Online assessment method for capacitive device status based on artificial intelligence and frequency domain interpolation

Abstract

The present invention relates to the technical field of power equipment status monitoring, specifically to an online evaluation method for capacitive equipment status based on artificial intelligence and frequency domain interpolation, including obtaining the complex spectrum of the equipment signal and extracting the operating condition characteristics; inputting the operating condition characteristics into the operating condition modulation model to generate dynamic parameters for real-time adjustment of the adversarial interpolation correction model; the correction model converts the spectrum slice into an optimized spectrum, thereby accurately calculating the dielectric loss factor. Afterwards, the factor is combined with other operating parameters into a multidimensional time series, input into a health status assessment model, outputs an abnormality score that characterizes the actual insulation degradation state of the equipment, and generates a diagnostic data object containing a measurement result consistency score. By introducing an artificial intelligence model, the present invention dynamically compensates for measurement errors according to real-time operating conditions and performs intelligent diagnosis, significantly improving the accuracy and reliability of online evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment status monitoring, and in particular to an online evaluation method for capacitive equipment status based on artificial intelligence and frequency domain interpolation. Background Art

[0002] Capacitive devices, such as capacitors, bushings, and cables, are critical components of power systems, and their insulation condition is directly related to the safe and stable operation of the entire power grid. Dielectric loss factor is a key metric for evaluating the insulation performance of these devices and determining their degree of aging and defects. Therefore, accurate, real-time online monitoring of this parameter is crucial for equipment condition-based maintenance and fault warning.

[0003] Currently, the mainstream method for online calculation of the dielectric loss factor is to collect the device's voltage and current signals, obtain the spectrum through a fast Fourier transform (FFT), and then use a frequency domain interpolation algorithm to accurately determine the amplitude and phase of the fundamental component to calculate the dielectric loss factor. However, this traditional method faces severe challenges in practical application. On-site power grid environments are often very complex, with a large amount of harmonic interference, background noise, and signal fluctuations. These interferences can seriously affect the accuracy of FFT spectrum analysis, causing traditional frequency domain interpolation algorithms to generate inherent calculation errors. These factors are unable to effectively filter out or compensate for these unfavorable factors, resulting in low accuracy and poor reliability of the calculated dielectric loss factor measurements, making it difficult to truly reflect the insulation degradation status of the equipment.

[0004] Therefore, an online assessment method for capacitive device status based on artificial intelligence and frequency domain interpolation is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide an online evaluation method for capacitive device status based on artificial intelligence and frequency domain interpolation, which compensates the inherent calculation error of the frequency domain interpolation algorithm through the artificial intelligence model to achieve accurate status evaluation.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The online assessment method of capacitive device status based on artificial intelligence and frequency domain interpolation includes:

[0008] Acquiring voltage and current signals of a capacitive device, performing a fast Fourier transform to obtain a complex spectrum, and performing real-time analysis on the complex spectrum to obtain operating condition characteristics;

[0009] Inputting the operating condition characteristics into an operating condition modulation model to generate a set of dynamic modulation parameters;

[0010] Inputting the complex spectrum slices into an adversarial interpolation correction model, and using the dynamic modulation parameters to adjust the internal data processing flow of the adversarial interpolation correction model in real time, converting the input spectrum slices into optimized spectra, and then inputting the optimized spectrum into a frequency domain interpolation algorithm to calculate a dielectric loss factor measurement value; combining the dielectric loss factor measurement value with the equipment operating parameters into a multidimensional time series;

[0011] The multidimensional time series is input into a health status assessment model; the health status assessment model is used to output an anomaly score to determine the true insulation degradation state of the capacitive device; a diagnostic data object is generated and output, and the diagnostic data object is used to encapsulate the state parameters of the capacitive device and the measurement result consistency score; the measurement result consistency score is a numerical value output by the interpolation posterior accuracy discriminator of the adversarial interpolation correction model, which is used to quantify the degree of consistency between the dielectric loss factor measurement value obtained after correction and calculation and the statistical distribution characteristics of the true health parameters in the training data set.

