Generator excitation system fault detection method and system

Abstract

The invention provides a generator excitation system fault detection method and system, and the method comprises the steps: firstly collecting the historical data of a generator excitation system, and forming a historical data set; then, constructing a stacked sparse auto-encoder network, taking the historical data set as input, and training the stacked sparse auto-encoder network in combination with a physical constraint equation of the generator excitation system; and finally, collecting real-time data of a generator excitation system, inputting the real-time data into the trained stacked sparse auto-encoder network, calculating an SPE index, and carrying out fault diagnosis through an SPE index control limit. Physical information is fused in the model training process, the detection result has high accuracy, the interpretability of the result is enhanced, and compared with a pure data driving method, the method has the advantages that dependence on large-scale fault samples is remarkably reduced, and the detection accuracy is greatly improved in an unknown scene or a scene without faults. High detection precision and robustness can still be kept, and the generalization ability and engineering applicability of the model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of generator fault detection, and in particular to a generator excitation system fault detection method and system. Background Art

[0002] Generators are critical equipment in power systems, and their operating status directly impacts the stable supply of electricity and system safety. The excitation system, a crucial component of a generator, directly influences its steady-state and dynamic performance, and thus, the stable supply of electricity and system safety. However, due to factors such as long-term operation, load fluctuations, temperature changes, and external interference, the excitation system is prone to failures such as excitation regulator failure, excitation winding short circuits, rectifier failures, and abnormal excitation currents. Failure to promptly detect and address these failures can lead to a failure in the generator's excitation control, causing voltage fluctuations, power oscillations, and even serious accidents such as demagnetization. Therefore, efficient and accurate excitation system fault detection methods are crucial to ensuring the safe and stable operation of generators.

[0003] Traditional methods for fault detection in generator excitation systems mainly include physical model-based methods and data-driven methods. Physical model-based methods rely on mathematical modeling and physical mechanism analysis of the generator. They establish state equations or equivalent circuit models and combine signals such as vibration, temperature, current, and voltage to perform fault diagnosis. However, this method is highly dependent on system parameters, has high modeling complexity, and is difficult to adapt to complex operating conditions. Data-driven methods collect operating data and use statistical analysis, signal processing, or machine learning techniques to identify faults, such as wavelet transforms, support vector machines, and random forests. These methods can overcome the limitations of physical modeling to a certain extent, but they rely on large amounts of fault data and have weak interpretability, making it difficult to improve detection accuracy using existing physical knowledge.

[0004] In recent years, deep learning technology has been widely studied in the field of fault detection. Autoencoders, due to their excellent feature learning capabilities, have been widely used for unsupervised anomaly detection. However, traditional autoencoder models fail to fully incorporate the physical characteristics of generators, resulting in insufficient generalization capabilities when using small samples or complex operating conditions. Summary of the Invention

[0005] Purpose of the invention: The first purpose of the present invention is to provide a generator excitation system fault detection method based on physical information combined with stacked sparse autoencoders to improve detection accuracy. The second purpose is to provide a generator excitation system fault detection system.

[0006] Technical solution: A method for detecting faults in a generator excitation system, comprising the following steps:

[0007] S1. Collect historical data of the generator excitation system to form a historical data set;

[0008] S2. Build a stacked sparse autoencoder network, use the historical data set obtained in step S1 as input, and train the stacked sparse autoencoder network in combination with the physical constraint equations of the generator excitation system;

[0009] S3. Collect real-time data of the generator excitation system, input the data into the stacked sparse autoencoder network trained in step S2, and calculate the SPE index. If the SPE index exceeds the residual abnormal control limit, it is determined that the generator excitation system has a fault. If the SPE index does not exceed the residual abnormal control limit, it is determined that the generator excitation system has not a fault, and the fault detection result is output.

[0010] Specifically, the physical constraint equations of the generator excitation system include: the relationship equation between excitation current and excitation voltage, the relationship equation between active power and reactive power, and the relationship equation between excitation current and stator current.

