Power plant start-up transformer open-phase detection method and device

Abstract

The application provides a power plant starting transformer open-phase detection method and device, and relates to the technical field of power failure detection. The method comprises the following steps: acquiring signal energy distribution of the starting transformer, constructing an energy feature vector reflecting the relative energy relationship among three-phase current signals and inputting the energy feature vector into an open-phase identification model to obtain initial probabilities of each-phase open-phase failure; combining real-time calculated zero sequence and negative sequence current amplitudes to perform state inversion calculation, inferring the change amount of equivalent circuit parameters of transformer internal windings and lead wires before and after the failure, correcting the initial probabilities to obtain confidence, and further generating an open-phase detection report containing fault phase and fault severity. The application completely depicts the three-phase current energy correlation characteristics through the energy feature vector, corrects the initial probabilities of the model by using the change amount of the equivalent circuit parameters, eliminates the random deviation of the single model output, realizes accurate matching between the fault determination result and the actual fault state of the transformer, and improves the fault phase identification accuracy.

Description

Technical Field

[0001] This application relates to the field of power fault detection technology, and in particular to a method and device for detecting phase loss in a power plant starting transformer. Background Technology

[0002] As a key piece of equipment in the power plant's auxiliary power system, the starting transformer's operational reliability directly affects the safe start-up and shutdown of the unit. Phase loss faults are a common and serious asymmetrical operating condition of starting transformers. If not detected in a timely and accurate manner, it may lead to transformer overheating and burnout, causing a power outage.

[0003] In related technologies, phase loss detection of power plant starting transformers commonly uses electromagnetic current transformers to collect phase current signals as discriminative features, and outputs fault judgment results through fixed threshold comparison or shallow classification models. However, electromagnetic current transformers are susceptible to interference in the strong electromagnetic environment of power plants, making it difficult to guarantee signal acquisition accuracy; the discriminative feature dimension based on single-phase amplitude or sequence component amplitude is too narrow, ignoring the energy correlation characteristics between the three-phase currents; the preliminary fault probability directly output by the shallow model has obvious random bias; the accuracy of fault phase identification is insufficient; and the assessment of fault severity is relatively coarse, making it difficult to meet the operational requirements of refined detection and rapid switching of starting transformers. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, the first aspect of this application proposes a method for detecting phase loss in a power plant startup transformer, comprising:

[0006] The current signals of each phase of the starting transformer are collected by an optical current transformer; Based on the correlation coefficient between the signal energy distribution curves of any two phase current signals, an energy feature vector reflecting the relative energy relationship between the three-phase current signals is constructed. The energy feature vector is input into a pre-trained phase failure identification model to obtain the initial probability of phase failure in each phase. The phase failure identification model is constructed using a support vector machine. An equivalent circuit model is constructed to describe the starting transformer under a phase loss fault. The zero-sequence current value and the negative-sequence current value are calculated from the current signal as observations. The impedance of each phase winding in the equivalent circuit model is used as the state variable to be inverted to perform state inversion calculation, and the relative change of the impedance of each phase winding before and after the phase loss fault occurs is obtained. The initial probability is corrected using the relative change to obtain the confidence level of each phase failure. Based on the confidence level and the preset fault threshold, determine whether each phase of the starting transformer has experienced a phase loss fault, and generate a phase loss detection report.

[0007] In some embodiments of this application, the energy feature vector is represented as follows:

[0008] in, The energy feature vector, , , The Pearson correlation coefficients are the signal energy distribution curves of phases AB, AC, and BC, respectively. , , These represent the average energy ratios of phases A, B, and C during the current time period.

[0009] In some embodiments of this application, the step of using the zero-sequence current value and negative-sequence current value obtained from the current signal as observations, and using the impedance of each phase winding in the equivalent circuit model as the state variable to be inverted to perform state inversion calculation, to obtain the relative change in the impedance of each phase winding before and after the phase loss fault, includes: using the zero-sequence current value and negative-sequence current value obtained from the current signal as observations, and using the impedance of each phase winding in the equivalent circuit model as the state variable to be inverted, constructing a state estimation optimization model with the objective of minimizing the observation residual, wherein the observation residual is the difference between the estimated value of the zero-sequence current amplitude and the zero-sequence current value. The difference, and the sum of squares of the differences between the estimated negative sequence current amplitude and the negative sequence current value; the estimated zero sequence current amplitude and the estimated negative sequence current amplitude are calculated based on the currently estimated impedance of each phase winding; the state estimation optimization model is solved using the gradient descent method, and the impedance values ​​of each phase winding in the equivalent circuit model are iteratively adjusted; when the observation residual is less than the preset convergence threshold, the state inversion process is determined to be converged, and the final estimated value of the impedance of each phase winding is determined; based on the final estimated value of the impedance of each phase winding and the standard impedance reference value, the relative change of the impedance of each phase winding before and after the phase failure is obtained.

[0010] In some embodiments of this application, the step of using the relative change to correct the initial probability and obtain the confidence level of each phase failure includes: comparing the relative change with a preset impedance change threshold to generate a hard decision indication; using a Bayesian update rule, taking the initial probability as a prior probability and the hard decision indication as new evidence, calculating the posterior probability of each phase failure under the hard decision indication evidence, and taking the posterior probability as the confidence level of each phase failure.

[0011] In some embodiments of this application, the Bayesian update rule is expressed as follows:

[0012] in, This indicates that evidence indicating a hard judgment was observed. Under the condition that, assuming The posterior probability of its validity. This indicates that a phase loss fault has occurred in a certain phase. Indicates a hypothesis The prior probability of its validity. Indicates in the assumption Under the conditions that it is established, hard judgment directive evidence is observed. The probability of.

[0013] In some embodiments of this application, determining whether a phase loss fault has occurred in each phase of the starting transformer based on the confidence level and a preset fault threshold includes: comparing the confidence level with the fault threshold, wherein the fault threshold includes an alarm threshold and a trip threshold, and the alarm threshold is less than the trip threshold; if the confidence level is greater than or equal to the alarm threshold and the confidence level is less than the trip threshold, the corresponding phase is determined to be a minor phase loss fault; if the confidence level is greater than or equal to the trip threshold, the corresponding phase is determined to be a severe phase loss fault.

