A fault recognition method based on ultra-high frequency phase characteristic change

Abstract

This application provides a fault identification method based on UHF phase characteristic changes, belonging to the field of fault identification technology. The method includes: acquiring the original electromagnetic signal of the target power equipment, performing preprocessing and phase demodulation to generate initial phase time series data; extracting the phase consistency coefficient, phase mutation index, and phase drift gradient within a preset time window to construct an initial three-dimensional phase feature vector; dynamically correcting the initial three-dimensional phase feature vector based on preset performance parameters of the phase-locked loop and historical data processing errors to generate an optimized three-dimensional phase feature vector; inputting the vector into a pre-trained fault classification model to calculate the matching probability between the output and each typical fault mode; performing multi-dimensional decision analysis based on the matching probability and the correlation weight matrix to identify and output the fault type identification result. This method solves the technical problems of insufficient identification accuracy and low precision in existing fault identification methods based on UHF phase characteristics.

Description

Technical Field

[0001] This invention relates to the field of fault identification, and in particular to a fault identification method based on ultra-high frequency phase characteristic changes. Background Technology

[0002] Early fault identification of power equipment is a core requirement for ensuring the safe operation of the power grid. Compared with parameters such as amplitude and frequency, ultra-high frequency phase characteristics are more sensitive to the early stages of faults and are less affected by the shielding of equipment structure, making them suitable for fault identification.

[0003] However, current fault identification technologies based on UHF phase characteristics mostly rely on a single phase parameter, which cannot distinguish the differentiated characteristics of different faults. Especially when facing multi-fault coupled scenarios, a single phase parameter is prone to confuse the dominant fault factors, leading to fault identification deviation. Moreover, in actual operation, there is complex electromagnetic interference, which will cause noise to be superimposed on the single phase parameter, further masking the true fault characteristics, resulting in insufficient identification accuracy and low accuracy, making it difficult to meet the needs of refined operation and maintenance of the power grid.

[0004] Therefore, there is an urgent need for a fault identification method based on UHF phase feature changes that can extract multi-dimensional phase features and has strong anti-interference capabilities, in order to break through the existing technical bottlenecks. Summary of the Invention

[0005] This invention addresses the technical problems of insufficient identification accuracy and low precision in existing fault identification methods based on ultra-high frequency phase characteristics, and provides a fault identification method based on changes in ultra-high frequency phase characteristics.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This invention provides a fault identification method based on ultra-high frequency phase characteristic changes, comprising: The original electromagnetic signal monitored by the ultra-high frequency sensing unit of the target power equipment is acquired, and the original electromagnetic signal is preprocessed and demodulated to generate initial phase time series data. Extract the phase consistency coefficient, phase abruptness index and phase drift gradient of the initial phase time series data within a preset time window to construct an initial three-dimensional phase feature vector; Based on the preset performance parameters of the phase-locked loop and the historical data processing errors during the phase demodulation process, the initial three-dimensional phase feature vector is dynamically corrected to generate an optimized three-dimensional phase feature vector. The optimized three-dimensional phase feature vector is input into the pre-trained fault classification model, and the matching probability of the output with each typical fault mode is calculated. Based on the matching probability and the association weight matrix, a multidimensional decision analysis is performed to identify and output the fault type identification result.

[0007] The beneficial effects of this invention are: Compared to existing technologies, this application first acquires the raw electromagnetic signals monitored by the UHF sensing unit of the target power equipment. The raw electromagnetic signals are preprocessed and demodulated to generate initial phase time-series data. This data is then accurately extracted from the complex electromagnetic environment of the power equipment, providing reliable input data for subsequent fault identification. Secondly, the phase consistency coefficient, phase mutation index, and phase drift gradient of the initial phase time-series data within a preset time window are extracted to construct an initial three-dimensional phase feature vector, providing a reliable data foundation for subsequent dynamic correction and model input. Thirdly, based on the preset performance parameters of the phase-locked loop (PLL) during phase demodulation and historical data processing errors, the initial three-dimensional phase feature vector is dynamically corrected to generate an optimized three-dimensional phase feature vector, improving its accuracy and better reflecting the complex hardware configuration and electromagnetic environment of the site. Furthermore, the optimized three-dimensional phase feature vector is input into a pre-trained fault classification model, calculating the matching probability between the output and each typical fault mode, improving the accuracy and precision of fault identification, and providing reliable quantitative basis for subsequent multi-dimensional decision analysis. Finally, multidimensional decision analysis is performed based on the matching probability and the association weight matrix to identify and output the fault type identification results. By combining the matching probability and feature contribution for multidimensional verification, misjudgment by relying solely on model probability is effectively avoided, ensuring that the output fault type identification results are both accurate and adaptable to the scenario.

[0008] Through the above technical solutions, this application obtains high-quality phase consistency coefficients, phase mutation indices, and phase drift gradients via preprocessing and phase demodulation, overcoming the limitation that single parameters cannot distinguish fault differences and accurately capturing the differentiated characteristics of different fault types. Based on the preset performance parameters of the phase-locked loop and historical data, the feature vector is dynamically corrected to effectively offset the impact of complex electromagnetic interference and hardware errors on the features. A pre-trained fault classification model is used to calculate the matching probability, and multi-dimensional decision analysis is conducted in conjunction with the correlation weight matrix, avoiding misjudgments based solely on probability and adapting to multi-fault coupling scenarios. Thus, the accuracy, reliability, and precision of power equipment fault identification are improved, while also considering adaptability to complex operating conditions and meeting the needs of refined power grid operation and maintenance. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a fault identification method based on ultra-high frequency phase characteristic changes provided by the present invention; Figure 2 This is a flowchart illustrating the pre-training steps of the fault classification model in a fault identification method based on ultra-high frequency phase characteristic changes provided by the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] In the description of this invention, 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 indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0012] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0013] Examples, such as Figure 1 As shown, this embodiment of the invention provides a fault identification method based on ultra-high frequency phase characteristic changes, including: S10: Acquire the original electromagnetic signal monitored by the UHF sensing unit of the target power equipment, preprocess and demodulate the original electromagnetic signal to generate initial phase timing data.

[0014] When power equipment malfunctions, its internal abnormal electromagnetic phenomena are accompanied by ultra-high frequency electromagnetic radiation in the 300MHz-3GHz frequency band. For example, partial discharge will generate nanosecond-level pulsed ultra-high frequency signals, and insulation degradation will release continuous low-frequency ultra-high frequency radiation. However, when power equipment is operating normally, it will only radiate extremely weak ultra-high frequency signals due to the influence of insulation medium and structural shielding, and these signals are irregular and have no fault correlation characteristics, thus having minimal interference to subsequent analysis.

[0015] To address the aforementioned issues, this application acquires the original electromagnetic signals monitored by the ultra-high frequency sensing unit of the target power equipment, preprocesses and demodulates the original electromagnetic signals, and generates initial phase timing data.

[0016] Specifically, step S10 in the method includes: The original electromagnetic signals radiated by the target power equipment during operation are collected by the ultra-high frequency sensing unit. The original electromagnetic signal is subjected to noise reduction and signal enhancement processing, and the instantaneous phase value of the signal is demodulated based on phase-locked loop technology; The instantaneous phase values ​​are sampled and recorded at equal time intervals to generate the initial phase time series data.