[0011] Specifically, the relationship equation between the excitation current and the excitation voltage is:

[0012]

[0013] Where: R f is the resistance of the excitation winding; L f is the inductance of the excitation winding; I f is the current of the excitation winding, t is the time;

[0014] The relationship equation between active power and reactive power is:

[0015]

[0016] Where: cos(φ) is the power factor, P is the active power, and Q is the reactive power;

[0017] The relationship equation between the excitation current and the stator current is:

[0018] ψ=L f I f +L s (I A +I B +I C )

[0019] Where: ψ is the magnetic flux of the generator, L s is the inductance of the stator winding, I A ,I B ,I C is the stator current.

[0020] Specifically, step S2 includes:

[0021] S21. Construct a stacked sparse autoencoder network consisting of an input layer, an output layer, and several hidden layers, and construct the physical constraint equations of the generator excitation system;

[0022] S22. Convert the physical constraint equation of the generator excitation system into a physical loss term, and introduce the differentiable physical loss term into the loss function of the stacked sparse autoencoder network;

[0023] S23. Complete the training of the stacked sparse autoencoder network through layer-by-layer unsupervised training and parameter optimization, and save the trained network parameters.

[0024] Specifically, the formula for the physical loss term is:

[0025]

[0026] Where: E p is the physical loss item; is the network output vector, is the network's reconstruction value for the input data, Data obtained by decoding the synthetic sample as a supplementary medium for physical constraints p l is the weight of each equality constraint in the physical loss term, l0 is the number of physical constraint equations, f l is the lth physical constraint equation.

[0027] Specifically, the loss function formula of the stacked sparse autoencoder network is:

[0028]

[0029] Where: Loss is the loss function, λ1,λ2,β,γ are weight factors, X (i) is the i-th input, Y (i) is the i-th output, N is the total number of samples, W is the weight value of the regularization term, is the average activation sparsity loss of the j-th neuron in the hidden layer, represents the average activation of the jth neuron; ρ is the expected sparsity set, and m is the total number of neurons.

[0030] Specifically, the formula of the SPE indicator is:

[0031]

[0032] Where: (xy) is the residual vector; To calculate the L2 norm of the residual vector.

[0033] Specifically, the formula for the residual abnormal control limit is:

[0034]

[0035] Where: is the residual normal control limit, z α is the upper quantile corresponding to the (1-α) confidence region, λ k is the kth eigenvalue of the sample covariance matrix.

[0036] Preferably, the above method further comprises the following steps:

[0037] S4. Calculate the root mean square error and determination coefficient of the stacked sparse autoencoder network fault detection results, and compare them with the corresponding set thresholds to judge the network accuracy. If the root mean square error is less than the set threshold and the determination coefficient is greater than the set threshold, the stacked sparse autoencoder network accuracy is determined to be qualified, and the fault detection accuracy rate is output.

[0038] The present invention also provides a generator excitation system fault detection system, comprising:

[0039] Data acquisition module: used to collect historical data of the generator excitation system to form a historical data set;

[0040] Network training module: used to build a stacked sparse autoencoder network. It uses the historical data set obtained by the data acquisition module as input and trains the stacked sparse autoencoder network based on the physical constraint equations of the generator excitation system.

[0041] Real-time detection module: used to collect real-time data of the generator excitation system, input the stacked sparse autoencoder network trained by the network training module, and calculate the SPE index. If the SPE index exceeds the residual abnormal control limit, it is determined that the generator excitation system has a fault. If the SPE index does not exceed the residual abnormal control limit, it is determined that the generator excitation system has not a fault and outputs the fault detection result.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following significant effects:

[0043] 1. This method utilizes a stacked sparse autoencoder that incorporates physical information for modeling. Building on traditional unsupervised feature extraction, it incorporates the physical constraints of the generator excitation system. This allows the model to not only mine deep-level features of the generator's operating state but also adhere to the physical laws of the engineering system. Compared to purely data-driven approaches, this method significantly reduces reliance on large-scale fault samples. It maintains high detection accuracy and robustness even in unknown or unseen fault scenarios, thereby improving the model's generalization and engineering applicability.