[0012] Preferably, the operating condition modulation model includes a feature extraction unit, an operating condition encoding unit and a parameter generation unit;

[0013] The feature extraction unit is used to calculate and quantify at least one preset operating condition characteristic index from the complex spectrum, wherein the operating condition characteristic index includes a total harmonic distortion rate, a signal-to-noise ratio, and an amplitude of a specified subharmonic component;

[0014] The working condition encoding unit is used to perform nonlinear mapping and fusion on the working condition characteristic indicators to generate a low-dimensional potential vector that can represent the current comprehensive working condition;

[0015] The parameter generation unit is used to decode the low-dimensional latent vector into n groups of dynamic modulation parameters, where the dynamic modulation parameters include a scaling factor and an offset factor.

[0016] Preferably, the adversarial interpolation correction model includes a frequency domain interpolation feedforward corrector and an interpolation a posteriori accuracy discriminator;

[0017] The frequency domain interpolation feedforward corrector is used to receive slices of the complex spectrum and convert them into an optimized spectrum;

[0018] During the adversarial training stage, the interpolation posterior accuracy discriminator is used to judge the accuracy of the parameters finally calculated after processing by the frequency domain interpolation feedforward corrector and the frequency domain interpolation algorithm, and thereby drive the frequency domain interpolation feedforward corrector to learn to compensate for the inherent calculation error of the frequency domain interpolation algorithm.

[0019] Preferably, the health status assessment model includes a time series feature encoding unit, a health pattern reconstruction unit and an anomaly score calculation unit;

[0020] The temporal feature encoding unit is used to compress and encode the input multidimensional time series into a low-dimensional latent vector that can represent its dynamic characteristics;

[0021] The health pattern reconstruction unit is used to decode and reconstruct the original multidimensional time series from the low-dimensional latent vector, wherein the health status assessment model is trained via health status data so that the health pattern reconstruction unit has a reconstruction error below a preset error threshold for the data pattern of the health condition;

[0022] The anomaly score calculation unit is used to calculate and quantify the difference between the input multidimensional time series and the time series reconstructed by the health model reconstruction unit, and use the difference as an anomaly score representing the actual insulation degradation state of the equipment.

[0023] Preferably, the step of performing real-time analysis on the complex spectrum to obtain the operating condition characteristics specifically includes:

[0024] In the complex spectrum, identifying and locating a fundamental wave spectrum line and a plurality of preset harmonic wave spectrum lines, and calculating the amplitude of the fundamental wave and the amplitude of the harmonic wave respectively;

[0025] The harmonic amplitudes of the plurality of harmonic spectral lines are squared and summed, and the square root of the square sum is taken to obtain a total harmonic equivalent amplitude; the total harmonic equivalent amplitude is compared with the fundamental wave amplitude to calculate the total harmonic distortion rate; the energy of the fundamental wave spectral line is compared with the average energy within a predetermined noise frequency band to calculate the signal-to-noise ratio.

[0026] Preferably, the real-time adjustment process is specifically as follows: by performing an affine transformation on the feature map output by the intermediate layer of the adversarial interpolation correction model, using the scaling factor and offset factor to perform element-by-element multiplication and addition operations on the feature map, respectively, to dynamically change the calculation behavior and feature response.

[0027] Preferably, the diagnostic data objects include state parameters, operating condition characteristics, dielectric loss factor measurement values ​​and measurement result consistency scores of the capacitive device.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. By introducing an adversarial interpolation correction model, this invention proactively learns and compensates for the inherent computational errors of traditional frequency-domain interpolation algorithms under complex operating conditions. Furthermore, a working condition modulation model analyzes real-time working condition characteristics such as total harmonic distortion and signal-to-noise ratio, and dynamically adjusts the correction model's behavior to effectively overcome strong interference from field harmonics and noise. Compared to traditional methods, this significantly improves the accuracy and robustness of online dielectric loss factor measurements.

[0030] 2. This invention goes beyond single parameter measurement and instead integrates dielectric loss factor with other operating parameters into a multidimensional time series, utilizing a health assessment model for in-depth analysis. This model automatically learns the healthy operating patterns of the equipment and quantifies the output of anomaly scores to characterize the true state of insulation degradation. This enables the transition from numerical measurement to intelligent diagnosis, providing a reliable basis for predicting equipment status trends and providing early warning of faults.

