Phase modifier static eccentricity fault diagnosis method based on improved deep belief network

Abstract

The invention discloses a phase modifier static eccentricity fault diagnosis method based on an improved deep belief network, particularly relates to the technical field of electrical fault diagnosis, and solves the technical problems of low accuracy and high sample dependence degree of a phase modifier fault diagnosis method in the prior art. According to the technical scheme, electromagnetic torque, stator current and rotor current signals of the phase modifier in different fault states are obtained, signal decomposition is carried out on collected multi-source signals through fast fourier transform to obtain pre-features, feature extraction is carried out on the pre-features through ensemble empirical mode decomposition and permutation entropy, and the pre-features are extracted through the permutation entropy; fusing the extracted characteristic parameters to construct a fault sample set; according to the method, the operation state of the phase modifier can be reflected more comprehensively, the defect that single signal diagnosis information is insufficient is overcome, and the accuracy and reliability of fault diagnosis are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical fault diagnosis, and particularly relates to a static eccentricity fault diagnosis method for a synchronous condenser based on an improved deep belief network. Background Art

[0002] As a key device for the stable operation of a power system, a synchronous condenser mainly undertakes the important task of reactive power compensation and plays a crucial role in maintaining the voltage stability of the power system. However, during long-term continuous operation, the synchronous condenser will inevitably be affected by various complex factors such as mechanical wear and environmental factors, thus triggering various faults. Static eccentricity fault is one of the more common fault types of synchronous condensers. Once a static eccentricity fault occurs, the operating state of the synchronous condenser will be severely affected. These abnormal phenomena will not only reduce the operating efficiency and service life of the synchronous condenser itself, but also pose a serious threat to the safe and stable operation of the entire power system, and may even trigger chain reactions such as voltage fluctuations and equipment damage in a local power grid, affecting the reliable supply of electricity.

[0003] Currently, there are many limitations in the fault diagnosis methods for synchronous condensers. Traditional fault diagnosis methods based on single-signal analysis, such as relying only on vibration signals or current signals for diagnosis, are difficult to comprehensively reflect the fault characteristics of synchronous condensers due to the limited information contained in the signals, resulting in low diagnostic accuracy. And some deep learning-based methods, such as deep belief network (DBN), although having certain feature learning capabilities, the traditional DBN has the problem of randomly setting the initial weights, which makes the model easily fall into a local optimum during the training process, thus affecting the accuracy and generalization ability of fault diagnosis. In addition, the existing diagnostic methods rely highly on samples, and the diagnostic performance will drop significantly when the sample quantity is insufficient or the sample distribution is uneven. Therefore, there is an urgent need for a new fault diagnosis method to improve the accuracy and reliability of static eccentricity fault diagnosis for synchronous condensers and reduce the dependence on samples. Summary of the Invention

[0004] Therefore, the present invention solves the technical problems in the prior art of low accuracy of the static eccentricity fault diagnosis method for synchronous condensers, high dependence on samples, and the traditional DBN being easily trapped in a local optimum; the static eccentricity fault diagnosis method for a synchronous condenser based on an improved deep belief network provided by the present invention overcomes the defect of insufficient diagnostic information of a single signal and significantly improves the accuracy and reliability of fault diagnosis.

[0005] The static eccentricity fault diagnosis method for a synchronous condenser based on an improved deep belief network provided by the present invention includes the following steps:

[0006] Step 1: Establish a finite element simulation model and collect multi-source signals of the synchronous condenser under different working conditions;

[0007] Step 2: Perform preprocessing operations such as denoising and normalization on the collected multi-source signals;

[0008] Step 3: Decompose the multi-source signals using the Fast Fourier Transform (FFT) to obtain pre-features.

[0009] Step 4: Use Ensemble Empirical Mode Decomposition (EEMD) and Permutation Entropy (PE) to extract features from the pre-features.

[0010] Step 5: Fuse the permutation entropy feature parameters extracted from the electromagnetic torque, stator current, and rotor current signals to construct a fault sample set.