[0044] 2. Real-time and explainable fault diagnosis capabilities:

[0045] This invention uses the SPE metric to perform fault detection and health assessment on real-time operating data, enabling rapid, online anomaly identification of generator excitation systems. Furthermore, because the model incorporates physical information during training, its detection results are not only highly accurate but also more interpretable, making it easier for operators to understand the possible sources and evolution paths of faults, thereby enhancing the credibility of intelligent diagnosis and its practical engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the method of the present invention.

[0047] Figure 2 This is a topological diagram of the stacked sparse autoencoder that integrates physical information in the present invention.

[0048] Figure 3 It is a scatter plot comparison of the first dimension parameters of the present invention.

[0049] Figure 4 It is a scatter comparison diagram of the 12th dimension parameter of the present invention.

[0050] Figure 5 This is a comparison diagram of the true value-predicted value curve of the first dimension parameter of the present invention. DETAILED DESCRIPTION

[0051] A preferred embodiment of the present invention is further described below with reference to the accompanying drawings.

[0052] Example 1

[0053] See also Figure 1 As shown, this embodiment provides a generator excitation system fault detection method based on physical information combined with a stacked sparse autoencoder, comprising the following steps:

[0054] Offline modeling part:

[0055] S1. Collect historical system operation data and pre-process the data.

[0056] In a specific implementation scenario, the generator excitation system is connected to the power plant automation system, such as the DCS system. Historical data of the generator excitation system over a certain period of time is collected from the DCS system to form a historical data set. The collected historical data is then normalized:

[0057]

[0058] Where: X is the original data, X min is the minimum value of the feature, X max is the maximum value of the feature, X′ is the normalized data, and its value range is [0,1].

[0059] The normalized data constitutes the historical data set for subsequent model training.

[0060] S2. Build a stacked sparse autoencoder network, use the historical data set obtained by the data acquisition module as input, and train the stacked sparse autoencoder network based on the physical constraint equations of the generator excitation system.

[0061] See also Figure 2 As shown, in this embodiment, the normalized data is first mapped to the hidden layer H through the encoder to achieve data dimensionality reduction and extraction, and then the nonlinear principal components of the hidden layer are reconstructed to the output layer.

[0062] Specifically, the encoding and decoding mapping formulas mentioned above are:

[0063] H=σ h (W h x+b h )

[0064] y=Wσ o (W o H+b o )

[0065] Where: H is the output of the hidden layer, y is the output of the output layer, W h Represents the weight matrix between the input layer and the hidden layer, W o represents the weight matrix between the hidden layer and the output layer; σ h represents the hidden layer activation function, σ o represents the output layer activation function; b h represents the neuron activation bias vector of the hidden layer, b o Represents the bias vector of neuron activations in the output layer.

[0066] Next, we improved the loss function by introducing a sparsity regularization term based on the existing autoencoder structure. This regularization term increases the sparsity of the encoded representation and improves the information extraction capability by adding constraints on the distribution of neuronal activations. Specifically, the sparsity constraint is introduced by adding a KL divergence term, which means stacking multiple sparse autoencoders to form a stacked sparse autoencoder.

[0067] The calculation formula of KL divergence is:

[0068]

[0069] Where: represents the average activation of the jth neuron; ρ is the expected sparsity, which is usually taken as a small value to encourage most neurons to remain in an inactive state, and m is the total number of neurons.

[0070] After the introduction of sparsification, the loss function expression of the stacked sparse autoencoder is:

[0071]

[0072] The first term on the right side of the equation is the reconstructed mean square error of the input and output, the second term on the right side is the regularized elastic network penalty, which is a weighted linear combination of the L1 norm and L2 norm of Lasso and Ridge, and the third term on the right side is the KL divergence loss. The parameters in the formula are: Loss is the loss function, X (i) is the i-th input, Y (i) is the i-th output, N is the total number of samples, W is the weight value of the regularization term; λ1 and λ2 are regularization weight factors, and β is the coefficient of the sparsity penalty term.