[0017] In this embodiment, the raw electromagnetic signals radiated by the target power equipment during operation are first acquired using an ultra-high frequency (UHF) sensing unit. The UHF sensing unit is a wideband, high-sensitivity sensing device used to acquire electromagnetic signals radiated by power equipment during operation. The wideband ensures that electromagnetic radiation from faults such as partial discharge and insulation degradation is not missed, while the high sensitivity ensures the capture of radiation signals with intensity as low as microwatts (μW) in the early stages of a fault. For example, the UHF sensing unit can be a UHF ultra-wideband antenna, a capacitively coupled sensor, etc. The UHF ultra-wideband antenna is adapted to open equipment, such as transformer bushings and switchgear, while the capacitively coupled sensor is adapted to enclosed equipment, such as GIS metal casings and cable terminals, ensuring the capture of weak fault radiation signals.

[0018] Secondly, since the original electromagnetic signal is mixed with interference signals such as electromagnetic noise and white noise from other devices, direct demodulation would result in severe phase distortion. Therefore, it is necessary to perform noise reduction and signal enhancement processing on the original electromagnetic signal, and then demodulate the instantaneous phase value of the signal based on phase-locked loop (PLL) technology. PLL technology is chosen for demodulation because ultra-high frequency signals have high frequencies, making it difficult for conventional demodulation methods to track phase changes in real time. A PLL, however, can accurately track the phase of the input signal through closed-loop control, outputting a stable instantaneous phase value.

[0019] Finally, the instantaneous phase values ​​are sampled and recorded at equal time intervals to generate initial phase time series data. The sampling interval needs to be set according to the time scale of the fault characteristics to avoid phase information aliasing or loss. For example, for transient faults such as partial discharge, the phase change time scale is in the nanosecond (ns) range, requiring a very small sampling interval, such as 100 ps / time or 1 ns / time, to ensure complete capture of phase jump details. For progressive faults such as insulation degradation and mechanical loosening, the phase drift time scale is in the millisecond (ms) or microsecond (μs) range, allowing for a more relaxed sampling interval, such as 1 μs / time or 10 μs / time, to reduce data redundancy while ensuring feature integrity and balancing storage pressure and real-time processing efficiency.

[0020] For example, the demodulated instantaneous phase values ​​are sampled and recorded according to a preset equal-interval time sequence to form initial phase time series data. For instance, within a 10ms analysis time window, 10,000 phase data points can be generated at a sampling interval of 1μs / sample, forming initial phase time series data: [phase 1 (t=0μs), phase 2 (t=1μs), ..., phase 10000 (t=10ms)]. The initial phase time series data can reflect the phase change trend over time, which is the key to distinguishing different faults.

[0021] Specifically, the step of "performing noise reduction and signal enhancement processing on the original electromagnetic signal, and demodulating the instantaneous phase value of the signal based on phase-locked loop technology" includes: The original electromagnetic signal is subjected to wavelet transform noise reduction processing. A wavelet basis function matching the characteristics of the UHF signal is selected, and environmental noise and white noise interference are suppressed by threshold filtering to obtain the first preprocessed signal. The Wiener filtering algorithm is used to adaptively filter the first preprocessed signal to obtain the second preprocessed signal; The second preprocessed signal is input to the phase-locked loop circuit, wherein the loop bandwidth of the phase-locked loop is preset according to the center frequency of the ultra-high frequency signal and the expected phase jitter range; The instantaneous phase value of the signal is tracked and demodulated in real time by comparing the phase-locked loop voltage-controlled oscillator output signal with the second preprocessed signal.

[0022] In this embodiment, the original electromagnetic signal is first subjected to wavelet transform denoising processing. Wavelet basis functions matching the characteristics of the UHF signal, such as db4, sym8, and coif5, are selected. A threshold filtering method is then used to suppress environmental noise and white noise interference, obtaining the first preprocessed signal. The UHF signal is characterized by high frequency and short-duration pulses. Wavelet basis functions such as db4 have short support lengths and high high-frequency resolution, enabling precise decomposition of the useful fault components and noise interference components in the signal. The threshold filtering method sets a noise threshold based on the noise statistical characteristics (such as noise variance and signal-to-noise ratio) of the UHF signal (300MHz-3GHz). Wavelet coefficients smaller than the noise threshold are mostly environmental noise and white noise, and are directly set to zero. Wavelet coefficients larger than the noise threshold mostly carry UHF fault signals and are retained or adjusted.

[0023] For example, for the acquired raw electromagnetic signal, a wavelet basis function matching the high-frequency pulse characteristics of the signal is first selected, such as db4 or sym8, and wavelet coefficients of different scales are obtained through 3-5 layers of wavelet decomposition. Then, a threshold filtering method is used to perform threshold filtering on the wavelet coefficients through a set noise threshold. Finally, the processed wavelet coefficients are reconstructed into an electromagnetic signal through inverse wavelet transform, resulting in a first preprocessed signal in which environmental noise and white noise are suppressed and fault characteristics are preserved.

[0024] Secondly, since wavelet transform denoising relies on a static noise threshold for filtering, it is insufficiently adaptable to non-stationary interference. Therefore, the Wiener filtering algorithm is needed to adaptively filter the first preprocessed signal to obtain a better second preprocessed signal. The advantage of the Wiener filtering algorithm lies in its dynamic adaptation of the noise threshold based on statistical characteristics. It can first analyze the statistical characteristics of the current electromagnetic signal and noise in real time, such as mean, variance, and autocorrelation function, and then dynamically construct the optimal filter coefficients based on the minimum mean square error criterion. Through this adaptive mechanism, Wiener filtering can accurately distinguish between useful UHF fault signals and residual non-stationary noise: while preserving key signal characteristics such as partial discharge pulses and insulation degradation radiation to the maximum extent, it effectively suppresses residual sudden interference and dynamic noise after wavelet denoising, further improving the signal-to-noise ratio and avoiding phase demodulation errors caused by noise.

[0025] Next, the second preprocessed signal is input to the phase-locked loop (PLL) circuit. The PLL's loop bandwidth is preset based on the center frequency of the UHF signal and the expected phase jitter range: if the UHF signal's center frequency is high, the PLL's loop bandwidth needs to be appropriately widened to avoid tracking lag; the expected phase jitter range is the PLL's own error range. If the expected phase jitter range is large, the PLL's loop bandwidth needs to be narrowed to reduce noise introduction; if the expected phase jitter range is small, the PLL's loop bandwidth can be widened to improve tracking speed. In this way, the PLL's loop bandwidth can balance phase tracking speed and anti-interference capability, avoiding demodulation phase distortion due to improper parameters.

[0026] Finally, the instantaneous phase value of the signal is tracked and demodulated in real time by comparing the phase of the voltage-controlled oscillator (VCO) output signal of the phase-locked loop (PLL) with the phase of the input second preprocessed signal. For example, the VCO outputs a reference signal, which is compared with the phase of the input second preprocessed signal. If there is a deviation, the PLL adjusts the VCO output until the phase of the VCO output signal is synchronized with the phase of the input second preprocessed signal. At this point, the phase of the VCO output signal approximately represents the instantaneous phase of the input second preprocessed signal, completing demodulation. In this way, the instantaneous phase value changing over time is accurately extracted from the second preprocessed signal, providing a data foundation for subsequent generation of timing data and extraction of fault features.