[0014] In some embodiments of this application, the phase failure detection report further includes the phase failure occurrence time; the method further includes: in the timestamp of the current signal, locating the time point at which the energy of the current signal under the phase that has experienced a phase failure first undergoes an abnormal change, and taking it as the corresponding phase failure occurrence time.

[0015] In some embodiments of this application, the method further includes: acquiring system operation status records during the process of the starting transformer performing the unit startup task; verifying the phase failure detection results in the phase failure detection report through the system operation status records; if the verification is successful, using the energy feature vector and the relative changes in the impedance of each phase winding before and after the phase failure as new training samples to incrementally train the phase failure identification model and update the internal parameters of the phase failure identification model.

[0016] A second aspect of this application provides a phase loss detection device for a power plant starting transformer, comprising: The acquisition module is used to acquire the current signals of each phase of the starting transformer through an optical current transformer; The processing module is used to construct an energy feature vector that reflects the relative energy relationship between the three-phase current signals based on the correlation coefficient between the signal energy distribution curves of any two phases of the current signals. The first identification module is used to input the energy feature vector into a pre-trained phase failure identification model to obtain the initial probability of phase failure in each phase. The phase failure identification model is constructed using a support vector machine. The determination module is used to construct an equivalent circuit model describing the starting transformer under a phase loss fault, to obtain the zero-sequence current value and negative-sequence current value as observations from the current signal, and to perform state inversion calculation using the impedance of each phase winding in the equivalent circuit model as the state quantity to be inverted, so as to obtain the relative change in the impedance of each phase winding before and after the phase loss fault occurs. The second identification module is used to correct the initial probability using the relative change amount to obtain the confidence level of each phase failure. The monitoring module is used to determine whether a phase failure has occurred in each phase of the starting transformer based on the confidence level and the preset fault threshold, and to generate a phase failure detection report.

[0017] A third aspect of this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method described in the first aspect above.

[0018] The phase loss detection method for power plant starting transformers provided in this application constructs an energy feature vector that reflects the relative energy relationship between the three-phase current signals. This overcomes the limitations of traditional detection methods that rely solely on the feature representation of single-phase current amplitude or sequence component amplitude. It can fully characterize the correlation characteristics and coordinated change law of the three-phase current at the energy level, providing fault feature inputs with higher discrimination and stronger anti-interference ability for the phase loss identification model, thereby reducing the distortion of the initial fault probability. Based on this, the zero-sequence current amplitude and negative-sequence current amplitude obtained by real-time calculation are combined to perform state inversion calculation, inferring the changes in the equivalent circuit parameters of the transformer's internal windings and leads before and after the occurrence of a phase-loss fault. This change is used to correct the initial probability output by the phase-loss identification model, and a dual-channel cross-validation mechanism integrating data-driven discrimination (phase-loss identification model) and physical mechanism verification (equivalent circuit model state inversion) is constructed. This effectively eliminates the random bias in the output results of a single model, enabling the fault confidence to accurately match the actual fault state of the transformer. It realizes the leap from simple qualitative judgment to multi-dimensional refined evaluation of phase-loss fault detection, providing sufficient information support for operation and maintenance personnel to quickly locate and handle faults, and effectively meeting the strict requirements of power plant start-up transformers for the accuracy and integrity of phase-loss detection.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A schematic flowchart illustrating a phase loss detection method for a power plant starting transformer provided in this application embodiment; Figure 2 This application provides a diagram showing the change in zero-sequence / negative-sequence current amplitude before and after a transformer phase-loss fault. Figure 3 This application provides a diagram showing the energy variation of three-phase signals of a transformer under different operating conditions. Figure 4 This application provides a diagram showing the energy variation of three-phase current signals under a phase-A failure. Figure 5 This is a schematic diagram of a phase loss detection device for a power plant starting transformer, provided in an embodiment of this application. Detailed Implementation

[0021] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0022] Specifically, the following describes the power plant starting transformer phase loss detection method and apparatus according to embodiments of this application with reference to the accompanying drawings.

[0023] Figure 1 This is a schematic flowchart illustrating a method for detecting phase loss in a power plant starting transformer, provided as an embodiment of this application. Figure 1 As shown, the method for detecting phase loss in the power plant's starting transformer may include the following steps: Step 101: Acquire the current signals of each phase of the starting transformer using an optical current transformer.

[0024] In some embodiments of this application, the starting transformer can be a starting step-down transformer for the unit's power supply system. Optical current transformers are installed at preset positions at the bushing roots of phases A, B, and C of the starting transformer. The optical current transformers are used to synchronously collect the current signals flowing through the grounding leads of each phase bushing.

[0025] Step 102: Based on the correlation coefficient between the signal energy distribution curves of any two phase current signals, construct an energy feature vector that reflects the relative energy relationship between the three-phase current signals.

[0026] In some embodiments of this application, the acquired raw current signal can be pre-conditioned. Pre-conditioning includes filtering the signal to suppress high-frequency interference and amplifying the signal to match the range of the subsequent analog-to-digital converter (ADC). In one implementation, a low-pass filter with a cutoff frequency of 2kHz can be used to filter out high-frequency noise, and then a programmable gain amplifier can be used to adjust the signal amplitude to a range of ±5V to match the input range of the 16-bit ADC. The pre-conditioned current signal is converted into a digital current sequence with a sampling frequency of 10kHz. The digital current sequence includes timestamp information and is sampled synchronously with the power grid frequency. A short-time Fourier transform is performed on the digital current sequence to extract the signal energy distribution within a characteristic frequency band centered on the power frequency fundamental wave.