[0073] Next, to further enhance the model's ability to express the internal structure and prior information of complex systems, the present invention designs a neural network modeling mechanism that integrates linear and nonlinear equality constraints, aiming to introduce relevant physical constraint information in engineering systems. In the case of linear constraints, the expression is:

[0074]

[0075] Where: m is the total number of samples, n is the number of physical constraint features corresponding to each sample, c is the sample index, r is the dimension index of each physical feature, α k is the coefficient of the linear equality constraint, z cr To constrain the relevant physical information characteristic variables, C is a constant term. Furthermore, its partial derivatives with respect to input parameters and network weights are derived as follows:

[0076]

[0077] Where: v ij represents the weighted sum of the input of the i-th sample to the j-th hidden layer neuron, w jk is the weight between the hidden layer and the output layer, b j is the bias term of the jth neuron, and σ is the activation function.

[0078] To further improve the modeling accuracy, the nonlinear equality constraints in the system are considered, and their expressions are as follows:

[0079]

[0080] Where: α rr is the quadratic term coefficient, q is another dimension index of the constraint feature, which is used to pair with the rth feature to construct the quadratic term z cr ,z cq. Furthermore, its partial derivatives with respect to input parameters and network weights can be derived as follows:

[0081]

[0082] In order to integrate the above constraints with the neural network training process, this method introduces a physical loss term, which is defined as follows:

[0083]

[0084] Where: E p is the physical loss item; is the network output vector, dimension is k, is the network's reconstruction value for the input data, Data obtained by decoding the synthetic sample as a supplementary medium for physical constraints p l is the weight of each equality constraint in the physical loss term, l0 is the number of physical constraint equations, f l is the lth physical constraint equation.

[0085] In this embodiment, the physical constraint equations include: an equation for the relationship between the excitation current and the excitation voltage, an equation for the relationship between the active power and the reactive power, and an equation for the relationship between the excitation current and the stator current.

[0086] The relationship equation between excitation current and excitation voltage is:

[0087]

[0088] Where: R f is the resistance of the excitation winding; L f is the inductance of the excitation winding; I f is the current of the excitation winding, and t is the time.

[0089] The relationship equation between active power and reactive power is:

[0090]

[0091] Where: cos(φ) is the power factor, P is the active power, and Q is the reactive power;

[0092] The relationship equation between the excitation current and the stator current is:

[0093] ψ=L f I f +L s (I A +I B +I C )

[0094] Where: ψ is the magnetic flux of the generator, L sis the inductance of the stator winding, I A ,I B ,I C is the stator current.

[0095] Finally, the physical information is integrated with the traditional reconstruction error and regularization term to construct a unified loss function as follows:

[0096]

[0097] Where: γ is the physical loss coefficient.

[0098] After training is completed, save the physical information stacked sparse autoencoder (PI-SSAE) network model.

[0099] Online monitoring part:

[0100] S3. Collect real-time data of the generator excitation system, input the data into the stacked sparse autoencoder network trained in step S2, and calculate the SPE index. If the SPE index exceeds the residual abnormal control limit, it is determined that the generator excitation system has a fault. If the SPE index does not exceed the residual abnormal control limit, it is determined that the generator excitation system has not a fault, and the fault detection result is output.

[0101] In this embodiment, real-time data of the generator excitation system operation is collected from the DCS system. After normalization, the real-time data is input into the physical information stacked sparse autoencoder (PI-SSAE) network model trained in step S2. The reconstruction error between the model output and input is then used to calculate the SPE indicator reflecting the current state of the generator excitation system.

[0102] The formula for the SPE indicator is:

[0103]

[0104] Where: (xy) is the residual vector; To calculate the L2 norm of the residual vector.

[0105] In order to realize the fault discrimination function of the SPE index, the present invention further sets the residual abnormal control limit of the SPE index abnormal discrimination The formula is:

[0106]

[0107] Where: is the residual normal control limit, z α is the upper quantile corresponding to the (1-α) confidence region, λ k is the kth eigenvalue of the sample covariance matrix.