[0027] It should be noted that wavelet transform and threshold filtering are existing technologies widely used in the field of signal denoising. Their technical principles and basic implementation processes have been well documented in professional literature such as "Wavelet Analysis and Its Engineering Applications" and in engineering practices of power equipment signal processing. Those skilled in the art can master their specific operational details based on existing publicly available information. Therefore, this application will not elaborate further on the specific technical details of wavelet transform and threshold filtering.

[0028] Further, the step of "comparing the phase of the voltage-controlled oscillator output signal of the phase-locked loop with the phase of the second preprocessed signal, and tracking and demodulating the instantaneous phase value of the signal in real time" includes: The second preprocessed signal and the output signal of the voltage-controlled oscillator are input to a phase detector for phase comparison to generate a phase error signal; The phase error signal is filtered by a loop filter, wherein the bandwidth of the loop filter is adaptively adjusted according to the phase noise characteristics of the ultra-high frequency signal. The filtered error signal is used as a control voltage input to the voltage-controlled oscillator to form a closed-loop control circuit. The phase information of the voltage-controlled oscillator output signal is collected in real time as the demodulated instantaneous phase value, wherein the sampling frequency of the instantaneous phase value is not less than twice the center frequency of the ultra-high frequency signal.

[0029] In this embodiment, the second preprocessed signal and the output signal of the voltage-controlled oscillator (VCO) are first input to a phase detector for phase comparison to generate a phase error signal. The phase detector is a phase comparator that compares the instantaneous phases of the second preprocessed signal and the VCO output signal to generate the phase error signal. For example, if the instantaneous phase of the second preprocessed signal is φ1 and the instantaneous phase of the VCO output signal is φ2, then the phase error signal = φ1 - φ2. The sign of the phase error signal depends on the magnitude of the instantaneous phases of the two signals. In this way, the phase deviation is converted into a measurable phase error signal, providing a basis for subsequent adjustment of the VCO.

[0030] Secondly, since the phase error signal still contains high-frequency noise, such as the noise from the phase detector itself, it needs to be filtered by a loop filter. The bandwidth of the loop filter is adaptively adjusted according to the phase noise characteristics of the ultra-high frequency signal. For example, depending on the phase noise characteristics, if the on-site phase noise is strong, the bandwidth of the loop filter needs to be narrowed to filter out high-frequency noise and make the error signal smoother. If the on-site phase noise is weak, the bandwidth of the loop filter can be widened to retain the details of the error signal and improve the adjustment speed. In this way, by filtering out high-frequency noise through the loop filter, a stable error signal is output, avoiding any impact on phase tracking stability.

[0031] Next, the filtered error signal is used as the control voltage input to the voltage-controlled oscillator (VCO), forming a closed-loop control circuit. For example, the output frequency / phase of the VCO changes with the control voltage. If the error signal is positive (the second preprocessed signal leads in phase), the control voltage increases, the VCO output frequency increases, and the phase catches up faster. If the error signal is negative (the second preprocessed signal lags in phase), the control voltage decreases, the VCO output frequency decreases, and the phase slows down. Ultimately, the phase of the VCO output signal is dynamically synchronized with the phase of the second preprocessed signal, avoiding phase demodulation interruption or distortion due to signal fluctuations.

[0032] Finally, the phase information of the voltage-controlled oscillator (VCO) output signal is acquired in real time as the instantaneous phase value after demodulation. The phase information of the VCO output signal is acquired because, under closed-loop control, the VCO output phase is synchronized with the input signal phase; therefore, the phase of the VCO output signal directly represents the instantaneous phase of the input signal. The sampling frequency of the instantaneous phase value is no less than twice the center frequency of the UHF signal. This is to comply with the Nyquist sampling theorem; if the sampling frequency is less than twice the center frequency of the UHF signal, phase aliasing will occur, meaning the sampled phase data cannot reconstruct the true phase change. For example, if the center frequency of the UHF signal is 1 GHz, the sampling frequency of the instantaneous phase value must be ≥2 GHz to completely preserve the phase information. In this way, aliasing-free, high-fidelity instantaneous phase values ​​are obtained, providing accurate raw data for generating the initial phase timing data.

[0033] It should be noted that phase-locked loop (PLL) technology and its core components, voltage-controlled oscillators (VCOs) and phase detectors, are mature existing technologies in the fields of electronic communication and phase control. Their basic working principles and hardware implementation methods have been thoroughly documented in professional works such as "PLL Design and Application" and in engineering practices of power equipment phase demodulation. Those skilled in the art can readily master the specific details of parameter configuration and performance debugging based on existing publicly available information. Therefore, this application will not elaborate on the specific technical details such as the loop bandwidth design of the PLL, the frequency response characteristics of the VCO, and the linear phase detection range of the phase detector.

[0034] In summary, compared to existing technologies, this application acquires the raw electromagnetic signals monitored by the ultra-high frequency sensing unit of the target power equipment, preprocesses and demodulates the raw electromagnetic signals to generate initial phase timing data. Thus, initial phase timing data is accurately extracted from the complex electromagnetic environment of the power equipment, providing reliable input data for subsequent fault identification.

[0035] S20: Extract the phase consistency coefficient, phase mutation index and phase drift gradient of the initial phase time series data within a preset time window, and construct an initial three-dimensional phase feature vector.

[0036] In the identification of ultra-high frequency faults in power equipment, traditional methods usually rely on a single phase parameter, such as focusing only on phase abrupt changes or amplitude variations, which makes it difficult to cover the differentiated characteristics of different faults.

[0037] To address the aforementioned issues, this application extracts the phase consistency coefficient, phase mutation index, and phase drift gradient of the initial phase time series data within a preset time window to construct an initial three-dimensional phase feature vector.

[0038] Specifically, step S20 in the method includes: Calculate the reciprocal of the standard deviation of the initial phase time series data within the preset time window, and use the normalized result as the phase consistency coefficient. The maximum absolute value of the first-order difference of the initial phase time series data within the preset time window is calculated as the phase change index; Linear fitting is performed on the initial phase time series data within the preset time window, and the absolute value of the fitting slope is used as the phase drift gradient. The phase consistency coefficient, phase mutation index, and phase drift gradient are combined in a preset order to construct the initial three-dimensional phase feature vector.

[0039] In this embodiment, the reciprocal of the standard deviation of the initial phase timing data within a preset time window is first calculated and normalized to serve as the phase consistency coefficient. The duration of the preset time window can be set according to the timescale of power equipment faults. Preferably, a preset time window of 10-20ms is suitable for sudden faults such as partial discharge to ensure complete coverage of pulse phase changes, while a preset time window of 50-100ms is suitable for gradual faults such as insulation degradation to capture long-term phase shift trends.