[0027] The process of performing a short-time Fourier transform on a digital current sequence may include the following steps: windowing the digital current sequence of length N, using a sliding time window to extract a fixed-length segment of the current signal, with a sliding window length L of 2000 sampling points and a sliding step size of 100 sampling points; applying a Hanning window function to each current signal segment to reduce spectral leakage; performing a fast Fourier transform on the windowed current signal segment to obtain the complex spectrum of the current signal segment; locating the spectral line corresponding to the fundamental frequency of 50Hz and the spectral lines corresponding to its adjacent preset number of harmonic frequencies from the complex spectrum, with a preset number of 5, i.e., calculating the spectral energy corresponding to the 50Hz fundamental, 150Hz third harmonic, 250Hz fifth harmonic, 350Hz seventh harmonic, and 450Hz ninth harmonic; calculating the sum of the squares of the amplitudes at the fundamental frequency spectral line and each harmonic frequency spectral line as the signal energy value within the characteristic frequency band, continuously performing this calculation on the sliding window to form a time-varying signal energy distribution.

[0028] For example, the energy value of a phase current signal calculated within a sliding window. This can be expressed by the following formula:

[0029] in, This represents the signal energy value calculated within a certain sliding window. This indicates that in the complex spectrum obtained by the Fast Fourier Transform, the th... The complex amplitude corresponding to each spectral line, subscript and These correspond to the spectral indices of the starting and ending harmonic orders within the characteristic frequency band, respectively. Taking phase A as an example, for a normal phase A current signal segment, the calculated... The value stabilizes within a reference range; when a phase A failure occurs, the phase A current signal is distorted or disappears, and the value calculated within the corresponding sliding window is... The values ​​will be significantly lower than the reference range, while the signal energy values ​​of phases B and C will be significantly lower. and This may manifest as fluctuations or increases of varying degrees. By continuously comparing the morphological and numerical differences in the energy distribution curves of the three-phase signals, the original time-frequency characteristics are provided for subsequent analysis.

[0030] Optionally, synchronous sampling can be implemented using a phase-locked loop circuit to ensure that the sampling clock of the analog-to-digital converter is strictly synchronized with the phase of the grid voltage, thereby ensuring that the sampling times of the three-phase digital current sequence are strictly aligned, making the signal energy distribution calculated subsequently comparable. The digital current sequence, along with a high-precision timestamp, is cached in a first-in-first-out memory for the signal energy analysis module to read. The method of performing a short-time Fourier transform on the digital current sequence and extracting the signal energy distribution within the characteristic frequency band can convert the time-domain current waveform into a time-frequency domain energy sequence. The signal energy distribution intuitively reflects the intensity changes of each phase current at the power frequency and its main harmonic components. The normal operation state and the phase loss fault state of the three-phase current signal will exhibit distinguishable differences in the signal energy distribution.

[0031] In some embodiments of this application, the signal energy distribution curves of the three-phase current signals can be obtained separately within the same time period. The time period can be selected as an interval containing 1024 consecutive sliding windows, corresponding to a current data segment lasting 10.24 seconds. The signal energy distribution curve of each phase current is normalized to eliminate absolute energy deviations caused by differences in the installation position or sensitivity of the current transformers. The normalization process uses the long-term average energy value of each phase obtained statistically under historical normal operating conditions as a benchmark for per-unit calculation. After processing, the energy distribution curve values ​​of the three phases are comparable on the same order of magnitude.

[0032] In one implementation, a 3×3 correlation matrix reflecting the synchronicity of energy changes among the three phases can be obtained based on the correlation coefficient between the signal energy distribution curves of any two phase current signals. The upper triangular elements of the correlation matrix are extracted, and the average energy ratio of each of the three phase currents is added to construct an energy feature vector reflecting the relative energy relationship between the three phase current signals.

[0033] As an example, the energy eigenvector can be represented as follows:

[0034] in, For energy eigenvectors, , , These are the Pearson correlation coefficients between the normalized signal energy distribution curves of phases AB, AC, and BC, respectively. , , These are the average energy ratios of phases A, B, and C during the current time period, which are the arithmetic mean of all values ​​on the normalized energy distribution curve. They reflect the relative level of the average intensity of the phase current signal during the observation period compared to its own historical benchmark. , , The three components quantify the synchronicity and correlation of the three-phase current signal energy in the time domain. Under normal operating conditions, the three-phase currents are in balance, and their energy curves show a highly consistent trend, so the three correlation coefficients are all close to 1. When a single-phase failure occurs, the energy curve shape of the faulty phase will differ significantly from that of the non-faulty phase, causing a sharp drop in the correlation coefficient between the faulty and non-faulty phases, thus directly and sensitively revealing the disruption of the interphase energy correlation. , , The three components characterize the average relative intensity level of each phase current signal during the observation period. Under normal operating conditions, the R values ​​of the three phases are all close to 1 and are relatively equal; when a single phase failure occurs, the R value of the faulty phase will be significantly lower than 1, and the R value of the non-faulty phase may also undergo an observable shift due to system asymmetry. When constructing the energy feature vector, the three upper triangular off-diagonal elements extracted from the correlation matrix and the three calculated average energy ratio values ​​are concatenated in a predetermined order to form a six-dimensional real vector. This vector fully encodes the relative relationship and absolute ratio information of the three-phase current signals at the energy level.

[0035] For example, within a certain time period, assuming the correlation coefficient between the energy curves of phase A and phase B is 0.95, the correlation coefficient between the energy curves of phase A and phase C is 0.92, the correlation coefficient between the energy curves of phase B and phase C is 0.94, and the average energy ratio of phase A is 1.05, the average energy ratio of phase B is 0.98, and the average energy ratio of phase C is 0.97, then the constructed energy feature vector is represented as [0.95, 0.92, 0.94, 1.05, 0.98, 0.97].

[0036] Therefore, the energy feature vector comprehensively and completely characterizes the relative energy relationship between the three-phase current signals from the perspectives of the correlation of the changing trends, the relativity of the first three components and the average intensity, and the two complementary dimensions of the last three components. Compared with traditional features such as single-phase amplitude or sequence components, the energy feature vector contains richer information on the coordinated operation of the three phases, reduces the impact of electromagnetic interference on the effectiveness of the features, and provides a fault feature input with higher discriminative power and stronger anti-interference capability for subsequent phase loss identification models.