[0011] Step 6: Construct a Deep Belief Network (DBN) model.

[0012] Step 7: Use the Particle Swarm Optimization (PSO) algorithm to optimize the weights and thresholds of the DBN.

[0013] Step 8: Diagnosis of static eccentricity fault of the synchronous condenser.

[0014] Further, the multi-source signals collected in Step 1 include electromagnetic torque, stator current, and rotor current signals under the normal state and different fault degrees of the synchronous condenser.

[0015] Further in Step 2: Perform preprocessing operations such as denoising and normalization on the collected multi-source signals; it includes the following steps:

[0016] Step 2.1 Perform denoising processing on the collected multi-source signals. For the electromagnetic torque signal, adopt the wavelet threshold denoising method. Select the wavelet basis function (db4 wavelet) and the decomposition level (5 levels). By performing threshold processing on the wavelet coefficients, remove the wavelet coefficients caused by noise and reconstruct to obtain the denoised electromagnetic torque signal. For the current signal, adopt the method combining mean filtering and median filtering. First, use mean filtering to remove the DC bias and low-frequency noise, and then use median filtering to remove the impulse noise to obtain a smooth current signal.

[0017] Step 2.2 Perform normalization processing on the collected multi-source information. Normalize the denoised electromagnetic torque and current signals. Adopt the min-max normalization method to map the amplitude of the signal to the interval [0, 1]. For the electromagnetic torque signal X V , the normalization formula is

[0018]

[0019] where X Tmin and X Tmax are the minimum and maximum values of the electromagnetic torque signal respectively, and Y Vis the normalized electromagnetic torque signal. Similarly, for the current signal X i a similar normalization process is performed to obtain the normalized current signal Y i .

[0020] Furthermore, in step 3: the multi-source signal is decomposed by fast Fourier transform (FFT) to obtain pre-features.

[0021] It includes the following steps:

[0022] The electromagnetic torque of the synchronous condenser in the normal state is expressed as

[0023]

[0024] where: p is the number of pole pairs of the synchronous condenser; L is the effective axial length of the motor; Λ0 is the constant value of the air-gap permeance; F s is the fundamental armature reaction magnetomotive force generated by the stator winding; F r is the fundamental main magnetomotive force generated by the rotor excitation winding; is the internal power angle of the generator; R0 is the inner radius of the stator.

[0025] The electromagnetic torque of the synchronous condenser in the static eccentricity fault state is expressed as

[0026]

[0027] where: δ s is the relative static eccentricity angle; t is time.

[0028] Comparing the synchronous condenser in the normal state and the static eccentricity fault state, in addition to the DC component, the electromagnetic torque of the synchronous condenser also increases the second harmonic component. The static air-gap eccentricity will cause the steady-state electromagnetic torque of the generator to show the fluctuation characteristics of twice the fundamental frequency. As the degree of static air-gap eccentricity increases, the degree of double-frequency fluctuation will also increase. The electromagnetic torque signal is decomposed by fast Fourier transform (FFT), and the second harmonic component of the electromagnetic torque signal is extracted as the pre-feature of the electromagnetic torque signal.

[0029] Step 3.2 The static eccentricity of the synchronous condenser will destroy the symmetry of the magnetic field in the air gap of the motor, causing the magnetic field distribution inside the motor to change, generating a pulsating magnetomotive force that changes at twice the power supply frequency, and inducing a current with a frequency of 2 times in the stator winding; the static eccentricity will destroy the three-phase symmetry of the magnetic field, affecting the three-phase current balance relationship, and 3rd harmonics will appear in the stator current; the static eccentricity will cause non-linear phenomena such as local saturation of the magnetic field, and 5th, 7th and other higher harmonics will appear in the stator current. The stator current signal is decomposed by fast Fourier transform (FFT), and the 2nd, 3rd, 5th, and 7th harmonic components of the stator current signal are extracted as the pre-features of the stator current signal.