[0040] For example, the standard deviation of the initial phase time series data can reflect the degree of phase dispersion. The larger the standard deviation of the initial phase time series data, the greater the dispersion, i.e., the worse the stability. Then, the reciprocal of the standard deviation of the initial phase time series data is calculated, positively correlated with stability and the phase consistency coefficient. Normalization eliminates dimensional differences between different devices, resulting in a higher final phase consistency coefficient, smaller phase fluctuations, and stronger stability. The phase consistency coefficient can be used to distinguish the operating state of equipment. Normal equipment typically has a high phase consistency coefficient due to a stable electromagnetic environment, while faults such as partial discharge and poor conductor contact can cause drastic phase fluctuations, resulting in a typically low phase consistency coefficient, which can serve as an early warning indicator for faults.

[0041] Secondly, the maximum absolute value of the first-order difference of the initial phase time series data within the preset time window is calculated as the phase jump index. The first-order difference of the initial phase time series data reflects the phase change between adjacent moments, and the maximum absolute value of the first-order difference represents the most dramatic phase jump. The phase jump index can be matched to sudden faults; for example, the nanosecond-level electromagnetic pulse generated by partial discharge will cause the phase jump index to be much higher than that of normal equipment or gradual faults, thus distinguishing between sudden and non-sudden fault types.

[0042] Next, linear fitting is performed on the initial phase time series data within the preset time window, and the absolute value of the fitting slope is used as the phase drift gradient. The fitting slope obtained from linear fitting reflects the phase change trend over time, while the absolute value represents the rate of gradual change. The larger the absolute value of the fitting slope, the faster the long-term phase shift. The phase drift gradient can match gradual faults. Insulation degradation and long-term equipment overload will cause a slow phase shift, with a phase drift gradient often >0.2° / ms. Normal equipment or sudden faults do not have a continuous shift, and the phase drift gradient is usually <0.05° / ms. This can be used to identify gradual faults such as equipment aging and degradation.

[0043] Finally, the phase consistency coefficient, phase mutation index, and phase drift gradient are combined in a preset order to construct an initial three-dimensional phase feature vector. For example, the initial three-dimensional phase feature vector can be formed by combining the phase consistency coefficient, phase mutation index, and phase drift gradient in a preset order. For instance, the initial three-dimensional phase feature vector obtained under a certain fault scenario is [0.7, 0.8, 0.03]. This initial three-dimensional phase feature vector breaks through the limitations of traditional fault identification relying on a single feature. It provides complementary characterization from three dimensions: stability, suddenness, and gradual trend, forming a unique description of different faults. For example, partial discharge corresponds to "low consistency, high mutation, low drift," while insulation degradation corresponds to "medium consistency, low mutation, high drift." This allows subsequent fault classification models to accurately match features with fault types, while providing structured and comparable input data for subsequent dynamic feature correction.

[0044] In summary, compared to existing technologies, this application extracts the phase consistency coefficient, phase mutation index, and phase drift gradient of the initial phase time series data within a preset time window to construct an initial three-dimensional phase feature vector. This yields a structured feature vector that can be used for fault classification, providing a reliable data foundation for subsequent dynamic correction and model input.

[0045] S30: Based on the preset performance parameters of the phase-locked loop and the historical data processing error during the phase demodulation process, the initial three-dimensional phase feature vector is dynamically corrected to generate an optimized three-dimensional phase feature vector.

[0046] While the initial three-dimensional phase feature vector obtained through the aforementioned steps can characterize phase changes, the inherent phase jitter of the phase-locked loop (PLL) causes random fluctuations in the demodulated instantaneous phase value. Even if the equipment is fault-free, the calculated three-dimensional phase feature vector will still have deviations. Furthermore, the feature vectors of normal signals will differ under different PLL parameter configurations, ultimately leading to potential feature distortion in the initial three-dimensional phase feature vector. Therefore, dynamic correction is necessary to counteract these two types of interference, ensuring that the three-dimensional phase feature vector accurately reflects fault characteristics rather than measurement errors.

[0047] To address the aforementioned issues, this application dynamically corrects the initial three-dimensional phase feature vector based on the preset performance parameters of the phase-locked loop and historical data processing errors during the phase demodulation process, thereby generating an optimized three-dimensional phase feature vector.

[0048] Specifically, step S30 in the method includes: The phase jitter variance of the phase-locked loop is obtained as a performance indicator of the phase-locked loop; Query the historical database for the baseline feature vector generated by normal signal processing under the same or similar phase-locked loop parameter configurations; Calculate the relative deviation of each component in the initial three-dimensional phase eigenvector and the reference eigenvector; The relative deviation is weighted and corrected according to the phase-locked loop performance index to obtain the corrected phase consistency coefficient, phase mutation index and phase drift gradient, and then recombined to obtain the optimized three-dimensional phase feature vector.

[0049] In this embodiment, the phase jitter variance of the phase-locked loop (PLL) is first obtained as a performance indicator of the PLL. For example, the PLL performance indicator can be obtained by consulting the PLL hardware manual or calculated based on real-time acquired instantaneous phase values. For example, if the target power equipment is a 110kV transformer with a UHF signal center frequency of 500MHz, and the sampling frequency is set to 1000MHz, five consecutive instantaneous phase values ​​(unit: rad) are collected: Y1=1.2, Y2=1.3, Y3=1.1, Y4=1.4, Y5=1.0. First, calculate the mean of the five instantaneous phase values ​​= (1.2+1.3+1.1+1.4+1.0) / 5 = 1.2rad. Then, calculate the sum of squared deviations = (1.2-1.2)²+(1.3-1.2)²+(1.1-1.2)²+(1.4-1.2)²+(1.0-1.2)² = 0.1. Finally, the phase jitter variance is calculated to be 0.02rad², which is used as the performance index of the phase-locked loop. Phase-locked loop (PLL) performance metrics quantify the inherent hardware errors of the PLL into calculable parameters. The larger the PLL performance metrics, the lower the demodulation accuracy of the PLL and the stronger the interference on the initial three-dimensional phase eigenvector.

[0050] Secondly, query the historical database for baseline feature vectors generated by normal signal processing under the same or similar phase-locked loop (PLL) parameter configurations. For example, select the three-dimensional phase feature vectors obtained after signal processing from power equipment with consistent key parameters such as loop bandwidth, sampling frequency, and phase jitter variance, and which is operating normally without faults, such as [0.9, 0.3, 0.02], as the baseline feature vectors. This eliminates the influence of differences in PLL parameter configurations on the feature vectors, ensuring that the deviation in subsequent calculations is a fault-related offset, rather than a normal difference caused by parameter differences.

[0051] Next, the relative deviations of each component in the initial three-dimensional phase eigenvector and the reference eigenvector are calculated. For example, the relative deviations of the phase consistency coefficient, phase abrupt change index, and phase drift gradient are calculated separately using the formula: Component relative deviation = (Initial component value - Reference component value) / Reference component value × 100%. For instance, if the initial three-dimensional phase eigenvector is [0.7, 0.8, 0.03] and the reference eigenvector is [0.9, 0.3, 0.02], then the relative deviation of the phase consistency coefficient = (0.7 - 0.9) / 0.9 × 100% = -22.2%, the relative deviation of the phase abrupt change index = (0.8 - 0.3) / 0.3 × 100% = 166.7%, and the relative deviation of the phase drift gradient = (0.03 - 0.02) / 0.02 × 100% = 50%. This quantifies the degree of deviation of each eigencomponent, allowing the location of the most severely disturbed eigencomponent.