[0108] The SPE index is continuously monitored. If the SPE index exceeds the residual abnormal control limit, it is determined that the generator excitation system has a fault. If the SPE index does not exceed the residual abnormal control limit, it is determined that the generator excitation system has not a fault and the fault detection result is output.

[0109] To verify the effectiveness of the present invention, the above method is applied in a specific implementation scenario.

[0110] Taking the generator excitation system in a power plant unit as an example, the generator excitation system has a total of 12 related parameters, namely: generator stator A phase current, generator stator B phase current, generator stator C phase current, generator stator AB phase line voltage, generator stator BC phase line voltage, generator stator CA phase line voltage, generator active power, generator reactive power, generator power factor, generator rotor temperature, generator excitation voltage and generator excitation current.

[0111] Collecting generator excitation system data from the DCS system The sampling interval is set to 60 seconds.

[0112] The PI-SSAE network model's training set consists of data from the 300MW-400MW and 700MW-1000MW sample load ranges, while the test set consists of data from the 400MW-700MW sample load range. These data are normalized and then fed into the PI-SSAE network model for training.

[0113] The PI-SSAE network model uses root mean square error and coefficient of determination to measure model accuracy. To demonstrate the improvement of the PI-SSAE network model over the traditional SSAE network model, the same dataset was used for training and testing. Figure 4 As shown, Figure 3 (a) is the scatter plot of the first-dimensional SSAE network model. Figure 3 (b) is a scatter plot of the first-dimensional PI-SSAE network model. It can be seen that the PI-SSAE network model provided by the present invention has good performance in two indicators (R 2 / RMSE) are much better than the SSAE network model without adding physical constraints; see Figure 5 As shown, Figure 4 (a) is a scatter plot of the 12th-dimensional SSAE network model. Figure 4 (b) is a scatter plot of the 12th-dimensional PI-SSAE network model. It can be seen that the PI-SSAE network model provided by the present invention has good performance in two indicators (R 2 / RMSE) are much better than the SSAE network model without adding physical constraints. Figure 5The figure below shows a true value-predicted value curve for the first dimension. It can be seen that the PI-SSAE network model provided by the present invention also significantly outperforms the SSAE network model without physical constraints in terms of prediction curve accuracy. A comparative test of fault diagnosis rates for common faults in motor excitation systems was conducted, yielding the results in Table 1 below.

[0114] Table 1

[0115]

[0116] As can be seen from Table 1, in the process of diagnosing seven different common faults, the diagnosis rate of the PI-SSAE network model provided by the present invention is much better than the existing PCA (principal component analysis) method and SSAE network model in all categories.

[0117] Example 2

[0118] This embodiment provides a generator excitation system fault detection system based on a physical information stacked sparse autoencoder, including:

[0119] Data acquisition module: used to collect historical data of the generator excitation system to form a historical data set;

[0120] Network training module: used to build a stacked sparse autoencoder network. It uses the historical data set obtained by the data acquisition module as input and trains the stacked sparse autoencoder network based on the physical constraint equations of the generator excitation system.

[0121] Real-time detection module: used to collect real-time data of the generator excitation system, input the stacked sparse autoencoder network trained by the network training module, and calculate the SPE index. If the SPE index exceeds the residual abnormal control limit, it is determined that the generator excitation system has a fault. If the SPE index does not exceed the residual abnormal control limit, it is determined that the generator excitation system has not a fault and outputs the fault detection result.

Claims

1. A method for detecting faults in a generator excitation system, characterized in that: The following steps are involved: S1. Collect historical data of the generator excitation system to form a historical data set; S2. Build a stacked sparse autoencoder network, use the historical data set obtained in step S1 as input, and train the stacked sparse autoencoder network in combination with the physical constraint equations of the generator excitation system; S3. Collect real-time data of the generator excitation system, input the data into the stacked sparse autoencoder network trained in step S2, and calculate the SPE index. If the SPE index exceeds the residual abnormal control limit, it is determined that the generator excitation system has a fault. If the SPE index does not exceed the residual abnormal control limit, it is determined that the generator excitation system has not a fault, and the fault detection result is output.