[0030] Step 3.3: The static eccentricity of the synchronous condenser will break the symmetry of the rotor magnetic field, causing a DC component to appear in the induced electromotive force of the rotor winding, thereby generating a DC component in the rotor current. Similar to the stator current, the rotor current may also have higher harmonics such as the 5th and 7th harmonics. The rotor current signal is decomposed using the Fast Fourier Transform (FFT) to extract the 5th and 7th harmonic components and the DC component of the rotor current signal as the pre-features of the rotor current signal.

[0031] Further, in Step 4: Ensemble Empirical Mode Decomposition (EEMD) and Permutation Entropy (PE) are used to extract features from the pre-features.

[0032] It includes the following steps:

[0033] Step 4.1: Input the pre-features collected from the electromagnetic torque and current signals into the EEMD algorithm. When the EEMD algorithm runs, white noise with different amplitudes is added to the original signal, and this white noise can make the signal show more obvious feature differences at different scales. Through multiple decompositions and averaging processes, several Intrinsic Mode Function (IMF) components are finally obtained.

[0034] Step 4.2: For each IMF component, calculate its permutation entropy. Taking a time series of an IMF component as an example, first perform phase space reconstruction, select appropriate embedding dimension m and time delay τ to obtain the phase space vector

[0035] X i =[x(i), x(i), …, x(i+(m - 1)τ)], i = 1, 2, …, N-(m - 1)τ (4)

[0036] For each vector X i Arrange its elements in ascending order to obtain

[0037] x(i1) ≤ x(i2) ≤ … ≤ x(i m ) (5)

[0038] where i1, i2…i m is a permutation of 1, 2…m, and this permutation is represented by π=(i1, i2…i m ). Calculate the permutation entropy from

[0039]

[0040] where p(π) is the probability of the permutation pattern occurring.

[0041] Further, in step 5, the permutation entropy characteristic parameters extracted from the electromagnetic torque, stator current, and rotor current signals are fused. In a series connection manner, the permutation entropy feature vectors of the second harmonic component of the electromagnetic torque signal, the 2nd, 3rd, 5th, and 7th harmonic components of the stator current signal, and the permutation entropy feature vectors of the 5th, 7th, and DC components of the rotor current are sequentially connected to form a high-dimensional feature vector. A fault sample set is constructed using the fused feature vector. The constructed fault sample set is divided into a training set, a validation set, and a test set according to a ratio of 7:2:1. The feature vectors in the fault sample set under different fault states (normal, slight static eccentricity, moderate static eccentricity, severe static eccentricity) are respectively marked with corresponding class labels, with the normal state marked as 0, the slight static eccentricity fault marked as 1, the moderate static eccentricity fault marked as 2, and the severe static eccentricity fault marked as 3.

[0042] Further, in step 6, a deep belief network (DBN) model is constructed, which is composed of multiple restricted Boltzmann machines (RBMs) stacked together. A DBN model containing two layers of RBMs is constructed. The number of visible layer nodes of the first layer of RBM is determined according to the dimension of the input feature vector, and the number of hidden layer nodes is set to 20; the number of visible layer nodes of the second layer of RBM is the number of hidden layer nodes of the first layer of RBM, and the number of hidden layer nodes is set to 15. Finally, a Softmax classification layer is added on top of the DBN model for outputting the fault diagnosis result, and the number of output layer nodes is 4 (corresponding to four fault states).

[0043] Further, in step 7: The particle swarm optimization (PSO) algorithm is used to optimize the weights and thresholds of the DBN.

[0044] It includes the following steps:

[0045] Step 7.1: Initially, 30 particles are randomly generated, and the position of each particle is randomly taken within a certain range. This position represents a set of weights and thresholds of the DBN model. The position of each particle contains the specific values of these weights and thresholds. The velocity of the particle is initialized to zero.

[0046] Step 7.2: Substitute the weights and thresholds corresponding to each particle into the DBN model, and use the training samples to train the model. During the training process, the cross-entropy loss function is used to calculate the error between the model prediction value and the true label, and the model parameters are updated through the backpropagation algorithm. After the training is completed, the validation samples are used to test the model, and the classification accuracy of the model on the validation samples is calculated as the fitness value.