[0052] Finally, the relative deviation is corrected by weighting according to the performance index of the phase-locked loop to obtain the corrected phase consistency coefficient, phase mutation index and phase drift gradient, and then recombined to obtain the optimized three-dimensional phase feature vector. The weighted correction formula is as follows: Corrected component = Initial component - Relative deviation of PLL performance index pair × Component relative deviation. For example, if the relative deviation of PLL performance index pair is 0.02, the initial three-dimensional phase feature vector is [0.7, 0.8, 0.03], and the component relative deviations are -22.2%, 166.7%, and 50% respectively, then the corrected phase consistency coefficient = 0.7 - 0.02 × (-22.2%) ≈ 0.704, the corrected phase mutation index = 0.8 - 0.02 × 166.7% ≈ 0.767, and the corrected phase consistency coefficient = 0.03 - 0.02 × 50% ≈ 0.02. The optimized three-dimensional phase feature vector is obtained by recombination: [0.704, 0.767, 0.02]. In this way, the feature vector distortion caused by hardware errors and parameter deviations of the PLL is eliminated, while the significant deviation caused by faults is retained, ensuring that the feature vector input to the fault classification model can accurately reflect the true state of the equipment.

[0053] In summary, compared to existing technologies, this application dynamically corrects the initial three-dimensional phase feature vector based on the preset performance parameters of the phase-locked loop (PLL) and historical data processing errors during the phase demodulation process, generating an optimized three-dimensional phase feature vector. Thus, based on the reference feature vector, the inherent hardware errors of the PLL are adaptively corrected, improving the accuracy of the phase feature vector and better suiting the complex hardware configurations and electromagnetic environments in the field.

[0054] S40: Input the optimized three-dimensional phase feature vector into the pre-trained fault classification model, and calculate the matching probability of the output with each typical fault mode.

[0055] The aforementioned steps, through corrective calculations, eliminate the influence of non-fault factors such as phase-locked loop jitter and parameter configuration differences, resulting in an optimized three-dimensional phase feature vector that retains the true phase fault characteristics of the power equipment. This vector can be combined with a pre-trained fault classification model for fault identification.

[0056] To address the aforementioned issues, this application inputs the optimized three-dimensional phase feature vector into a pre-trained fault classification model and calculates the matching probability of the output with each typical fault mode.

[0057] Specifically, such as Figure 2 As shown, the pre-training step of the fault classification model in step S40 of the method includes: Collect three-dimensional phase feature vector samples corresponding to multiple sets of UHF signals from historical fault records, and label their actual fault types to form a training sample set; Before supervised training, correlation analysis is performed on the three-dimensional phase feature vector samples in the training sample set. The correlation strength between the phase consistency coefficient, phase mutation index, phase drift gradient and specific fault type in each sample is calculated, and a correlation weight matrix is ​​generated. Using the training sample set, with the three-dimensional phase feature vector weighted according to the correlation weight matrix as input and the fault type as output label, supervised training is performed on the fault classification model built based on the multilayer perceptron. After the model converges, the model parameters and the associated weight matrix are saved to obtain the pre-trained fault classification model.

[0058] In this embodiment, firstly, three-dimensional phase feature vector samples corresponding to multiple sets of UHF signals from historical fault records are collected from a historical database, and their actual fault types are labeled to form a training sample set. The actual fault types of the training samples should cover typical fault types such as partial discharge, insulation degradation, and poor conductor contact as much as possible, and the dimensional order of the three-dimensional phase feature vector samples should be consistent with the initial three-dimensional phase feature vector.

[0059] Secondly, since different phase features contribute differently to the identification of different types of faults—for example, the phase abrupt change index is the core for distinguishing partial discharge, and the phase drift gradient is the key to identifying insulation degradation—a correlation analysis is performed on the three-dimensional phase feature vector samples in the training sample set before supervised training. This calculates the correlation strength between the phase consistency coefficient, phase abrupt change index, phase drift gradient, and specific fault type in each sample, and generates a correlation weight matrix. For example, the correlation strength between the phase consistency coefficient, phase abrupt change index, phase drift gradient, and each specific fault type can be quantified using mutual information or chi-square tests.

[0060] For example, suppose the training sample set contains 200 samples, divided into two categories according to fault type: 100 samples labeled as partial discharge faults and 100 samples labeled as insulation degradation faults. The mutual information method is used to quantify the correlation strength between features and faults. Statistical analysis of partial discharge fault samples shows that the mean of their phase abrupt change index is significantly higher than that of normal samples, and the phase abrupt change index within the sample is positively correlated with the discharge intensity. Mutual information calculations show that the correlation strength between the phase abrupt change index and partial discharge is 0.6, the correlation strength of the phase consistency coefficient is 0.3, and the correlation strength of the phase drift gradient is 0.1. Similarly, statistical analysis of insulation degradation faults yields correlation strengths of 0.3, 0.1, and 0.6, respectively. Finally, a correlation weight matrix is ​​generated, where rows represent fault types and columns represent three-dimensional phase features. The weights of each row in the correlation weight matrix sum to 1. This allows the model to focus on highly correlated features during training, weakening interference from irrelevant features and improving the ability to distinguish between different faults.

[0061] Next, using the training sample set, with the three-dimensional phase feature vector weighted according to the correlation weight matrix as input and the fault type as output label, the fault classification model based on the multilayer perceptron is trained in a supervised manner.

[0062] For example, a fault classification model can be constructed using a multilayer perceptron (MLP), which mainly consists of an input layer, a hidden layer, and an output layer. The input layer has three neurons, each corresponding to a weighted three-dimensional phase feature vector. The hidden layer has 1-2 layers, each with 64-128 neurons, and ReLU is used as the activation function. The number of neurons in the output layer is the same as the number of fault types, and the probability of the fault type is output through the Softmax activation function.

[0063] For example, during training, the training sample set is first weighted according to the correlation weight matrix to obtain a weighted three-dimensional phase feature vector. The weighted three-dimensional phase feature vector and the corresponding fault type are divided into training set, validation set and test set in a ratio of 7:1.5:1.5. The weighted three-dimensional phase feature vector in the training set is used as input and the corresponding fault type is used as label. The prediction error is calculated using the cross-entropy loss function. The model parameters are adjusted by the Adam optimizer. At the same time, Dropout or early stopping strategy is used to prevent overfitting. The training set loss is stable and the validation set accuracy fluctuates less than 0.5% for multiple rounds. This is considered convergence, and the trained fault classification model is obtained.

[0064] Finally, after the model converges, the model parameters and associated weight matrix are saved to obtain the pre-trained fault classification model.

[0065] For example, a pre-trained fault classification model is invoked, and an optimized three-dimensional phase feature vector [0.704, 0.767, 0.02] is input. Combined with the correlation weight matrix, the matching probability of the output with each typical fault mode is calculated. For example, partial discharge fault is 68%, poor conductor contact fault is 21%, insulation deterioration fault is 7%, and mechanical loosening fault is 3%. Through this probability distribution, the most likely fault type can be quickly located, realizing an efficient mapping from feature vector to fault type.

