maxpool means maximum pooling.
[0150] The goal of the sample generator model is to transform samples in the source domain into samples in the target domain while maintaining the same spatial dimension. The goal of the discriminator model is to extract features from the input samples and achieve class differentiation, that is, to determine whether the input samples are real samples or fake samples generated by the generator. The model structure uses the Inception module and channel attention mechanism. The Inception module contains multiple parallel branches, each branch uses convolution kernels and pooling operations of different sizes, and the outputs of these branches are spliced in the channel dimension and then passed to the next layer.
[0151] Preferably, the generator (G, F) and the distinguisher (Dt, Ds) of the bidirectional generation have the same structure.
[0152] To train a bidirectional generative adversarial network, four networks (G, F, Dt, Ds) need to be trained simultaneously, and a cycle consistency loss is used to ensure that the generated samples remain consistent across domains.
[0153] Preferably, step 4 comprises the following steps:
[0154] Step 4-1: Establish Wasserstein-GP loss, for the source domain sample x s , target domain sample x t , For the generated samples, the loss function with gradient penalty is as follows:
[0155]
[0156] where ∈~U[0,1],L D is the discriminator loss, LG is the generator loss;
[0157] Step 4-2: Establish the least squares loss based on the last dimension element y of the output vector of the discriminator D e , the formula is:
[0158] L GLS =(D(G(x s ))-1) 2
[0159] L DLS =(D(x t )-1) 2 ;
[0160] Step 4-3: Fusion of two sets of loss functions, the total loss target of the generator and the discriminator is:
[0161] L D =L DGP +L DLS
[0162] L G =L GGP +L GLS .
[0163] Discriminator loss: the ability to distinguish between real samples and generated samples;
[0164] Generator loss: encourages the generator to generate samples that can deceive the discriminator;
[0165] Gradient penalty term: ensures that the gradient of the discriminator in the sample space meets certain conditions (for example, the gradient norm is close to 1), which helps to maintain the smoothness of the generated sample space; the least squares loss is used to replace the traditional cross entropy loss, which can provide more stable training. Finally, the total loss of the generator and the discriminator is some combination of these two sets of loss functions.
[0166] Preferably, the step 5 comprises the following steps:
[0167] Step 5-1: Model training starts with the input sample. The source domain sample is forward transformed by the generator to obtain the generated target domain sample.
[0168] Step 5-2: According to the real samples and generated samples of the target domain, calculate the loss function L of the generator and the discriminator according to step 4 G , L D ;
[0169] Step 5-3: Use the deep learning gradient optimization algorithm Adam, set the learning rate to 0.0002, train the iterative generator and discriminator, and select 300-500 iterations;
[0170] Step 5-4: Swap the source domain and the target domain, change the data generation direction, and repeat steps 5-1 to
[0171] In step 5-3, train another pair of generators and discriminators, convert the generated samples in each data domain into samples in another domain, train the overall model again, cyclically convert between domain samples, supplement the minority samples in the unbalanced dataset, construct a new balanced dataset, and train the model using the balanced sample set generated across domains.
[0172] Training other models with balanced sample sets generated across domains will improve the performance of the models on the original imbalanced dataset.
[0173] In order to compare the effectiveness of the fault diagnosis method for hydropower unit unbalanced data set generated based on adaptive domain transformation provided by the present invention, in the experiment, the trained model was applied to the cross-domain data generation transformation, and the results are shown in the figure. Figure 2-5 , Figure 2 , 4 are the sample time domain waveforms of domain 1 and domain 2 respectively, Figure 3 , 5 is the spectral characteristics of the corresponding signal, where (a) is the real result in the data domain, and (b) is the generated sample result converted from another data domain by the generative model. It can be seen that the generated samples are very similar to the real samples and have diversity. They are not exactly the same data, indicating that the common pattern collapse problem of the generative model does not occur.
[0174] The generated samples are further used for fault identification, and the generated domain sample data set is obtained using the generated model. The deep fault diagnosis model is trained directly and completely using the generated data set without using the real data set for training. The model is tested and evaluated using the real data set to analyze the model diagnosis accuracy. The experimental results of the method provided by the present invention and the latest research method are shown in Table 1:
[0175] Table 1:
[0176]
[0177]
[0178] From the results in the table, we can see that the accuracy of the proposed method reaches 95%, while other methods can basically only reach about 90%, which means that the samples generated by the proposed method have a higher similarity with the real data, which verifies the performance of the model. Figure 6 It is the result of visualizing the features obtained after model training through T-SNE dimensionality reduction. The features of the real data and the generated data basically overlap in each subclass, showing that the generated data and the real data are consistent in spatial distribution.
[0179] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the field without departing from the purpose of the present invention. These changes involve related technologies well known to those skilled in the art, which all fall within the scope of protection of the patent of the present invention.
[0180] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.
Claims
1. A fault diagnosis method for hydropower unit unbalanced data set generated by adaptive domain transformation, characterized in that: The method comprises the following steps: Step 1: Construct basic sample sets of source domain and target domain; The source domain includes several different types of fault samples and well-labeled data sets, and the target domain is a data set with an unbalanced number of samples; Step 2: Data preprocessing; Perform data normalization on the basic sample set in step 1 to obtain data that is distributed in [-1,1] as a whole, and then use a deep convolutional network to extract signal features and construct a two-dimensional data input model; Step 3: Establish a domain transformation sample generation model based on a cross-domain adversarial generative network; The domain transformation sample generation model includes a forward generator G, a target domain distinguisher Dt, a reverse generator F, and a source domain distinguisher Ds, which can realize the forward conversion of source domain samples into target domain samples and the reverse conversion of target domain samples into source domain samples; Step 4: Construct the loss functions of the generator and the discriminator, and establish the loss functions of the composite Wasserstein-GP loss and the least squares loss; Used to evaluate the difference between generated samples and real samples; Step 5: Train the model. A gradient optimization algorithm is used to optimize the bidirectional paired generator and discriminator, and the training model realizes cross-domain bidirectional sub-class sample conversion and cross-domain sub-domain corresponding alignment generation.