[0047] Step 7.3: Update the position and velocity of the particle according to the current position and velocity of the particle, as well as the information of the global optimal solution and the individual optimal solution.

[0048] The velocity update formula is

[0049]

[0050] The position update formula is

[0051]

[0052] where and are respectively the velocity and position of the $i$-th particle in the $d$-th dimension at the $k$-th iteration; $\omega$ is the inertia weight, set to 0.8; $c_1$ and $c_2$ are learning factors, both 1.5; $r$ 1id and $r$ 2id are random numbers between [0, 1]; $p$ id is the individual optimal position of the $i$-th particle, and $g$ d is the global optimal position of the entire particle swarm.

[0053] Step 7.4: Set the maximum number of iterations to 50. When the number of iterations reaches 50 or the fitness value no longer improves in consecutive iterations, stop the iteration.

[0054] Step 7.5: Take the weights and thresholds corresponding to the global optimal solution as the weights and thresholds of the optimized DBN model to obtain the PSO - DBN model.

[0055] Furthermore, in step 8, input the test set into the trained PSO - DBN model. The model calculates and judges according to the learned feature patterns and outputs the diagnostic result. If the output result is 0, it means the synchronous condenser is in normal operation; if the output result is 1, 2, or 3, it means the synchronous condenser is in a slight, moderate, or severe static eccentricity fault state respectively.

[0056] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0057] 1. Through multi - source information fusion, the present invention comprehensively utilizes the characteristics of various signals such as electromagnetic torque and rotor current, can more comprehensively reflect the operating state of the synchronous condenser, overcomes the defect of insufficient diagnostic information of a single signal, and significantly improves the accuracy and reliability of fault diagnosis.

[0058] 2. The present invention adopts a feature extraction method combining EEMD and permutation entropy, can effectively extract the non - linear features in the signal, better capture the characteristic information of the static eccentricity fault of the synchronous condenser, and provides strong support for accurate diagnosis.

[0059] 3. The present invention optimizes the weights and thresholds of the DBN using the PSO algorithm, avoiding the problem of being easily trapped in local optima caused by randomly initialized weights in the traditional DBN, improving the diagnostic accuracy and generalization ability of the model, and enabling it to adapt to the fault diagnosis of synchronous condensers under different working conditions.

[0060] 4. The fault diagnosis method and system of the present invention have good generality and scalability, and are not only applicable to the static eccentricity fault diagnosis of synchronous condensers, but also can provide reference for the fault diagnosis of other types of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings.

[0062] Figure 1 It is a flowchart of the static eccentricity fault diagnosis method for synchronous condensers of the present invention.

[0063] Figure 2 It is an architecture diagram of the fault diagnosis model based on PSO-DBN of the present invention.

[0064] Figure 3 It is a finite element model of the synchronous condenser established.

[0065] Figure 4 It is a mesh segmentation diagram of the finite element model of the synchronous condenser established.

[0066] Figure 5 It is a finite element coupled stator external circuit model.

[0067] Figure 6 It is a finite element coupled rotor external circuit model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] In order to make the objectives, technical solutions and advantages of the present application more clear, the following will further describe the present application in detail with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0069] Embodiment 1:

[0070] In this embodiment, in step 1, a finite element simulation model of the synchronous condenser is established using a professional electromagnetic simulation software, such as ANSYS Maxwell. In the model, various parameters of the synchronous condenser are accurately set, including the geometric dimensions, material properties, number of winding turns and connection methods of the stator and rotor.

[0071] To simulate different degrees of static eccentricity faults, it is achieved by adjusting the eccentricity distance between the rotor and the stator in the model. Under normal conditions, the rotor is concentric with the stator; in the case of a slight static eccentricity fault, a small eccentricity distance is set, such as 1% of the stator radius; in the case of a moderate static eccentricity fault, the eccentricity distance is set to 3% of the stator radius; in the case of a severe static eccentricity fault, the eccentricity distance is set to 5% of the stator radius.