[0066] In summary, compared to existing technologies, this application inputs the optimized three-dimensional phase feature vector into a pre-trained fault classification model to calculate the matching probability between the output and each typical fault mode. Thus, by strengthening key features through the association weight matrix, the accuracy and precision of fault classification are improved, and a reliable quantitative basis is provided for subsequent multi-dimensional decision analysis.

[0067] S50: Perform multidimensional decision analysis based on the matching probability and association weight matrix to identify and output the fault type identification result.

[0068] The aforementioned steps, based on the fault classification model, obtained the matching probabilities between the optimized three-dimensional phase feature vector and each typical fault mode, providing a preliminary quantitative basis for fault identification. However, non-fault factors may interfere in field operations, such as instantaneous electromagnetic disturbances and minor fluctuations in equipment parameters, potentially leading to situations where the fault classification model outputs a high probability but the feature logic is inconsistent. Therefore, multi-dimensional decision analysis is necessary to ensure that the output fault type identification results are both efficient and reliable, avoiding misjudgments.

[0069] To address the aforementioned issues, this application performs multidimensional decision analysis based on the matching probability and association weight matrix to identify and output fault type identification results.

[0070] Specifically, step S50 in the method includes: Obtain the matching probability of each fault type output by the fault classification model and sort them in descending order; Query the association weight matrix of the fault type corresponding to the current highest matching probability, and analyze the contribution of the three components of phase consistency coefficient, phase mutation index and phase drift gradient in the optimized three-dimensional phase feature vector; If the contribution analysis shows that the contribution distribution of the three feature components is consistent with the preset ideal weight distribution of the corresponding fault type in the correlation weight matrix, then the corresponding fault type is directly output as the identification result. If the contribution analysis shows that only one feature component has a significantly higher contribution than the other two, and is highly similar to the typical characteristics of similar failures in the past, then the weight enhancement mechanism will be activated. If the contribution analysis shows that the contribution distribution of the three feature components is scattered and does not match the ideal weight distribution of any known fault type, then the corresponding identification event will be marked as an uncertain state, and the identification result will be pushed to the manual diagnosis terminal.

[0071] In this embodiment, the matching probabilities of each fault type output by the fault classification model are first obtained and sorted in descending order. For example, the matching probabilities of each fault type output by the fault classification model are sorted in descending order as follows: partial discharge fault 68%, insulation degradation fault 21%, poor conductor contact fault 7%, and mechanical loosening fault 3%.

[0072] Secondly, query the association weight matrix of the fault type corresponding to the highest matching probability, and analyze the contribution of the three components of phase consistency coefficient, phase mutation index and phase drift gradient in the three-dimensional phase feature vector. For example, if the three-dimensional phase feature vector [0.704, 0.767, 0.02] is optimized, and the weights of the phase consistency coefficient, phase mutation index, and phase drift gradient of the partial discharge fault obtained from the correlation weight matrix are 0.3, 0.6, and 0.1 respectively, then the contribution of the phase consistency coefficient is approximately 31% (0.3×0.704 / (0.3×0.704+0.6×0.767+0.1×0.02)), the contribution of the phase mutation index is approximately 68% (0.6×0.767 / (0.3×0.704+0.6×0.767+0.1×0.02)), and the contribution of the phase drift gradient is approximately 1% (0.1×0.02 / (0.3×0.704+0.6×0.767+0.1×0.02)).

[0073] Furthermore, if the contribution analysis shows that the contribution distribution of the three feature components is consistent with the preset ideal weight distribution of the corresponding fault type in the correlation weight matrix, then the corresponding fault type can be directly output as the identification result. For example, the relative deviation between the actual contribution of the three components in the optimized three-dimensional phase feature vector and the ideal weight of the corresponding fault type in the correlation weight matrix can be calculated using the formula: Relative deviation = |(Actual contribution - Ideal weight) / Ideal weight| × 100%. A reasonable deviation threshold, such as 10%, can then be preset. When all three calculated relative deviations meet the preset deviation threshold, it indicates that the actual contribution logic of the optimized three-dimensional phase feature vector highly matches the typical characteristic pattern of the corresponding fault type. In this case, the fault type can be directly output as the identification result, ensuring the reliability and rigor of the result.

[0074] Furthermore, if contribution analysis shows that only one feature component has a significantly higher contribution than the other two, and is highly similar to typical characteristics of similar faults in the past, then a weight enhancement mechanism is activated. For example, if the actual contributions of the three components in the optimized three-dimensional phase feature vector are 18%, 81%, and 1%, and the contribution of the phase abrupt change index is significantly higher than the other two, and is highly similar to typical characteristics of similar faults in the past, then a weight enhancement mechanism is activated. By enhancing the weight of the core feature, interference is eliminated, and the fault type is ultimately confirmed.

[0075] Finally, if the contribution analysis shows that the contribution distribution of the three feature components is dispersed and does not match the ideal weight distribution for any known fault type, the corresponding identified event is marked as uncertain, and the identification result is pushed to the manual diagnostic terminal. For example, if the contribution of each of the three feature components is between 20% and 40%, and the deviation from the ideal weight distribution for all known fault types is greater than 15%—for instance, if the contribution distribution of the three feature components is: partial discharge (35%), insulation degradation (32%), and poor conductor contact (33%), and does not match the ideal weight distribution for any known fault type—then it is marked as uncertain and pushed to the manual diagnostic terminal for comprehensive judgment based on equipment maintenance records, infrared detection, and other data. This avoids misjudgments of rare faults, complex faults, and strong interference scenarios, improving the reliability of identification in complex scenarios through manual intervention.

[0076] Specifically, the "weight enhancement mechanism" includes: Obtain all samples from the historical database that correspond to the fault type with the highest current matching probability, and construct a reference feature set; Calculate the average similarity between the three components of the optimized three-dimensional phase feature vector—phase consistency coefficient, phase abrupt change index, and phase drift gradient—and the corresponding feature components in the reference feature set; The feature components with the highest similarity to the reference feature set are selected, and the weight ratio of the corresponding feature components in the feature vector is increased according to the similarity percentage. Using the improved weights, the distance between the optimized three-dimensional phase feature vector and the center of the reference feature set for various fault types is recalculated and converted into a new matching probability; Based on the new matching probability, the decision analysis is re-executed. If the fault type corresponding to the highest matching probability remains unchanged, the corresponding fault type identification result is output.

[0077] In this embodiment, all samples of the fault type corresponding to the current highest matching probability in the historical database are first obtained to form a reference feature set. For example, if the fault type corresponding to the current highest matching probability is a partial discharge fault, all samples of partial discharge faults are obtained from the historical database as the reference feature set. The reference feature set can provide a feature benchmark for similar faults for the current feature vector.