[0037] As an example, the correlation coefficient can be calculated using the following formula:

[0038] in, This represents the Pearson correlation coefficient between the normalized energy sequences of phases X and Y. This indicates the total number of sliding windows within the selected time period. and They represent the first The normalized signal energy values ​​calculated for the X and Y phases within a sliding window. and They represent phase X and phase Y respectively throughout the entire time period. The normalized signal energy average value within each window. The calculated... , , The three values ​​constitute the off-diagonal elements of the upper triangular region of the correlation matrix. Under normal operating conditions, the three-phase currents are balanced, and the calculated correlation coefficient is close to 1, indicating a high correlation in the energy feature vector. When a single-phase failure occurs, the energy change patterns of the fault phase lose synchronization with the energy changes of the other two phases, and the correlation coefficient corresponding to the fault phase drops significantly. The change in the pattern of the energy feature vector can be captured by the subsequent phase failure identification model.

[0039] Step 103: Input the energy feature vector into the pre-trained phase failure identification model to obtain the initial probability of phase failure for each phase. The phase failure identification model is constructed using a support vector machine.

[0040] In some embodiments of this application, the training samples of the phase failure identification model contain a large number of energy feature vectors under normal operating conditions and various preset phase failure fault conditions. The training samples can be derived from historical data records or generated by simulation models, covering various load levels and initial fault angles. During the training phase, the phase failure identification model can use optimization algorithms to find an optimal classification hyperplane or nonlinear decision boundary to distinguish the energy feature vectors corresponding to "normal" and various "phase failure" states. In the application phase, the phase failure identification model performs nonlinear mapping and classification decisions on the input energy feature vectors. The nonlinear mapping can be implemented using kernel functions to map the input low-dimensional energy feature vectors to a high-dimensional feature space to improve classification performance. The output of the phase failure identification model is a probability vector containing three elements, where each element corresponds to the initial probability of a phase failure. The sum of the three probability values ​​does not have to be 1. Each probability value independently represents the probability of a corresponding phase failure. For example, the model may output a probability vector of [0.85, 0.12, 0.08], which means that the model judges the initial probability of a phase failure in phase A to be 0.85, phase B to be 0.12, and phase C to be 0.08.

[0041] Step 104: Construct an equivalent circuit model to describe the starting transformer under a phase loss fault. The zero-sequence current value and the negative-sequence current value are calculated from the current signal as observations. The state inversion calculation is performed using the impedance of each phase winding in the equivalent circuit model as the state quantity to be inverted, and the relative change of the impedance of each phase winding before and after the phase loss fault occurs is obtained.

[0042] In one implementation, a symmetrical component method transformation can be performed on the three-phase digital current sequence to calculate the instantaneous amplitudes of the zero-sequence current component and the negative-sequence current component in real time, thus obtaining the zero-sequence current value and the negative-sequence current value. The calculation process is performed in units of sliding time windows, synchronized with the signal energy feature extraction window.

[0043] In some embodiments of this application, an equivalent circuit model describing the starting transformer under a phase-loss fault can be pre-established. The equivalent circuit model equates the phase-loss fault to a step increase in the impedance of a specific phase winding. The model includes the self-impedance of the three-phase windings and the mutual impedance parameters between each pair of windings. The zero-sequence current value and the negative-sequence current value obtained from the current signal are used as observations. The impedance of each phase winding in the equivalent circuit model is used as the state variables to be inverted. A state estimation optimization model is constructed with the goal of minimizing the observation residual. The observation residual is the sum of the squares of the difference between the estimated zero-sequence current amplitude and the zero-sequence current value, and the difference between the estimated negative-sequence current amplitude and the negative-sequence current value. The estimated zero-sequence current amplitude and the estimated negative-sequence current amplitude are calculated based on the currently estimated impedance of each phase winding.

[0044] Optionally, the equivalent circuit model can employ the positive-sequence, negative-sequence, and zero-sequence equivalent circuits of the transformer windings. In the inversion calculation, it is typically assumed that the positive-sequence impedance and negative-sequence impedance are equal, and that the mutual impedance relationship is known or can be obtained through offline testing. This allows the state variables to be inverted to focus primarily on the self-impedance changes directly related to the faulty phase, thereby reducing the complexity of the inversion problem and improving the convergence speed. The standard impedance reference value is obtained by collecting three-phase current and voltage data over a long period of known normal operation of the transformer and averaging them using a similar state inversion method or offline parameter identification method. It can be understood that the actual observed amplitudes of the zero-sequence current and negative-sequence current obtained through the symmetrical component method provide a direct measure of the system asymmetry, while establishing the equivalent circuit model links the physical fault (phase loss) with changes in circuit parameters (impedance). The state inversion calculation process is essentially a process of using observable electrical quantities (zero-sequence and negative-sequence currents) to estimate internal parameters (impedances of each phase winding) that cannot be directly measured.

[0045] As an example, the state estimation optimization model can be represented as follows:

[0046] in, The objective function is the observed residuals. This represents the vector of state quantities to be inverted, composed of the impedances of each phase winding in the equivalent circuit model. and These represent the zero-sequence current value and the negative-sequence current value (actual observed amplitude) calculated from the three-phase digital current sequence within the current sliding window using the symmetrical component method, respectively. and Represents the state vector based on the current estimate. Based on the current system voltage conditions, the estimated values ​​of the zero-sequence current amplitude and the negative-sequence current amplitude are calculated using an equivalent circuit model. Under normal three-phase symmetrical conditions, the zero-sequence current and negative-sequence current values ​​are theoretically close to zero; when a single-phase loss occurs, the zero-sequence current and negative-sequence current values ​​will increase significantly. The goal of state inversion is to find a set of state variables (winding impedance) that makes the zero-sequence current amplitude and negative-sequence current amplitude calculated by the model as close as possible to this increased actual observed value.

[0047] The gradient descent method is used to solve the state estimation optimization model, and the impedance values ​​of each phase winding in the equivalent circuit model are iteratively adjusted. When the observation residual is less than the preset convergence threshold, the state inversion process is determined to be converged, and the final estimated value of the impedance of each phase winding is determined. Based on the final estimated value of the impedance of each phase winding and the standard impedance reference value, the relative change of the impedance of each phase winding before and after the phase loss fault is obtained.