[0072] Run the simulation model. Under different fault conditions, collect multi-source signals at a sufficiently high sampling frequency (such as 10 kHz) to ensure that the subtle changes in the signals can be captured. The duration of each collected signal is 10 s to obtain stable signal characteristics.

[0073] In step 2, for the electromagnetic torque signal, wavelet threshold denoising is performed using the wavelet analysis toolbox in MATLAB. First, import the collected electromagnetic torque signal into the MATLAB environment. Then, select the db4 wavelet basis function because the db4 wavelet has good time-frequency localization characteristics and is suitable for processing the mutation information in the electromagnetic torque signal. Set the decomposition level to 5 layers and perform wavelet decomposition on the electromagnetic torque signal to obtain wavelet coefficients at different frequency scales. According to the noise characteristics of the signal, select an appropriate threshold method, such as the soft threshold method, to perform threshold processing on the wavelet coefficients and remove the wavelet coefficients caused by noise. Finally, reconstruct the denoised electromagnetic torque signal using the processed wavelet coefficients.

[0074] For the current signal, write programs for mean filtering and median filtering using Python. First, process the collected current signal using the mean filtering algorithm. By setting an appropriate window size (such as 5 sampling points), calculate the average value of the signal within the window and replace the sampling point value at the center of the window with this average value to remove the DC bias and low-frequency noise. Then, use the median filtering algorithm. Similarly, set an appropriate window size (such as 3 sampling points), sort the sampling point values within the window, and replace the sampling point value at the center of the window with the median value to remove the impulse noise and obtain a smooth current signal.

[0075] In Python, for the denoised electromagnetic torque signal and current signal, the minimum-maximum normalization method is used for processing. By calculating the minimum and maximum values of the signal, map the amplitude of the signal to the interval [0, 1] according to the normalization formula to ensure comparability between different signals.

[0076] In steps 3 and 4, the FFT calculation is performed using the numpy and scipy libraries in Python. The denoised electromagnetic torque signal is subjected to FFT transformation to obtain the spectrum of the signal. According to the characteristics of the electromagnetic torque during the static eccentricity fault of the synchronous condenser, the second harmonic component is extracted from the spectrum. The frequency of this component is twice the power supply frequency. For example, if the power supply frequency is 50 Hz, the harmonic component at 100 Hz is extracted as the pre-feature of the electromagnetic torque signal. Similarly, the FFT transformation is performed on the denoised stator current and rotor current signals using the numpy and scipy libraries. According to the influence of the static eccentricity fault on the stator current harmonics and rotor current, the 2nd (100 Hz), 3rd (150 Hz), 5th (250 Hz), and 7th (350 Hz) harmonic components are extracted from the stator current spectrum as pre-features. The 5th (250 Hz), 7th (350 Hz) harmonic components and the DC component (frequency of 0 Hz) are extracted from the rotor current spectrum as pre-features. By extracting the amplitudes at the corresponding frequencies in the spectrum, the corresponding pre-feature values are obtained.

[0077] In steps 5 and 6, the EEMD decomposition of the multi-source signals is performed using the PyEMD library in Python. The denoised and normalized electromagnetic torque signal, stator current signal, and rotor current signal are respectively input into the EEMD algorithm. During the operation of EEMD, the amplitude of the added white noise is set to 0.2 times the standard deviation of the signal, and the decomposition times are set to 100 times to ensure that the signal can be fully decomposed. Through multiple decompositions and averaging processes, several IMF components are obtained.

[0078] For each IMF component, a permutation entropy calculation program is written in Python. First, phase space reconstruction is performed, and the embedding dimension and time delay are selected to obtain the phase space vector. For each vector, its elements are arranged in ascending order to determine the permutation pattern. The number of occurrences of each permutation pattern is counted, the probability of the permutation pattern occurrence is calculated, and the permutation entropy of each IMF component is calculated according to the permutation entropy formula.

