Hydroelectric unit fault diagnosis method and system based on transfer learning and PSO algorithm

Abstract

This invention discloses a method and system for fault diagnosis of hydropower units based on transfer learning and the PSO algorithm, belonging to the field of hydropower unit fault diagnosis technology. It includes: using hydropower unit fault vibration data under design operating conditions as source domain data, and hydropower unit fault vibration data under varying operating conditions as auxiliary domain data, with fault type as label, to obtain a source domain sample set and a target domain / auxiliary domain sample set; using the source domain sample set for supervised training of a fault diagnosis model, determining the connection weights of the entire model; using the target domain auxiliary sample set for supervised training of a fault diagnosis transfer learning model, adjusting the connection weights of the fully connected layers and the classification layer; using the particle swarm optimization algorithm to optimize the hyperparameters of the fault diagnosis model's GRU network model, determining the optimal hyperparameters; and inputting the test samples into the trained fault diagnosis model to perform hydropower unit fault diagnosis. This invention addresses the problem of low diagnostic accuracy due to a lack of samples under certain operating conditions of hydropower units through transfer learning.

Description

Technical Field

[0001] This invention belongs to the field of hydropower unit fault diagnosis technology, and more specifically, relates to a hydropower unit fault diagnosis method and system based on transfer learning and PSO algorithm. Background Technology

[0002] Hydropower units are the most critical core equipment in a hydropower system. Their operational status directly affects the safe and stable operation of the entire power system, as well as the economic benefits and safe production of the hydropower station. Therefore, research on condition monitoring and fault diagnosis technologies for hydropower units is of significant theoretical and engineering application value for ensuring the safety and economic benefits of the power station and maintaining the safe and stable operation of the power grid. With the development of signal processing and artificial intelligence technologies, fault diagnosis of hydropower units is gradually developing towards intelligent diagnosis, online diagnosis, and remote diagnosis.

[0003] Deep learning-based fault diagnosis techniques can directly establish diagnostic models for hydropower units from signal data. For example, Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks have shown good results in equipment fault diagnosis. However, these methods require sufficient labeled data for model training and must include complete health status information of the hydropower equipment. Due to the significant losses associated with shutdowns for health checks, the number of fault samples obtainable during actual operation of hydropower units is extremely small compared to normal samples, and the obtained samples are usually unlabeled, lacking historical training data for the target fault type.

[0004] Transfer learning uses readily available fault samples (source domain) for training. Provided the source and target domains follow the same data distribution, it can help identify and diagnose fault data (target domain) where labels are difficult to obtain. This makes it effective in real-world engineering scenarios where fault samples are scarce or lack labels. Currently, the most common transfer learning method for handling fault data with missing labels is fine-tuning. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for fault diagnosis of hydropower units based on transfer learning and PSO algorithm, in order to solve the problem of low diagnostic accuracy caused by lack of samples under certain operating conditions of hydropower units.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for fault diagnosis of hydropower units based on transfer learning and the PSO algorithm, the method comprising:

[0007] S1. Take the fault vibration data of the hydropower unit under design conditions as the source domain data, take the fault vibration data of the hydropower unit under variable conditions as the auxiliary domain data, and take the fault type as the label to obtain the source domain sample set and the target domain auxiliary domain sample set.

[0008] S2. A fault diagnosis model is trained under supervision using a source domain sample set, and the connection weights of the entire model are determined. The fault diagnosis model includes a GRU network model, a fully connected layer, and a Softmax regression model. The GRU network model is used to extract features from the preprocessed input sample data to obtain a set of reconstructed feature vectors. The fully connected layer is used to transform the reconstructed feature vectors into a linear combination of features and output them to the Softmax regression model. The Softmax regression model is used to classify faults in hydropower units and calculate the probability of each type of fault occurring.

[0009] S3. Supervised training of the fault diagnosis transfer learning model is conducted using a target domain auxiliary sample set, and the connection weights of the fully connected layer and the classification layer are adjusted.