[0048] The gradient descent method starts with a set of initial impedance values, calculates the gradient of the objective function J(Θ) with respect to each impedance parameter in the state vector Θ, updates the impedance values ​​along the gradient descent direction with a preset learning step size, and iterates the process until a stopping condition is met. When the observed residual is less than a preset convergence threshold, the state inversion process is considered converged, and the final estimated value of the impedance of each phase winding is recorded. The convergence threshold is a small positive number, for example, 0.01 (per unit). The final estimated value of the impedance of each phase winding is compared with the standard impedance reference value obtained statistically from the transformer's historical normal operating conditions to obtain the relative change in the impedance of each phase winding before and after the phase loss fault. This relative change is the change in equivalent circuit parameters. For example, assuming the standard impedance reference value of phase A winding is... The final estimated value of the A-phase winding impedance obtained after inversion convergence is Then the relative change in phase A impedance Calculated as If a phase-loss fault occurs in phase A, its equivalent impedance will increase significantly, and the calculated... It is a positive value much greater than zero, while the relative change in impedance between phase B and phase C is... and The expected value is close to zero. Therefore, by inferring the relative changes in the equivalent self-impedance of each phase winding of the transformer through state inversion calculation, the actual impact of phase loss faults on the electrical parameters of the transformer body is directly quantified. This provides fault evidence based on the physical model that is independent of the initial probability of the model. The relative change in impedance of the faulty phase will show a significant increase, while the relative change in impedance of the non-faulty phase will remain basically unchanged. This provides an objective basis for verifying and correcting the initial probability of the model from the perspective of circuit physics.

[0049] Figure 2 This application provides a diagram illustrating the changes in zero-sequence / negative-sequence current amplitude before and after a transformer phase-loss fault. Figure 2 As shown, the curve reflects the amplitude changes of zero-sequence current and negative-sequence current before and after the fault. Under normal operating conditions (00:00–00:30), both the zero-sequence current amplitude and the negative-sequence current amplitude are close to 0, with minimal fluctuations, consistent with the theoretical characteristics of three-phase symmetrical operation. The current amplitude remains stable below 0.05 per unit, without significant abnormal fluctuations. At the fault occurrence time (00:30), marked by the black dashed line, the fault trigger point, the current amplitude undergoes a step-like change. The zero-sequence current amplitude jumps sharply from near 0 to the 1.2–1.5 per unit range, while the negative-sequence current amplitude simultaneously jumps to the 0.7–0.9 per unit range. After the fault condition (00:30), the zero-sequence current amplitude is consistently significantly higher than the negative-sequence current amplitude, both remaining at a high level, reflecting the system asymmetry caused by the phase loss fault. The curve exhibits small random fluctuations, simulating the effects of measurement noise and load fluctuations in actual engineering.

[0050] Step 105: Correct the initial probability using the relative change to obtain the confidence level of each phase failure.

[0051] In some embodiments of this application, the relative change can be compared with a preset impedance change threshold to generate a hard decision indication; a Bayesian update rule is adopted, the initial probability is used as the prior probability, the hard decision indication is used as new evidence, the posterior probability of each phase having a phase failure is calculated under the hard decision indication evidence, and the posterior probability is used as the confidence level of each phase having a phase failure. The corrected confidence level can filter out misjudged data in the initial probability.

[0052] The Bayesian update rule can be represented as follows:

[0053] in, This indicates that evidence indicating a hard judgment was observed. Under the condition that, assuming The posterior probability of its validity. This indicates that a phase loss fault has occurred in a certain phase. Indicates a hypothesis The prior probability of its validity. Indicates in the assumption Under the conditions that it is established, hard judgment directive evidence is observed. The probability, which can be set in advance based on historical data or simulation.

[0054] For example, it can be set that when a phase is indeed faulty, the conditional probability of generating "suspected fault" evidence by its impedance change exceeding the threshold is 0.9; and when a phase is normal, the conditional probability of generating "suspected fault" evidence by its impedance change mistakenly exceeding the threshold is 0.1. The marginal probability of observing evidence E can be expressed using the law of total probability. Calculation, where For example, suppose the initial probability of the phase loss identification model outputting phase A is... The value is 0.75, and the relative change in phase A impedance obtained from the example is 0.45 (greater than the threshold of 0.3), thus generating hard decision evidence for phase A. Assign a value of 1 to "suspected fault". (Default) It is 0.9. The value is 0.1. Then calculate... Substituting into Bayes' theorem, we obtain the posterior probability, or confidence level, of phase A. The same method is applied to phases B and C, resulting in an updated probability vector, such as [0.964, 0.032, 0.021], which serves as the final confidence level.

[0055] As an example, the impedance change threshold can be set to 0.3 (per unit). That is, when the relative change in impedance of a certain phase is greater than 0.3, a hard decision indicator (assigned a value of 1) for "suspected fault" of that phase is generated; otherwise, a hard decision indicator (assigned a value of 0) for "normal" is generated. Using a Bayesian update rule, the initial probability output by the phase failure identification model is used as the prior probability, and the hard decision indicator is used as new evidence. The posterior probability of a phase failure occurring in each phase under the hard decision indicator evidence is calculated, and this posterior probability is the corrected confidence level.

[0056] Step 106: Based on the confidence level and the preset fault threshold, determine whether each phase of the starting transformer has experienced a phase loss fault, and generate a phase loss detection report.

[0057] In some embodiments of this application, the confidence level can be compared with a fault threshold, which includes an alarm threshold and a tripping threshold, wherein the alarm threshold is less than the tripping threshold; if the confidence level is greater than or equal to the alarm threshold and less than the tripping threshold, the corresponding phase is determined to be a minor phase loss fault; if the confidence level is greater than or equal to the tripping threshold, the corresponding phase is determined to be a severe phase loss fault.