[0079] The permutation entropy feature vectors of the second harmonic component of the electromagnetic torque signal, the 2nd, 3rd, 5th, and 7th harmonic components of the stator current signal, and the 5th, 7th, and DC components of the rotor current are sequentially connected using the numpy library in Python to form a high-dimensional feature vector. The constructed fault sample set is divided into a training set, a validation set, and a test set according to the ratio of 7:2:1, and the class labels of different fault states are marked respectively. The normal state is marked as 0, the slight static eccentricity fault is marked as 1, the moderate static eccentricity fault is marked as 2, and the severe static eccentricity fault is marked as 3.

[0080] In step 7, a DBN model with two layers of RBM is constructed using the TensorFlow library in Python. The number of visible layer nodes of the first layer of RBM is determined according to the dimension of the input feature vector, and the number of hidden layer nodes is set to 20. The number of visible layer nodes of the second layer of RBM is the number of hidden layer nodes of the first layer of RBM, and the number of hidden layer nodes is set to 15. A Softmax classification layer is added on top of the DBN model for outputting the fault diagnosis results, and the number of output layer nodes is 4, corresponding to four fault states.

[0081] In step 8, a PSO algorithm program is written in Python. 30 particles are randomly generated, and the position of each particle is randomly valued within a certain range, and this position represents a set of weights and thresholds of the DBN model. The velocity of the particles is initialized to zero. The weights and thresholds corresponding to each particle are substituted into the DBN model, and the model is trained using the training samples. During the training process, the cross-entropy loss function is used to calculate the error between the model prediction value and the true label, and the model parameters are updated through the backpropagation algorithm. After the training is completed, the model is tested using the validation samples, and the classification accuracy of the model on the validation samples is calculated as the fitness value. The particles are iteratively updated according to the velocity and position update formulas, and the maximum number of iterations is set to 50. When the number of iterations reaches 50 or the fitness value no longer improves in consecutive multiple iterations, the iteration stops. The weights and thresholds corresponding to the global optimal solution are used as the weights and thresholds of the optimized DBN model to obtain the PSO-DBN model.

[0082] In step 9, the test set is input into the trained PSO-DBN model, and the model calculates and judges according to the learned feature patterns. In Python, the trained model is called to predict the test set and the diagnostic results are output. If the output result is 0, it means that the synchronous condenser is in normal operation; if the output result is 1, 2, or 3, it means that the synchronous condenser is in a slight, moderate, or severe static eccentricity fault state respectively. According to the diagnostic results, corresponding maintenance measures are taken in a timely manner to ensure the safe and stable operation of the synchronous condenser.

[0083] Only some exemplary embodiments of the present invention are described above by way of illustration. Undoubtedly, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network, characterized in that, The steps include: S1: Establish a finite element simulation model to collect multi-source signals of the condenser under different working conditions; S2: performing preprocessing operations including denoising and normalization on the collected multi-source signals; S3: Decompose multi-source signals using fast Fourier transform to obtain pre-features; S4: Extract features from pre-features using ensemble empirical mode decomposition and permutation entropy; S5: Fusing the permutation entropy feature parameters extracted from the electromagnetic torque, stator current and rotor current signals to construct a fault sample set; S6: Build a deep belief network model; S7: Particle swarm optimization algorithm is used to optimize the weights and thresholds of the deep belief network model; S8: Diagnosis of static eccentricity fault of phase regulator.

2. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein The multi-source signals collected in step S1 include electromagnetic torque, stator current and rotor current signals in the normal state and at different fault levels of the phase regulator.

3. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein, The pre-processing operation in step S2 comprises the following steps: S21: De-noising the collected multi-source signals; for the electromagnetic torque signal, adopt the wavelet threshold denoising method; select the wavelet basis function and the number of decomposition layers; perform threshold processing on the wavelet coefficients to remove the wavelet coefficients caused by noise, and reconstruct the denoised electromagnetic torque signal; for the current signal, adopt a method combining mean filtering and median filtering; First, use mean filtering to remove DC bias and low-frequency noise, and then use median filtering to remove pulse noise to obtain a smooth current signal; S22: Normalize the multi-source information collected: normalize the denoised electromagnetic torque and current signals; use the min-max normalization method to map the amplitude of the signals to the interval [0, 1]; for the electromagnetic torque signal X T , the normalization formula is where X Tmin and X Tmax are the minimum and maximum values of the electromagnetic torque signal respectively, and Y v is the normalized electromagnetic torque signal; similarly, similar normalization processing is performed on the current signal X i to obtain the normalized current signal Y i .

4. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein In step S3, fast Fourier transform is used to decompose the multi-source signals of electromagnetic torque, stator current, and rotor current, and the double frequency component of the electromagnetic torque signal, the 2nd, 3rd, 5th, and 7th harmonic components of the stator current, and the 5th, 7th, and DC components of the rotor current are extracted as multi-source signals.

5. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 4, wherein The step S3 comprises the following steps: S31: The electromagnetic torque of the condenser in normal state is expressed as where: δ s is the relative static deflection angle; t is the time; Comparing the phase regulator under normal state and static eccentricity fault state, the phase regulator electromagnetic torque has a second harmonic component in addition to the DC component; the static eccentricity of the air gap will make the steady-state electromagnetic torque of the generator present a fluctuation characteristic of twice the fundamental frequency. As the static eccentricity of the air gap increases, the fluctuation degree of the second harmonic frequency will also increase; the second harmonic component of the electromagnetic torque signal is used as the pre-characteristic of the electromagnetic torque signal; S32: The static eccentricity of the phase regulator will destroy the symmetry of the air gap magnetic field of the motor, causing the magnetic field distribution inside the motor to change, generating a pulsating magnetomotive force that changes at twice the power supply frequency, and inducing a current of twice the frequency in the stator winding; static eccentricity will destroy the three-phase symmetry of the magnetic field, affecting the balance relationship of the three-phase current, and the third harmonic will appear in the stator current; static eccentricity will cause nonlinear phenomena such as local saturation of the magnetic field, and the fifth and seventh harmonics will appear in the stator current; the second, third, fifth and seventh harmonic components of the stator current signal are used as the pre-features of the stator current signal; The stator current signal is expressed as: Where: p represents the number of pole pairs of the motor, δ represents the initial rotor phase angle, ω c represents the fault characteristic frequency of the bearing, ω s represents the power supply frequency of the motor; The static eccentricity of the synchronous condenser will break the symmetry of the rotor magnetic field, causing a DC component to appear in the induced electromotive force of the rotor winding, thereby generating a DC component in the rotor current; similar to the stator current, the rotor current may also have higher harmonics such as the 5th and 7th harmonics; the 5th and 7th harmonic components and the DC component of the rotor current signal are used as the pre-features of the rotor current signal; The current of phase A of the motor is expressed as: Where: θ0 is the initial position of the rotor, f e0 is the fundamental frequency of the motor current, q tz is the modulation component corresponding to the angular velocity fluctuation, i q0 represents the average value of the q-axis current, i qi represents the amplitude of the component with a fluctuation frequency of f i , and is its corresponding phase.

6. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein The feature extraction in step S4 includes the following steps: S41: Input the pre-features collected from the electromagnetic torque and current signals into the ensemble empirical mode decomposition algorithm. When the ensemble empirical mode decomposition algorithm runs, white noise with different amplitudes is added to the original signal, and these white noises make the signal show more obvious feature differences at different scales; through multiple decompositions and averaging processes, several intrinsic mode function components are finally obtained; S42: Calculate the permutation entropy for each intrinsic mode function component; Taking a time series of an intrinsic mode function component as an example, first perform phase space reconstruction, select appropriate embedding dimension m and time delay τ, and obtain the phase space vector: X i = [x(i), x(i), …, x(i+(m-1)τ)], i = 1, 2, …, N-(m-1)τ For each vector X i Arrange its elements in ascending order to obtain: x(i1) ≤ x(i2) ≤ … ≤ x(i m ) where i1, i2... i m is a permutation of 1, 2... m, and this permutation is represented by π = (i1, i2... i m ), which is composed of Calculate the permutation entropy, where p(π) is the probability of the permutation pattern appearing.

7. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein In step S5, the permutation entropy feature parameters extracted from the electromagnetic torque, stator current, and rotor current are fused; in a cascaded manner, the permutation entropy feature vectors of the second harmonic component of the electromagnetic torque signal, the 2nd, 3rd, 5th, and 7th harmonic components of the stator current signal, and the permutation entropy feature vectors of the 5th, 7th, and DC components of the rotor current are connected in sequence to form a high-dimensional feature vector; use the fused feature vector to construct a fault sample set; divide the constructed fault sample set into a training set, a validation set, and a test set according to a ratio of 7:2:1; mark the feature vectors in different fault states in the fault sample set with corresponding category labels, mark the normal state as 0, the slight static eccentricity fault as 1, the moderate static eccentricity fault as 2, and the severe static eccentricity fault as 3.

8. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein In step S6, a deep belief network model is constructed, which is composed of multiple restricted Boltzmann machines stacked; construct a deep belief network model containing two layers of restricted Boltzmann machines. The number of visible layer nodes of the first layer of restricted Boltzmann machines is determined according to the dimension of the input feature vector, and the number of hidden layer nodes is set to 20; the number of visible layer nodes of the second layer of restricted Boltzmann machines is the number of hidden layer nodes of the first layer of restricted Boltzmann machines, and the number of hidden layer nodes is set to 15; finally, add a Softmax classification layer on the top of the deep belief network model to output the fault diagnosis result, and the number of output layer nodes is 4, corresponding to four fault states.

9. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein The optimization of step S7 specifically includes the following steps: S71: Initially randomly generate 30 particles, and the position of each particle is randomly taken within a certain range. This position represents a set of weights and thresholds of the deep belief network model; the position of each particle contains the specific values of these weights and thresholds; the velocity of the particle is initialized to zero; S72: Substitute the weights and thresholds corresponding to each particle into the deep belief network model, and use the training samples to train the model; During the training process, the cross-entropy loss function is used to calculate the error between the model prediction value and the true label, and the model parameters are updated through the backpropagation algorithm; after the training is completed, the validation samples are used to test the model, and the classification accuracy of the model on the validation samples is calculated as the fitness value; S73: Update the position and velocity of the particle according to the current position and velocity of the particle, as well as the information of the global optimal solution and the individual optimal solution; The velocity update formula is: The position update formula is: Among them and are the velocity and position of the $i$-th particle in the $d$-th dimension at the $k$-th iteration respectively; $\omega$ is the inertia weight, set to 0.8; $c_1$ and $c_2$ are learning factors, both being 1.5; $r$ 1id and $r$ 2id are random numbers between [0, 1]; $p$ id is the individual best position of the $i$-th particle, and $g$ d is the global best position of the entire particle swarm; S74: Set the maximum number of iterations to 50. When the number of iterations reaches 50 or the fitness value no longer improves in consecutive multiple iterations, stop the iteration; S75: Use the weights and thresholds corresponding to the global optimal solution as the weights and thresholds of the optimized deep belief network model to obtain the deep belief network model optimized by the particle swarm optimization algorithm.

10. The method for diagnosing the static eccentricity fault of a synchronous condenser based on an improved deep belief network according to claim 1, wherein In step S8, the test set is input into the deep belief network model optimized by the trained particle swarm optimization algorithm. The model calculates and judges according to the learned feature patterns and outputs the diagnostic results. If the output result is 0, it means that the synchronous condenser is in normal operation; if the output result is 1, 2, or 3, it means that the synchronous condenser is in a slight, moderate, or severe static eccentricity fault state, respectively.

Images