[0010] S4. The hyperparameters of the fault diagnosis model GRU network model are optimized using the particle swarm optimization algorithm to determine the optimal hyperparameters;

[0011] S5. Input the sample to be tested into the trained fault diagnosis model to perform fault diagnosis of hydropower units.

[0012] Preferably, step S1 specifically includes:

[0013] S11. The collected hydropower unit fault signals are divided into source domain sample set and target domain auxiliary sample set according to different operating conditions. The two sample sets are merged to obtain a combined set. The combined set samples are randomly divided to obtain the target domain test sample set.

[0014] S12. Perform wavelet decomposition on the fault signals of hydropower units in the above sample set, and normalize each decomposed signal to obtain sample data.

[0015] S13. Perform One-Hot encoding on the fault signal tags of the hydropower unit.

[0016] Preferably, the wavelet decomposition decomposes the original fault signal of the hydropower unit into N simple wavelets, which are then normalized and input into the GRU network model for feature extraction.

[0017] Preferably, the GRU network model uses a unidirectional structure and incorporates residual connections.

[0018] Preferably, in step S2, the Dropout training method is incorporated to control the parameter changes of the neural units:

[0019] When fine-tuning the parameters of the GRU network model, a neural network unit is temporarily discarded from the network with a certain probability; that is, the probability determines whether to ignore the network unit in the current round of input calculation, so that it does not participate in the current round of calculation, and in the next round of calculation, the probability determines whether to participate in the calculation again.

[0020] Preferably, the functional expression of the Softmax regression model is:

[0021]

[0022] Where z represents the feature vector of various faults, σ(z) j Let z be the fuzzy evaluation value for the j-th type of fault. j z represents the feature combination value of the j-th fault type. k represents the feature combination value of the k-th fault type, and K represents the total number of categorizable faults.

[0023] Preferably, the hyperparameters include the number of neurons in the GRU layer, the learning rate, the time step, and the number of nodes in the fully connected layer.

[0024] Preferably, the fitness function of the particle swarm optimization algorithm is:

[0025]

[0026] Where fitness represents fitness, n represents the number of iterations of the hydropower unit fault diagnosis pre-model, and accuracy represents accuracy. i This represents the training accuracy of the source domain sample set in the i-th iteration.

[0027] To achieve the above objectives, in a second aspect, the present invention provides a hydropower unit fault diagnosis system based on transfer learning and PSO algorithm, comprising: a processor and a memory; the memory for storing computer execution instructions; and the processor for executing the computer execution instructions, such that the method described in the first aspect is executed.

[0028] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0029] (1) To get rid of the dependence on labeled fault data and target fault type training data, this invention proposes a hydropower unit fault diagnosis method and system based on transfer learning and PSO algorithm. Under the design working condition, a large number of labeled samples can be obtained and used as source domain data for training. Under the variable working condition, only a small number of labeled samples can be obtained and used as target domain auxiliary samples for fine-tuning of the transfer model. Thus, hydropower unit fault diagnosis can be effectively performed when fault data labels are lacking or missing. Furthermore, the PSO algorithm is applied to the hyperparameter tuning of deep learning to reduce the manual work of hyperparameter selection, thereby enhancing the intelligent learning ability of the whole method and achieving the ability to generalize to a variety of similar problems.

[0030] (2) In the preferred embodiment of the present invention, the Dropout method is used in the unsupervised learning process to avoid the problem of overfitting in the deep learning model. Attached Figure Description

[0031] Figure 1 This is a flowchart of a hydropower unit fault diagnosis method based on transfer learning and PSO algorithm provided by the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] like Figure 1 As shown, this invention provides a fault diagnosis method for hydropower units based on transfer learning and the PSO algorithm. The method includes:

[0034] Step S1. Take the hydropower unit fault vibration data under design conditions as the source domain data, take the hydropower unit fault vibration data under variable conditions as the auxiliary domain data, and take the fault type as the label to obtain the source domain sample set and the target domain auxiliary domain sample set.