[0031] 3. This invention proposes and generates a measurement result consistency score. This score, output by the interpolation posterior accuracy discriminator in the adversarial interpolation correction model, quantifies the degree of similarity between the measurement result and the healthy sample in terms of data pattern, providing an objective reference indicator for the stability of the measurement process. This provides key evidence for operators to determine whether the diagnosis conclusion is based on a reliable measurement process, effectively avoiding misjudgments caused by data quality issues, and making the output of the entire evaluation system more transparent and trustworthy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of a method for online assessment of capacitive device status based on artificial intelligence and frequency domain interpolation proposed in an embodiment of the present invention;

[0033] Figure 2 This is a flow chart of the adversarial interpolation correction model proposed in an embodiment of the present invention;

[0034] Figure 3 Schematic diagram of the structure of the health status assessment model in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] Example 1

[0037] See also Figures 1 to 2 The present invention provides an online evaluation method for capacitive device status based on artificial intelligence and frequency domain interpolation. The technical solution is as follows:

[0038] Online assessment method of capacitive device status based on artificial intelligence and frequency domain interpolation, such as Figure 1 Shown, including:

[0039] Acquiring voltage and current signals of a capacitive device, performing a fast Fourier transform to obtain a complex spectrum, and performing real-time analysis on the complex spectrum to obtain operating condition characteristics;

[0040] Inputting the operating condition characteristics into an operating condition modulation model to generate a set of dynamic modulation parameters;

[0041] Inputting the complex spectrum slices into an adversarial interpolation correction model, and using the dynamic modulation parameters to adjust the internal data processing flow of the adversarial interpolation correction model in real time, converting the input spectrum slices into optimized spectra, and then inputting the optimized spectrum into a frequency domain interpolation algorithm to calculate a dielectric loss factor measurement value; combining the dielectric loss factor measurement value with the equipment operating parameters into a multidimensional time series;

[0042] The multidimensional time series is input into a health status assessment model; the health status assessment model is used to output an anomaly score to determine the true insulation degradation state of the capacitive device; a diagnostic data object is generated and output, and the diagnostic data object is used to encapsulate the state parameters of the capacitive device and the measurement result consistency score; the measurement result consistency score is a numerical value output by the interpolation posterior accuracy discriminator of the adversarial interpolation correction model, which is used to quantify the degree of consistency between the dielectric loss factor measurement value obtained after correction and calculation and the statistical distribution characteristics of the true health parameters in the training data set.

[0043] Furthermore, the operating condition modulation model includes a feature extraction unit, an operating condition encoding unit and a parameter generation unit;

[0044] The feature extraction unit is used to calculate and quantify at least one preset operating condition characteristic index from the complex spectrum, wherein the operating condition characteristic index includes a total harmonic distortion rate, a signal-to-noise ratio, and an amplitude of a specified subharmonic component;

[0045] The working condition encoding unit is used to perform nonlinear mapping and fusion on the working condition characteristic indicators to generate a low-dimensional potential vector that can represent the current comprehensive working condition;

[0046] The parameter generation unit is used to decode the low-dimensional latent vector into n groups of dynamic modulation parameters, where the dynamic modulation parameters include a scaling factor and an offset factor.

[0047] The operating condition encoding unit consists of a multilayer perceptron (MLP) with three fully connected layers, each with 64, 32, and 16 neurons, respectively, using ReLU as the activation function. The parameter generation unit consists of a linear layer that maps the 16-dimensional latent vector to scaling and offset factors.

[0048] The parameter generation unit is used to decode the low-dimensional latent vector into four sets of dynamic modulation parameters, each set including a scaling factor and an offset factor. The four sets of parameters correspond to four skip connections in the frequency domain interpolation feedforward corrector U-Net structure.