[0078] Next, the average similarity between the three components of the optimized 3D phase feature vector—phase consistency coefficient, phase mutation index, and phase drift gradient—and their corresponding feature components in the reference feature set is calculated. For example, the cosine similarity between the three components of the optimized 3D phase feature vector and their corresponding feature components in all reference features in the reference feature set is calculated, and then the mean is calculated to obtain the average similarity between the three components and their corresponding feature components in the reference feature set. For instance, the average similarity of the phase consistency coefficient is 0.85, the average similarity of the phase mutation index is 0.71, and the average similarity of the phase drift gradient is 0.95. The closer the average similarity is to 1, the more consistent the component patterns of the optimized 3D phase feature vector are with the components of historical similar samples.

[0079] Next, the feature components with the highest similarity to the reference feature set are selected, and their weights in the feature vector are increased based on the similarity percentage. For example, the weights of the feature components in the feature vector can be adjusted using preset rules. For instance, if the similarity is ≥90%, the weight is increased by 20%-30%; if the similarity is ≤80% and <90%, the weight is increased by 10%-20%, and the sum of the weights of all feature components remains 1 after adjustment.

[0080] For example, if the weights of the phase consistency coefficient, phase abrupt change index, and phase drift gradient of a partial discharge fault are 0.3, 0.6, and 0.1 respectively, and the phase drift gradient has the highest similarity to the reference feature set (0.95), then according to a preset rule, the weight of the phase drift gradient is increased by 25%, i.e., the new weight of the phase drift gradient = 0.1 × (1 + 25%) = 0.125. At this time, the other two feature components need to be adaptively reduced according to their original weight ratios, with a total reduction of 0.125 - 0.1 = 0.025. The original weight ratio of the phase consistency coefficient and the phase abrupt change index is 0.3:0.6 = 1:2, and the components are allocated according to this ratio. The adjustment range of 0.025: The phase consistency coefficient needs to be reduced by 0.025 × (1 / (1+2)) ≈ 0.0083, so the new weight = 0.3 - 0.0083 ≈ 0.2917; the phase mutation index needs to be reduced by 0.025 × (2 / (1+2)) ≈ 0.0167, so the new weight = 0.6 - 0.0167 ≈ 0.5833. The weight distribution of the three-dimensional features after adjustment is [0.2917, 0.5833, 0.125], which highlights the influence of high similarity features and ensures that the total weight remains unchanged, providing a reasonable feature priority basis for recalculating the matching probability in the future.

[0081] Furthermore, using the improved weights, the distance between the optimized three-dimensional phase feature vector and the reference feature set centers for various fault types is recalculated and converted into a new matching probability. For example, if the improved weights are [0.2917, 0.5833, 0.125], and the optimized three-dimensional phase feature vector is [0.704, 0.767, 0.02], the reference feature set centers for partial discharge faults [0.70, 0.76, 0.21] and insulation degradation faults [0.61, 0.30, 0.86] are extracted from the historical database. The distance to the fault type reference feature set centers is calculated using the weighted Euclidean distance formula: Distance to the reference feature set center for partial discharge faults = ≈0.0674, and the distance from the center of the insulation degradation fault reference feature set = The probability of partial discharge fault is approximately 0.4669. Then, the probability of insulation degradation fault is normalized using the inverse of the distance to convert it into a matching probability: the probability of partial discharge fault is approximately (1 / 0.0674) / (1 / 0.0674+1 / 0.4669)≈87.4%, and the probability of insulation degradation fault is 1-87.4%=12.6%. It can be seen that, based on the enhanced weights, partial discharge fault is still the highest matching item, and the distinction between it and insulation degradation fault is significantly increased. This verifies the reliability of the original fault identification result and further eliminates the possibility of misjudgment caused by random fluctuations in features.

[0082] Finally, based on the new matching probabilities, the decision analysis is re-executed. If the fault type corresponding to the highest matching probability remains unchanged, the corresponding fault type identification result is output; otherwise, if the fault type changes, it is marked as uncertain, and the analysis process is pushed to the human diagnostic terminal. In this way, through secondary verification, the stability of the fault type is ensured, and the risk of random anomalies of a single feature is eliminated.

[0083] In summary, compared to existing technologies, this application performs multi-dimensional decision analysis based on the matching probability and the association weight matrix to identify and output fault type identification results. Thus, on the one hand, by combining the association weight matrix, the actual contribution of feature components is verified to match the ideal weight of the corresponding fault, determining whether a high probability is driven by the core fault feature; on the other hand, differentiated decision strategies are formulated for different scenarios such as prominent contributions from a single feature or dispersed contributions from multiple features. Finally, through multi-dimensional verification combining matching probability and feature contribution, misjudgments based solely on model probability are effectively avoided, ensuring that the output fault type identification results are both accurate and adaptable to different scenarios.

[0084] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this application first acquires the raw electromagnetic signals monitored by the ultra-high frequency sensing unit of the target power equipment, and then preprocesses and demodulates the raw electromagnetic signals to generate initial phase timing data. In this way, initial phase timing data is accurately extracted from the complex electromagnetic environment of the power equipment, providing reliable input data for subsequent fault identification.

[0085] Secondly, this application extracts the phase consistency coefficient, phase abrupt change index, and phase drift gradient of the initial phase time series data within a preset time window to construct an initial three-dimensional phase feature vector. This yields a structured feature vector that can be used for fault classification, providing a reliable data foundation for subsequent dynamic correction and model input.

[0086] Furthermore, this application dynamically corrects the initial three-dimensional phase feature vector based on the preset performance parameters of the phase-locked loop (PLL) and historical data processing errors during the phase demodulation process, generating an optimized three-dimensional phase feature vector. Thus, based on the reference feature vector, the inherent hardware errors of the PLL are adaptively corrected, improving the accuracy of the phase feature vector and better suiting the complex hardware configuration and electromagnetic environment of the field.

[0087] Furthermore, this application inputs the optimized three-dimensional phase feature vector into the pre-trained fault classification model to calculate the matching probability between the output and each typical fault mode. In this way, by strengthening key features through the association weight matrix, the accuracy and precision of fault classification are improved, and a reliable quantitative basis is provided for subsequent multi-dimensional decision analysis.

[0088] Finally, this application performs multidimensional decision analysis based on the matching probability and association weight matrix to identify and output the fault type identification result. Thus, by combining multidimensional verification with matching probability and feature contribution, misjudgments that rely solely on model probability are effectively avoided, ensuring that the output fault type identification result is both accurate and adaptable to different scenarios.

[0089] Through the above technical solutions, this application obtains high-quality phase consistency coefficients, phase mutation indices, and phase drift gradients via preprocessing and phase demodulation, overcoming the limitation that single parameters cannot distinguish fault differences and accurately capturing the differentiated characteristics of different fault types. Based on the preset performance parameters of the phase-locked loop and historical data, the feature vector is dynamically corrected to effectively offset the impact of complex electromagnetic interference and hardware errors on the features. A pre-trained fault classification model is used to calculate the matching probability, and multi-dimensional decision analysis is conducted in conjunction with the correlation weight matrix, avoiding misjudgments based solely on probability and adapting to multi-fault coupling scenarios. Thus, the accuracy, reliability, and precision of power equipment fault identification are improved, while also considering adaptability to complex operating conditions and meeting the needs of refined power grid operation and maintenance.