[0035] Preferably, step S1 specifically includes:

[0036] S11. The collected hydropower unit fault signals are divided into source domain sample set and target domain auxiliary sample set according to different operating conditions. The two sample sets are merged to obtain a combined set. The combined set samples are randomly divided to obtain the target domain test sample set.

[0037] S12. Perform wavelet decomposition on the fault signals of hydropower units in the above sample set, and normalize each decomposed signal to obtain sample data.

[0038] S13. Perform One-Hot encoding on the fault signal tags of the hydropower unit.

[0039] Some of the codes are shown in Table 1.

[0040] Table 1 Fault Type Coding

[0041]

[0042] Preferably, the wavelet decomposition decomposes the original fault signal of the hydropower unit into N simple wavelets, which are then normalized and input into a GRU (Gated Recursive Unit) network model for feature extraction.

[0043] Normalize the data:

[0044]

[0045] Where, x max x is the maximum value of the i-th sample data. min Let x be the minimum value of the i-th sample data. i For the i-th sample, x′ i These are the samples after normalization.

[0046] Finally, the processed sample set is divided into a source domain training set, a target domain auxiliary sample set, and a target domain test set. The specific division process should be determined based on the actual data collection situation.

[0047] Wavelet transform is a transform method that inherits and develops the idea of ​​localization in short-time Fourier transform while overcoming its shortcomings. Wavelet transform performs localized analysis of time (space) and frequency by progressively refining the signal (function) at multiple scales through scaling and translation operations, ultimately achieving time subdivision at high frequencies and frequency subdivision at low frequencies, automatically adapting to the requirements of time-frequency signal analysis. The specific transform is as follows:

[0048] 1. Take the inner product of the initial part of the wavelet w(t) and the original function f(t), calculate the similarity between this part of the function and the wavelet, and represent it with coefficient C.

[0049] 2. Shift the wavelet to the right by k units to obtain the wavelet w(tk). Repeat step 1 until the function f ends.

[0050] 3. Extend the wavelet w(t) to obtain the wavelet w(t / 2), and repeat steps 1-2.

[0051] 4. Continue to expand the wavelet and repeat steps 1-3.

[0052] The general form of a wavelet function is:

[0053]

[0054] Wavelet transform formula:

[0055]

[0056] Step S2. Supervised training of the fault diagnosis model using the source domain sample set, determining the connection weights of the entire model. The fault diagnosis model includes: a GRU network model, a fully connected layer, and a Softmax regression model. The GRU network model is used to extract features from the preprocessed input sample data to obtain a set of reconstructed feature vectors. The fully connected layer is used to transform the reconstructed feature vectors into a linear combination of features and output them to the Softmax regression model. The Softmax regression model is used to classify faults in hydropower units and calculate the probability of each type of fault occurring.

[0057] The GRU includes two gating mechanisms: a reset gate and a resetting gate. t and update gate z t When the input sequence is x t At that time, the outputs of the two gated and hidden states are as follows:

[0058] r t =σ(x t W r +H t-1 U r +b r )

[0059] z t =σ(x t W z +H t-1 U z +b z )

[0060] H t =(1-Z) t )⊙H t-1 +Z t ∩tanh(x t W h +R t ⊙H t-1 U h +b h )

[0061] Among them, W r W z W h U represents the weight matrix between the reset gate and the input, the weight matrix between the update gate and the input, and the weight matrix between the current hidden state and the input, respectively. r U z U hH represents the weight matrix between the reset gate and the previous hidden layer, the weight matrix between the update gate and the previous hidden layer, and the weight matrix between the current hidden state and the previous hidden layer, respectively. t-1 Let σ be the hidden state at the previous time step, and σ be the Logistic function.

[0062] Preferably, the GRU network model uses a unidirectional structure and incorporates residual connections, with the step size T set to N.