2. The method for fault diagnosis of hydropower unit unbalanced data set generated by adaptive domain transformation according to claim 1 is characterized in that: The step 1 comprises the following steps: Step 1-1: Get the original data, build a data set, divide the vibration signal into several sample segments, the data length is 1024, each sample signal must contain several vibration cycles to fully express the fault information; Step 1-2: The source domain and the target domain have the same number of sample categories, wherein the source domain has a large number of samples in each category, and the target domain has a large number of normal category samples.
3. The method for fault diagnosis of unbalanced data set of hydropower units generated by adaptive domain transformation according to claim 2 is characterized in that: The step 2 comprises the following steps: Step 2-1: Map the monitoring data in step 1 to the specified range. The data normalization process adopts the Min-Max normalization method to obtain data with an overall distribution in [-1,1]. The formula is as follows: Among them, x max is the maximum value in the data set, x min is the minimum value in the data set, x new is the normalized data, and x is the original data; Step 2-2: Convert each sample segment of length 1024 into 32×32 graph structure data.
4. The method for fault diagnosis of unbalanced data set of hydropower units generated by adaptive domain transformation according to claim 3 is characterized in that: The step 3 comprises the following steps: A bidirectional generative network deep model and a domain sample distinguishing deep model are established. First, a bidirectional generative network deep model is constructed, which is built with different structural layers, including a downsampling layer, a residual structure layer, and an upsampling layer; secondly, a domain sample distinguishing deep model is constructed, and the domain sample distinguishing deep model includes a sample generator model and a distinguisher model.
5. The method for fault diagnosis of unbalanced data set of hydropower units generated by adaptive domain transformation according to claim 4 is characterized in that: The downsampling layer consists of a 2D convolution structure, the input is the constructed two-dimensional data matrix x, and the output is x', and its formula is as follows: Where conv represents convolution operation, bn represents regularization, relu represents nonlinear processing using relu activation function, w, b represent convolution kernel and bias, and the main parameters of convolution are: kernel_size = (3,3), strides = (2,2), padding = 'same' during convolution; The residual structure layer consists of the main convolution channel and the residual connection, and its formula is as follows: Where x is the module input feature, o is the residual feature output, and the main convolution parameters are: kernel_size = (3, 3), strides = (1, 1); The upsampling layer performs feature upsampling on the input feature map, and its formula is as follows: Where x is the module input feature, x' is the residual feature output, and the main convolution parameters are: kernel_size = (3, 3), strides = (2, 2).
6. The method for fault diagnosis of hydropower unit unbalanced data set generated by adaptive domain transformation according to claim 5 is characterized in that: The sample generator model is composed of basic convolution, regularization layer, and independent modules in the order of input and output. The input of the sample generator model is the data of the sample in the source domain processed by step 2, and its dimension is (32, 32). The output is the sample in the target domain, and its dimension is (32, 32). The structure of the sample generator model is as follows: Where LeakyReLU(x)=max(0,x)+a*min(0,x), a is used to control the slope of the negative part and change the output range; The discriminator model is matched with the sample generator model, and the sample input in the data set is subjected to feature extraction to achieve class differentiation. The structure of the discriminator model is as follows: The parameters of each branch of the discriminator model are set to (1) 1×1conv (2) 3×3conv, 5×5conv (3)7×7conv,11×11conv (4) 3×3maxpool, 1×1conv Among them, Conv means convolution and maxpool means maximum pooling.
7. The method for fault diagnosis of unbalanced data set of hydropower units generated by adaptive domain transformation according to claim 1 is characterized in that: The generator (G, F) and the discriminator (Dt, Ds) of the bidirectional generation have the same structure.
8. The method for fault diagnosis of unbalanced data set of hydropower units generated by adaptive domain transformation according to claim 6 is characterized in that: The step 4 comprises the following steps: Step 4-1: Establish Wasserstein-GP loss, for the source domain sample x s , target domain sample x t , For the generated samples, the loss function with gradient penalty is as follows: where ∈~U[0,1],L D is the discriminator loss, L G is the generator loss; Step 4-2: Establish the least squares loss based on the last dimension element y of the output vector of the discriminator D e , the formula is: L GLS =(D(G(x s ))-1) 2 L LDS =(D(x t )-1) 2 ; Step 4-3: Fusion of two sets of loss functions, the total loss target of the generator and the discriminator is: L D =L DGP +L DLS L G =L GGP +L GLS .
9. The method for fault diagnosis of unbalanced data set of hydropower units generated by adaptive domain transformation according to claim 8 is characterized in that: The step 5 comprises the following steps: Step 5-1: Model training starts with the input sample. The source domain sample is forward transformed by the generator to obtain the generated target domain sample. Step 5-2: According to the real samples and generated samples of the target domain, calculate the loss function L of the generator and the discriminator according to step 4 G , L D ; Step 5-3: Use the deep learning gradient optimization algorithm Adam, set the learning rate to 0.0002, train the iterative generator and discriminator, and select 300-500 iterations; Step 5-4: Swap the source domain and the target domain, change the data generation direction, repeat steps 5-1 to 5-3, train another pair of generators and discriminators, convert the generated samples in each data domain into samples in another domain, train the overall model again, cyclically convert between domain samples, supplement the minority samples in the unbalanced dataset, build a new balanced dataset, and train the model using the balanced sample set generated across domains.
Citation Information
Cited By
Motor fault diagnosis method for unbalanced data in noise environment based on dual-scale network
CN121350417A