2. The generator excitation system fault detection method according to claim 1, characterized in that: The physical constraint equations of the generator excitation system include: an equation for the relationship between the excitation current and the excitation voltage, an equation for the relationship between the active power and the reactive power, and an equation for the relationship between the excitation current and the stator current.

3. The generator excitation system fault detection method according to claim 2, characterized in that: The relationship equation between the excitation current and the excitation voltage is: Where: R f is the resistance of the excitation winding; L f is the inductance of the excitation winding; I f is the current of the excitation winding, t is the time; The relationship equation between active power and reactive power is: Where: cos(φ) is the power factor, P is the active power, and Q is the reactive power; The relationship equation between the excitation current and the stator current is: ψ=L f I f +L s (I A +I B +I C ) Where: ψ is the magnetic flux of the generator, L s is the inductance of the stator winding, I A ,I B ,I C is the stator current.

4. The generator excitation system fault detection method according to claim 1, characterized in that: The step S2 comprises: S21. Construct a stacked sparse autoencoder network consisting of an input layer, an output layer, and several hidden layers, and construct the physical constraint equations of the generator excitation system; S22. Convert the physical constraint equation of the generator excitation system into a physical loss term, and introduce the differentiable physical loss term into the loss function of the stacked sparse autoencoder network; S23. Complete the training of the stacked sparse autoencoder network through layer-by-layer unsupervised training and parameter optimization, and save the trained network parameters.

5. The generator excitation system fault detection method according to claim 4, characterized in that: The formula for the physical loss term is: Where: E p is the physical loss item; is the network output vector, is the network's reconstruction value for the input data, Data obtained by decoding the synthetic sample as a supplementary medium for physical constraints p l is the weight of each equality constraint in the physical loss term, l0 is the number of physical constraint equations, f l is the lth physical constraint equation.

6. The generator excitation system fault detection method according to claim 5, characterized in that: The loss function formula of the stacked sparse autoencoder network is: Where: Loss is the loss function, λ1,λ2,β,γ are weight factors, X (i) is the i-th input, Y (i) is the i-th output, N is the total number of samples, W is the weight value of the regularization term, is the average activation sparsity loss of the jth neuron in the hidden layer, and m is the total number of neurons.

7. The generator excitation system fault detection method according to claim 1, characterized in that: The formula of the SPE index is: Where: (xy) is the residual vector; To calculate the L2 norm of the residual vector.

8. The generator excitation system fault detection method according to claim 1, characterized in that: The formula for the residual abnormal control limit is: Where: is the residual normal control limit, z α is the upper quantile corresponding to the (1-α) confidence region, λ k is the kth eigenvalue of the sample covariance matrix.

9. The generator excitation system fault detection method according to claim 1, characterized in that: The following steps are also included: S4. Calculate the root mean square error and determination coefficient of the stacked sparse autoencoder network fault detection results, and compare them with the corresponding set thresholds to judge the network accuracy. If the root mean square error is less than the set threshold and the determination coefficient is greater than the set threshold, the stacked sparse autoencoder network accuracy is determined to be qualified, and the fault detection accuracy rate is output.

10. A generator excitation system fault detection system, characterized in that: include: Data acquisition module: used to collect historical data of the generator excitation system to form a historical data set; Network training module: used to build a stacked sparse autoencoder network. It uses the historical data set obtained by the data acquisition module as input and trains the stacked sparse autoencoder network based on the physical constraint equations of the generator excitation system. Real-time detection module: used to collect real-time data of the generator excitation system, input the stacked sparse autoencoder network trained by the network training module, and calculate the SPE index. If the SPE index exceeds the residual abnormal control limit, it is determined that the generator excitation system has a fault. If the SPE index does not exceed the residual abnormal control limit, it is determined that the generator excitation system has not a fault and outputs the fault detection result.

Images