[0063] Preferably, the functional expression of the Softmax regression model is:

[0064]

[0065] Where z represents the feature vector of various faults, σ(z) j Let z be the fuzzy evaluation value for the j-th type of fault. j z represents the feature combination value of the j-th fault type. k represents the feature combination value of the k-th fault type, and K represents the total number of categorizable faults.

[0066] The Adam algorithm is used to update the parameters of the GRU network model, ensuring that the model converges as quickly as possible. The Adam optimization algorithm is as follows:

[0067] Based on the first-moment and second-moment estimates of the calculated gradient, the learning rate of each parameter in the network can be adaptively adjusted.

[0068] m t =β1m t-1 +(1-β1)g t

[0069]

[0070]

[0071]

[0072]

[0073] Where, m′ t For moment estimation, v′ t For second-order moment estimation; m′ t For m t t correction, v′ t For v t Correction; β1 is the exponential decay rate estimated by the first moment; β2 is the exponential decay rate estimated by the second moment; θ t These are the weights and thresholds for the neural network model. Typically, the default values ​​are ε = 10⁻⁸, α = 0.01, β₁ = 0.9, and β₂ = 0.999.

[0074] Preferably, in step S2, the Dropout training method is incorporated to control the parameter changes of the neural units:

[0075] To avoid overfitting, when fine-tuning the parameters of the GRU network model, neural network units are temporarily discarded from the network with a certain probability; that is, the probability determines whether to ignore the network unit in the current round of input calculation, so that it does not participate in the current round of calculation, and in the next round of calculation, the probability determines whether to participate in the calculation again.

[0076] Step S3. Supervised training of the fault diagnosis transfer learning model is conducted using the target domain auxiliary sample set, and the connection weights of the fully connected layer and the classification layer are adjusted.

[0077] The parameters of the GRU neural network layers in the model are frozen, the target domain dataset is input into the model, and the connection weights of the fully connected layers and classification layers are fine-tuned using gradient descent and backpropagation algorithms to obtain the transfer model. The GRU neural network layer parameters are then unfrozen to determine the final fault diagnosis transfer model.

[0078] Backpropagation is a gradient-based optimization process. The results of Softmax classification are compared with the training data labels to find the feature combination value j corresponding to the correct fault type. The gradient is then calculated using the following formula:

[0079]

[0080] Where, Δ i Let z be the i-th element of the gradient vector Δ, j be the index of the actual fault type corresponding to the label, and z be the index of the gradient vector Δ. j z is the feature combination value corresponding to the actual fault type number. i To be with Δ i The characteristic combination value of the fault type corresponding to the serial number, z k This represents the feature combination value corresponding to the kth fault type.

[0081] Step S4. Use the particle swarm optimization algorithm to optimize the hyperparameters of the fault diagnosis model GRU network model and determine the optimal hyperparameters.

[0082] Preferably, the hyperparameters include the number of neurons in the GRU layer, the learning rate, the time step, and the number of nodes in the fully connected layer.

[0083] 1) Initialization: Set the maximum number of iterations, the number of independent variables in the objective function, and the maximum velocity of the particles; set the particle swarm size to M, and randomly initialize a flight velocity for each particle.

[0084] 2) Calculate fitness:

[0085]

[0086] Where fitness represents fitness, n represents the number of iterations of the hydropower unit fault diagnosis pre-model, and accuracy represents accuracy. i This represents the training accuracy of the source domain sample set in the i-th iteration.

[0087] 3) Particle position and velocity update:

[0088] V id =ωV id +C1r(0,1)(P id -X id )+C2r(0,1)(P gd -X id )

[0089] X id =X id +V id

[0090] Among them, V id Let X be the velocity of the individual in the d-th dimension of the i-th variable. id Let ω be the position of the individual in the d-th dimension of the i-th variable; ω is the inertia factor, r(0,1) represents any number between 0 and 1, C1 is the individual optimization factor, and C2 is the global optimization factor of the particle, usually set C1 = C2 = 2; P id Let P be the individual extreme value of the i-th variable in the d-th dimension. gd Let d represent the d-th dimension of the global optimal solution.