[0058] In some embodiments of this application, the phase failure detection report may include not only the phase failure detection result indicating whether a phase failure has occurred in each phase, but also the phase failure occurrence time. In the timestamp of the current signal, the time point at which the energy of the current signal first abnormally changes under the phase experiencing a phase failure is located, and this is taken as the corresponding phase failure occurrence time. The phase failure detection report can comprehensively present the multi-dimensional information of the fault, refining the representation of the fault detection results. The criterion for abnormal change can be that the phase energy value is lower than 50% of its historical average energy within three consecutive sliding windows, and the starting time point of the first sliding window that meets the condition is taken as the fault occurrence time. The assessment of fault severity integrates the final confidence level of the phase determined to be faulty and the magnitude of the change in equivalent circuit parameters; for example, a severity index can be defined. The results are then mapped to three levels: “mild,” “moderate,” and “severe.”

[0059] As an example, the alarm threshold can be set to 0.7, and the tripping threshold can be set to 0.9. Summarize the fault phase, fault occurrence time, and fault severity information, fill in the data according to a fixed format, and generate a phase loss detection report as shown in Table 1.

[0060] Table 1

[0061] Figure 3This application provides a diagram illustrating the energy variation of three-phase transformer signals under different operating conditions. It shows the time-varying patterns of the three-phase current signal energy under normal conditions and single-phase failures (A / B / C phases). This visualization is based on the signal energy feature extraction stage of phase failure detection using an optical current transformer (CT). Under normal conditions, the signal energy exhibits periodic sinusoidal fluctuations with a stable amplitude between 30 and 50, reflecting the symmetrical characteristics of balanced three-phase operation. During a phase failure (A phase), a step drop occurs at approximately 50, with the energy plummeting from about 40 to around 8, before maintaining a low energy level with minor fluctuations. During a phase failure (B phase), a step drop occurs at approximately 50, with the energy falling from about 40 to around 20, before slowly recovering to 40 and continuing periodic fluctuations. During a phase failure (C phase), the signal energy remains consistently high, fluctuating between 40 and 60, without a significant drop, and is highly similar to the waveform under normal conditions.

[0062] In some embodiments of this application, the system operation status record of the starting transformer during the unit start-up task can also be obtained; the phase failure detection result in the phase failure detection report can be verified by the system operation status record; if the verification is successful, the energy feature vector and the relative change of the impedance of each phase winding before and after the phase failure are used as new training samples to incrementally train the phase failure identification model and update the internal parameters of the phase failure identification model.

[0063] In one implementation, current data collected by the optical current transformer during the active monitoring and startup operation of the starting transformer is continuously collected. This data is archived and stored in fixed time periods, including normal operation data and data from suspected fault periods triggered by this method. After the starting transformer performs the unit startup task and carries the load, the accuracy of the phase loss detection report is verified using the load current data and corresponding system operation status records during this process. The verification method involves comparing the fault conclusions in the phase loss detection report generated by this method with actual protection device operation records, manual inspection records, or offline electrical test results.

[0064] In some embodiments, correctly verified detection cases and their corresponding energy feature vectors and equivalent circuit parameter changes can be stored as new training samples in the historical sample library. A detection case is considered "correctly verified" only if it meets specific conditions. For example, if the method reports a severe phase-loss fault in phase A, and subsequent operation records or protection information show that the main protection trips and on-site inspection confirms that the phase A lead connection point is burned out, then this case, along with its complete data stream from before the fault occurred to the fault being cleared (including the calculated energy feature vector and the ultimately derived equivalent circuit parameter changes), is marked as a positive sample (fault sample). Conversely, if the method reports a fault but subsequent verification shows that the transformer is actually fault-free, then this case is marked as a negative sample (normal sample) and is also stored in the historical sample library for model correction. The pre-trained phase-loss identification model is periodically incrementally trained using newly added training samples to update the model's internal parameters. The incremental training process employs an online learning or small-batch update strategy, for example, initiating a model parameter update every 100 new valid training samples. The update of the internal weight parameter θ of the pre-trained phase discontinuity recognition model follows the rules of the following form:

[0065] in, This represents the updated model parameter vector. This represents the model parameter vector before the update. The learning rate is a preset small positive number used to control the step size for each update. Represents the loss function In old parameters Below, regarding the newly added training sample set The gradient. Loss function. Hinge loss or cross-entropy loss can be used, with the same form as that used during the initial offline training of the model. By solving the above update rules, the internal parameters of the phase loss identification model are adapted to changes in equipment aging and the field environment. For example, the transformer winding resistance increases slowly with the increase of the number of years of operation, or the characteristics of the optical current transformer drift slightly with the ambient temperature. These slow changes will be gradually absorbed into the updated model parameters.

[0066] Optionally, add a new training sample set. When used for incremental training, it can be mixed with a subset of old samples randomly sampled from the historical sample library to prevent the model from "forgetting" previously learned knowledge when adapting to new data; this strategy is called experience replay. Learning rate The value can be set to a sequence that decays over time, for example... ,in It is the initial learning rate. It is the attenuation coefficient. It is the sequence number of the incremental training round, which helps the model update process to be more stable in the later stages.

[0067] It is understandable that the online adaptive update step constructs a closed-loop learning system. The phase loss identification model not only plays a role in production, but its detection results, after being verified in actual field operation, are fed back as training data to optimize the model itself. Continuously collecting current data from optical current transformers during active monitoring and start-up operation of the transformer ensures the continuity of data sources. Verifying the accuracy of the phase loss detection report using load current data is a key step in associating the model's judgment with the real state of the physical world; only verified data can serve as reliable training samples. Using correctly verified detection cases and their corresponding features and parameters as new training samples allows the training set to cover new patterns that may not be fully reflected in the original training set during actual equipment operation. Regularly using new samples for incremental training enables the model to track the slow drift of equipment status and the external environment, thereby maintaining and potentially improving its long-term detection performance and reliability.