[0090] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0091] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0096] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A fault identification method based on ultra-high frequency phase characteristic changes, characterized in that, The method includes: The original electromagnetic signal monitored by the ultra-high frequency sensing unit of the target power equipment is acquired, and the original electromagnetic signal is preprocessed and demodulated to generate initial phase time series data. Extract the phase consistency coefficient, phase abruptness index and phase drift gradient of the initial phase time series data within a preset time window to construct an initial three-dimensional phase feature vector; Based on the preset performance parameters of the phase-locked loop and the historical data processing errors during the phase demodulation process, the initial three-dimensional phase feature vector is dynamically corrected to generate an optimized three-dimensional phase feature vector. The optimized three-dimensional phase feature vector is input into the pre-trained fault classification model, and the matching probability of the output with each typical fault mode is calculated. Based on the matching probability and the association weight matrix, a multidimensional decision analysis is performed to identify and output the fault type identification result.

2. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 1, characterized in that, Acquire the raw electromagnetic signal monitored by the ultra-high frequency sensing unit of the target power equipment, preprocess and demodulate the raw electromagnetic signal to generate initial phase time series data, including: The original electromagnetic signals radiated by the target power equipment during operation are collected by the ultra-high frequency sensing unit. The original electromagnetic signal is subjected to noise reduction and signal enhancement processing, and the instantaneous phase value of the signal is demodulated based on phase-locked loop technology; The instantaneous phase values ​​are sampled and recorded at equal time intervals to generate the initial phase time series data.

3. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 2, characterized in that, The original electromagnetic signal undergoes noise reduction and signal enhancement processing, and the instantaneous phase value of the signal is demodulated based on phase-locked loop technology, including: The original electromagnetic signal is subjected to wavelet transform noise reduction processing. A wavelet basis function matching the characteristics of the UHF signal is selected, and environmental noise and white noise interference are suppressed by threshold filtering to obtain the first preprocessed signal. The Wiener filtering algorithm is used to adaptively filter the first preprocessed signal to obtain the second preprocessed signal; The second preprocessed signal is input to the phase-locked loop circuit, wherein the loop bandwidth of the phase-locked loop is preset according to the center frequency of the ultra-high frequency signal and the expected phase jitter range; The instantaneous phase value of the signal is tracked and demodulated in real time by comparing the phase-locked loop voltage-controlled oscillator output signal with the second preprocessed signal.

4. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 3, characterized in that, By comparing the phase of the voltage-controlled oscillator output signal of the phase-locked loop with the phase of the second preprocessed signal, the instantaneous phase value of the signal is tracked and demodulated in real time, including: The second preprocessed signal and the output signal of the voltage-controlled oscillator are input to a phase detector for phase comparison to generate a phase error signal; The phase error signal is filtered by a loop filter, wherein the bandwidth of the loop filter is adaptively adjusted according to the phase noise characteristics of the ultra-high frequency signal. The filtered error signal is used as a control voltage input to the voltage-controlled oscillator to form a closed-loop control circuit. The phase information of the voltage-controlled oscillator output signal is collected in real time as the demodulated instantaneous phase value, wherein the sampling frequency of the instantaneous phase value is not less than twice the center frequency of the ultra-high frequency signal.

5. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 1, characterized in that, Extract the phase consistency coefficient, phase abrupt change index, and phase drift gradient of the initial phase time series data within a preset time window to construct an initial three-dimensional phase feature vector, including: Calculate the reciprocal of the standard deviation of the initial phase time series data within the preset time window, and use the normalized result as the phase consistency coefficient. The maximum absolute value of the first-order difference of the initial phase time series data within the preset time window is calculated as the phase change index; Linear fitting is performed on the initial phase time series data within the preset time window, and the absolute value of the fitting slope is used as the phase drift gradient. The phase consistency coefficient, phase mutation index, and phase drift gradient are combined in a preset order to construct the initial three-dimensional phase feature vector.

6. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 1, characterized in that, Based on the preset performance parameters of the phase-locked loop and historical data processing errors during the phase demodulation process, the initial three-dimensional phase feature vector is dynamically corrected to generate an optimized three-dimensional phase feature vector, including: The phase jitter variance of the phase-locked loop is obtained as a performance indicator of the phase-locked loop; Query the historical database for the baseline feature vector generated by normal signal processing under the same or similar phase-locked loop parameter configurations; Calculate the relative deviation of each component in the initial three-dimensional phase eigenvector and the reference eigenvector; The relative deviation is weighted and corrected according to the phase-locked loop performance index to obtain the corrected phase consistency coefficient, phase mutation index and phase drift gradient, and then recombined to obtain the optimized three-dimensional phase feature vector.

7. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 1, characterized in that, The pre-training steps of the fault classification model include: Collect three-dimensional phase feature vector samples corresponding to multiple sets of UHF signals from historical fault records, and label their actual fault types to form a training sample set; Before supervised training, correlation analysis is performed on the three-dimensional phase feature vector samples in the training sample set. The correlation strength between the phase consistency coefficient, phase mutation index, phase drift gradient and specific fault type in each sample is calculated, and a correlation weight matrix is ​​generated. Using the training sample set, with the three-dimensional phase feature vector weighted according to the correlation weight matrix as input and the fault type as output label, supervised training is performed on the fault classification model built based on the multilayer perceptron. After the model converges, the model parameters and the associated weight matrix are saved to obtain the pre-trained fault classification model.

8. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 1, characterized in that, Based on the matching probability and association weight matrix, a multidimensional decision analysis is performed to identify and output the fault type identification result, including: Obtain the matching probability of each fault type output by the fault classification model and sort them in descending order; Query the association weight matrix of the fault type corresponding to the current highest matching probability, and analyze the contribution of the three components of phase consistency coefficient, phase mutation index and phase drift gradient in the optimized three-dimensional phase feature vector; If the contribution analysis shows that the contribution distribution of the three feature components is consistent with the preset ideal weight distribution of the corresponding fault type in the correlation weight matrix, then the corresponding fault type is directly output as the identification result. If the contribution analysis shows that only one feature component has a significantly higher contribution than the other two, and is highly similar to the typical characteristics of similar failures in the past, then the weight enhancement mechanism will be activated. If the contribution analysis shows that the contribution distribution of the three feature components is scattered and does not match the ideal weight distribution of any known fault type, then the corresponding identification event will be marked as an uncertain state, and the identification result will be pushed to the manual diagnosis terminal.

9. The fault identification method based on ultra-high frequency phase characteristic changes according to claim 8, characterized in that, The weight enhancement mechanism includes: Obtain all samples from the historical database that correspond to the fault type with the highest current matching probability, and construct a reference feature set; Calculate the average similarity between the three components of the optimized three-dimensional phase feature vector—phase consistency coefficient, phase abrupt change index, and phase drift gradient—and the corresponding feature components in the reference feature set; The feature components with the highest similarity to the reference feature set are selected, and the weight ratio of the corresponding feature components in the feature vector is increased according to the similarity percentage. Using the improved weights, the distance between the optimized three-dimensional phase feature vector and the center of the reference feature set for various fault types is recalculated and converted into a new matching probability; Based on the new matching probability, the decision analysis is re-executed. If the fault type corresponding to the highest matching probability remains unchanged, the corresponding fault type identification result is output.

Images