[0091] 4) Termination condition: Stop when the number of iterations (epochs) is reached or the error requirement is met. The parameters obtained are the optimal parameters of the GRU model.

[0092] Step S5. Input the sample to be tested into the trained fault diagnosis model to perform fault diagnosis of the hydropower unit.

[0093] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of hydroelectric generating units based on transfer learning and PSO algorithm, characterized in that, The method includes: S1. Take the fault vibration data of the hydropower unit under design conditions as the source domain data, take the fault vibration data of the hydropower unit under variable conditions as the auxiliary domain data, and take the fault type as the label to obtain the source domain sample set and the target domain auxiliary domain sample set. S2. A fault diagnosis model is trained under supervision using a source domain sample set, and the connection weights of the entire model are determined. The fault diagnosis model includes a GRU network model, a fully connected layer, and a Softmax regression model. The GRU network model is used to extract features from the preprocessed input sample data to obtain a set of reconstructed feature vectors. The fully connected layer is used to transform the reconstructed feature vectors into a linear combination of features and output them to the Softmax regression model. The Softmax regression model is used to classify faults in hydropower units and calculate the probability of each type of fault occurring. S3. Supervised training of the fault diagnosis transfer learning model is conducted using a target domain auxiliary sample set, and the connection weights of the fully connected layer and the classification layer are adjusted. S4. The hyperparameters of the fault diagnosis model GRU network model are optimized using the particle swarm optimization algorithm to determine the optimal hyperparameters; S5. Input the sample to be tested into the trained fault diagnosis model to perform fault diagnosis of hydropower units; The fitness function of the particle swarm optimization algorithm is: in, Indicates fitness. This indicates the number of iterations in the pre-model for fault diagnosis of hydropower units. Indicates the first Accuracy of training the source domain sample set in the next iteration.

2. The method as described in claim 1, characterized in that, Step S1 specifically includes: S11. The collected hydropower unit fault signals are divided into source domain sample set and target domain auxiliary sample set according to different operating conditions. The two sample sets are merged to obtain a combined set. The combined set samples are randomly divided to obtain the target domain test sample set. S12. Perform wavelet decomposition on the fault signals of hydropower units in the above sample set, and normalize each decomposed signal to obtain sample data. S13. Perform One-Hot encoding on the fault signal tags of the hydropower unit.

3. The method of claim 2, wherein, The wavelet decomposition process decomposes the original fault signal of the hydropower unit into N simple wavelets, which are then normalized and input into the GRU network model for feature extraction.

4. The method of claim 1, wherein, The GRU network model uses a unidirectional structure and incorporates residual connections.

5. The method of claim 1, wherein, In step S2, the Dropout training method is incorporated to control the parameter changes of the neural units: When fine-tuning the parameters of the GRU network model, a neural network unit is temporarily discarded from the network with a certain probability; that is, the probability determines whether to ignore the network unit in the current round of input calculation, so that it does not participate in the current round of calculation, and in the next round of calculation, the probability determines whether to participate in the calculation again.

6. The method of claim 1, wherein, The functional expression of the Softmax regression model is: in, Feature vectors representing various faults For the first Fuzzy evaluation values ​​for various types of faults Indicates the first Feature combination values ​​for various fault types Indicates the first Feature combination values ​​for various fault types This represents the total number of categorizable faults.

7. The method of claim 1 wherein, The hyperparameters include the number of neurons in the GRU layer, the learning rate, the time step, and the number of nodes in the fully connected layer.

8. A hydroelectric unit fault diagnosis system based on transfer learning and a PSO algorithm, characterized in that, include: Processor and memory; The memory is used to store computer-executed instructions; The processor is configured to execute the computer execution instructions, causing the method described in any one of claims 1 to 7 to be executed.

Images