[0068] Figure 4 This application provides a diagram illustrating the energy variation of three-phase current signals under a phase-A failure, demonstrating the impact of the phase-A failure on the three-phase energy distribution. From 0 to 6 seconds, the energies of phases A, B, and C all fluctuate synchronously within the range of 80 to 130, exhibiting highly similar waveforms, reflecting the energy synchronicity of balanced three-phase operation. The phase difference of the three-phase energy curves is stable, conforming to the energy distribution law of a symmetrical three-phase system. At 6 seconds, the energy of phase A experiences a step-like drop, plummeting from approximately 90 to around 30, and then fluctuating slightly within the range of 20 to 40. The energies of phases B and C do not show a significant drop, remaining within the normal fluctuation range of 80 to 130, with only slight phase shifts caused by system asymmetry. The sudden drop in phase A energy and the disruption of energy synchronicity among the three phases are the core basis for constructing the energy feature vector, which can be directly used as input for the phase failure identification model. By comparing the amplitude of the three-phase energy changes, the faulty phase can be accurately located, avoiding misjudgment of other phases.

[0069] By implementing the embodiments of this application, by constructing an energy feature vector that reflects the relative energy relationship between three-phase current signals, the traditional detection method is freed from the limitation of relying only on the feature representation of single-phase current amplitude or sequence component amplitude. It can fully characterize the correlation characteristics and coordinated change law of three-phase current at the energy level, and provide fault feature input with higher discrimination and stronger anti-interference ability for the phase failure identification model, thereby reducing the degree of distortion of the initial fault probability. Based on this, the zero-sequence current amplitude and negative-sequence current amplitude obtained by real-time calculation are combined to perform state inversion calculation, inferring the changes in the equivalent circuit parameters of the transformer's internal windings and leads before and after the occurrence of a phase-loss fault. This change is used to correct the initial probability output by the phase-loss identification model, and a dual-channel cross-validation mechanism integrating data-driven discrimination (phase-loss identification model) and physical mechanism verification (equivalent circuit model state inversion) is constructed. This effectively eliminates the random bias in the output results of a single model, enabling the fault confidence to accurately match the actual fault state of the transformer. It realizes the leap from simple qualitative judgment to multi-dimensional refined evaluation of phase-loss fault detection, providing sufficient information support for operation and maintenance personnel to quickly locate and handle faults, and effectively meeting the strict requirements of power plant start-up transformers for the accuracy and integrity of phase-loss detection.

[0070] Figure 5 This is a schematic diagram of a phase loss detection device for a power plant starting transformer, provided as an embodiment of this application. Figure 5 As shown, the power plant's start-up transformer phase loss detection device may include: a data acquisition module 501, a processing module 502, a first identification module 503, a determination module 504, a second identification module 505, and a monitoring module 506.

[0071] The acquisition module 501 is used to acquire the current signals of each phase of the starting transformer through an optical current transformer; Processing module 502 is used to construct an energy feature vector that reflects the relative energy relationship between three-phase current signals based on the correlation coefficient between the signal energy distribution curves of any two-phase current signals. The first identification module 503 is used to input the energy feature vector into the pre-trained phase failure identification model to obtain the initial probability of phase failure in each phase. The phase failure identification model is constructed using a support vector machine. The determination module 504 is used to construct an equivalent circuit model to describe the starting transformer under a phase loss fault, so as to obtain the zero-sequence current value and the negative-sequence current value as observations in the current signal, and to perform state inversion calculation using the impedance of each phase winding in the equivalent circuit model as the state quantity to be inverted, so as to obtain the relative change of each phase winding impedance before and after the phase loss fault occurs. The second identification module 505 is used to correct the initial probability using the relative change amount to obtain the confidence level of each phase failure. The monitoring module 506 is used to determine whether a phase failure has occurred in each phase of the starting transformer based on the confidence level and the preset fault threshold, and to generate a phase failure detection report.

[0072] In some embodiments of this application, the determining module 504 is specifically used to: use the zero-sequence current value and negative-sequence current value obtained from the current signal as observations, and use the impedance of each phase winding in the equivalent circuit model as the state variable to be inverted, construct a state estimation optimization model with the goal of minimizing the observation residual, wherein the observation residual is the sum of the squares of the difference between the zero-sequence current amplitude estimate and the zero-sequence current value, and the difference between the negative-sequence current amplitude estimate and the negative-sequence current value; the zero-sequence current amplitude estimate and the negative-sequence current amplitude estimate are calculated based on the currently estimated impedance of each phase winding; solve the state estimation optimization model using the gradient descent method, and iteratively adjust the impedance values ​​of each phase winding in the equivalent circuit model; when the observation residual is less than a preset convergence threshold, determine that the state inversion process has converged, and determine the final estimated value of the impedance of each phase winding; and obtain the relative change in the impedance of each phase winding before and after the phase loss fault based on the final estimated value of the impedance of each phase winding and the standard impedance reference value.

[0073] In some embodiments of this application, the second identification module 505 is specifically used to: compare the relative change with a preset impedance change threshold to generate a hard decision indication; adopt a Bayesian update rule, take the initial probability as the prior probability, take the hard decision indication as new evidence, calculate the posterior probability of each phase experiencing a phase failure under the hard decision indication evidence, and take the posterior probability as the confidence level of each phase experiencing a phase failure.

[0074] In some embodiments of this application, the monitoring module 506 is specifically used to: compare the confidence level with the fault threshold, the fault threshold including the alarm threshold and the trip threshold, the alarm threshold being less than the trip threshold; if the confidence level is greater than or equal to the alarm threshold and less than the trip threshold, determine that the corresponding phase is a minor phase loss fault; if the confidence level is greater than or equal to the trip threshold, determine that the corresponding phase is a serious phase loss fault.

[0075] In some embodiments of this application, the phase failure detection report also includes the phase failure occurrence time; the monitoring module 506 is further configured to: locate the time point in the timestamp of the current signal where the phase failure occurs and where the energy of the current signal first undergoes an abnormal change, and use it as the corresponding phase failure occurrence time.

[0076] In some embodiments of this application, such as Figure 5Based on the illustrated embodiment, the power plant starting transformer phase loss detection device may further include a model update module. The model update module is used to: acquire system operation status records during the starting transformer's unit startup process; verify the phase loss fault detection results in the phase loss detection report using the system operation status records; and if the verification is successful, incrementally train the phase loss identification model using the energy feature vector and the relative changes in the impedance of each phase winding before and after the phase loss fault as new training samples, thereby updating the internal parameters of the phase loss identification model.

[0077] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0078] To implement the above embodiments, this application also proposes an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0079] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0080] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0081] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0082] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0083] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0084] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0085] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0086] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0087] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0088] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for detecting phase loss in a power plant starting transformer, characterized in that, Includes the following steps: The current signals of each phase of the starting transformer are collected by an optical current transformer; Based on the correlation coefficient between the signal energy distribution curves of any two phase current signals, an energy feature vector reflecting the relative energy relationship between the three-phase current signals is constructed. The energy feature vector is input into a pre-trained phase failure identification model to obtain the initial probability of phase failure in each phase. The phase failure identification model is constructed using a support vector machine. An equivalent circuit model is constructed to describe the starting transformer under a phase loss fault. The zero-sequence current value and the negative-sequence current value are calculated from the current signal as observations. The impedance of each phase winding in the equivalent circuit model is used as the state variable to be inverted to perform state inversion calculation, and the relative change of the impedance of each phase winding before and after the phase loss fault occurs is obtained. The initial probability is corrected using the relative change to obtain the confidence level of each phase failure. Based on the confidence level and the preset fault threshold, determine whether each phase of the starting transformer has experienced a phase loss fault, and generate a phase loss detection report.

2. The method according to claim 1, characterized in that, The energy feature vector is represented as follows: in, The energy feature vector, , , The Pearson correlation coefficients are the signal energy distribution curves of phases AB, AC, and BC, respectively. , , These represent the average energy ratios of phases A, B, and C during the current time period.

3. The method according to claim 1, characterized in that, The zero-sequence current value and negative-sequence current value obtained from the current signal are used as observations. State inversion calculations are performed using the impedance of each phase winding in the equivalent circuit model as the state variables to be inverted, to obtain the relative changes in the impedance of each phase winding before and after the phase loss fault, including: Using the zero-sequence current value and negative-sequence current value obtained from the current signal as observations, and the impedance of each phase winding in the equivalent circuit model as the state variables to be inverted, a state estimation optimization model is constructed with the goal of minimizing the observation residual. The observation residual is the sum of the squares of the differences between the zero-sequence current amplitude estimate and the zero-sequence current value, and the negative-sequence current amplitude estimate and the negative-sequence current value. The zero-sequence current amplitude estimate and the negative-sequence current amplitude estimate are calculated based on the currently estimated impedance of each phase winding. The state estimation optimization model is solved using the gradient descent method, and the impedance values ​​of each phase winding in the equivalent circuit model are iteratively adjusted. When the observation residual is less than the preset convergence threshold, the state inversion process is determined to be converged, and the final estimated value of the impedance of each phase winding is determined. Based on the final estimated values ​​of the impedance of each phase winding and the standard impedance reference value, the relative changes in the impedance of each phase winding before and after the phase loss fault occur are obtained.

4. The method according to claim 1, characterized in that, The step of correcting the initial probability using the relative change to obtain the confidence level of each phase failure includes: The relative change is compared with a preset impedance change threshold to generate a hard decision indication; Using a Bayesian update rule, the initial probability is taken as the prior probability, and the hard decision indication is taken as the new evidence. The posterior probability of each phase experiencing a phase failure is calculated under the hard decision indication evidence, and the posterior probability is taken as the confidence level of each phase experiencing a phase failure.

5. The method according to claim 4, characterized in that, The Bayesian update rule is expressed as follows: in, This indicates that evidence indicating a hard judgment was observed. Under the condition that, assuming The posterior probability of its validity. This indicates that a phase loss fault has occurred in a certain phase. Indicates a hypothesis The prior probability of its validity. Indicates in the assumption Under the conditions that it is established, hard judgment directive evidence is observed. The probability of.

6. The method according to claim 1, characterized in that, The step of determining whether a phase loss fault has occurred in each phase of the starting transformer based on the confidence level and a preset fault threshold includes: The confidence level is compared with the fault threshold, which includes an alarm threshold and a trip threshold, wherein the alarm threshold is less than the trip threshold; If the confidence level is greater than or equal to the alarm threshold and the confidence level is less than the tripping threshold, the corresponding phase is determined to be a minor phase loss fault. If the confidence level is greater than or equal to the tripping threshold, the corresponding phase is determined to be a severe phase loss fault.

7. The method according to claim 6, characterized in that, The phase loss detection report also includes the time of occurrence of the phase loss fault; the method further includes: In the timestamp of the current signal, the time point at which the energy of the current signal under the phase in which the phase failure occurs first is located and taken as the corresponding time of occurrence of the phase failure.

8. The method according to claim 1, characterized in that, Also includes: Obtain the system operation status record during the process of the starting transformer performing the unit startup task; The phase failure detection results in the phase failure detection report are verified by the system operation status record; If the verification is successful, the energy feature vector and the relative changes in the impedance of each phase winding before and after the phase failure are used as new training samples to incrementally train the phase failure identification model and update the internal parameters of the phase failure identification model.

9. A phase loss detection device for a power plant starting transformer, characterized in that, include: The acquisition module is used to acquire the current signals of each phase of the starting transformer through an optical current transformer; The processing module is used to construct an energy feature vector that reflects the relative energy relationship between the three-phase current signals based on the correlation coefficient between the signal energy distribution curves of any two phases of the current signals. The first identification module is used to input the energy feature vector into a pre-trained phase failure identification model to obtain the initial probability of phase failure in each phase. The phase failure identification model is constructed using a support vector machine. The determination module is used to construct an equivalent circuit model describing the starting transformer under a phase loss fault, to obtain the zero-sequence current value and negative-sequence current value as observations from the current signal, and to perform state inversion calculation using the impedance of each phase winding in the equivalent circuit model as the state quantity to be inverted, so as to obtain the relative change in the impedance of each phase winding before and after the phase loss fault occurs. The second identification module is used to correct the initial probability using the relative change amount to obtain the confidence level of each phase failure. The monitoring module is used to determine whether a phase failure has occurred in each phase of the starting transformer based on the confidence level and the preset fault threshold, and to generate a phase failure detection report.

10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-8.

Images