[0049] By establishing an operating condition modulation model, this invention achieves deep perception and dynamic adaptation to complex power grid environments. This model accurately quantifies multiple key operating condition indicators, such as total harmonic distortion (THD) and signal-to-noise ratio (SNR), and intelligently integrates these indicators into a low-dimensional latent vector that represents the current comprehensive operating condition. Ultimately, this vector is decoded into precise dynamic modulation parameters (scaling and offset factors) to accurately guide the internal data processing flow of subsequent AI models in real time. This enables the entire evaluation system to possess exceptional environmental adaptability, ensuring high accuracy and robustness even in a variety of harsh operating conditions.

[0050] Furthermore, the slicing of the complex spectrum specifically refers to: after performing a fast Fourier transform (FFT), taking the identified fundamental frequency point as the center, cutting N frequency points on the left and right sides thereof to form a complex vector with a total length of (2N+1) as input.

[0051] If the FFT point count is 4096 and the fundamental frequency is 50 Hz, the fundamental peak point and N = 255 frequency points to its left and right can be selected to form a spectrum slice with a length of 511. This width is sufficient to cover the main lobe and major sidelobe information of the fundamental, providing sufficient interpolation basis for the correction model.

[0052] Furthermore, if Figure 2 As shown, it is specifically a flow chart of an adversarial interpolation correction model, wherein the adversarial interpolation correction model includes a frequency domain interpolation feedforward corrector and an interpolation posterior accuracy discriminator;

[0053] The frequency domain interpolation feedforward corrector is used to receive slices of the complex spectrum and convert them into an optimized spectrum;

[0054] During the adversarial training stage, the interpolation posterior accuracy discriminator is used to judge the accuracy of the parameters finally calculated after processing by the frequency domain interpolation feedforward corrector and the frequency domain interpolation algorithm, and thereby drive the frequency domain interpolation feedforward corrector to learn to compensate for the inherent calculation error of the frequency domain interpolation algorithm.

[0055] The frequency domain interpolation feedforward corrector can adopt a U-Net network structure based on one-dimensional convolution, which includes 4 downsampling modules and 4 upsampling modules to effectively capture the multi-scale characteristics of the spectrum.

[0056] Specifically, each module in the downsampling module consists of two consecutive one-dimensional convolutional layers with a kernel size of 3, a stride of 1, and "same" padding, followed by a ReLU activation function and a batch normalization layer. The number of output channels for each module is [64, 128, 256, 512], respectively. After each module, a max pooling layer with a stride of 2 is used for downsampling.

[0057] Each module in the upsampling module first performs upsampling using a transposed convolution with a stride of 2, reducing the number of channels by half. This upsampling is then concatenated with the feature map from the corresponding encoder layer (passed via a skip connection). The concatenated feature map then passes through two convolutional layers, similar to those in the encoder, followed by a ReLU activation function and a batch normalization layer.

[0058] The last layer of the network is a convolutional layer with a convolution kernel size of 1, which is used to map the number of channels of the feature map back to the number of channels of the input spectrum slice.

[0059] The input of the interpolation a posteriori accuracy discriminator is a two-dimensional vector formed by concatenating the dielectric loss factor measurement value calculated by the frequency domain interpolation algorithm and the dielectric loss factor 'true value' corresponding to the sample in the training data set.

[0060] The interpolation posterior accuracy discriminator consists of a three-layer fully connected network. The input layer receives a two-dimensional vector, the number of neurons in each hidden layer is 128 and 64, respectively, and the output layer is a single neuron. It uses a sigmoid function to output a scalar value between 0 and 1, representing the probability that the input parameter is "true." The hidden layer uses LeakyReLU as the activation function.

[0061] The adversarial training process in this example uses the WGAN-GP loss function to enhance training stability. Training uses the Adam optimizer, with the learning rate for both the generator and the discriminator set to 0.0001. The training dataset contains 10,000 simulated spectrum samples covering various signal-to-noise ratio (SNR) and total harmonic distortion (THD) conditions. The batch size is 64, and training is repeated for 200 epochs.

[0062] The adversarial interpolation correction model designed in this paper achieves deep optimization over traditional algorithms through an adversarial game between its internal feedforward corrector and a posteriori accuracy discriminator. During the training phase, the discriminator focuses on the accuracy of the final calculated parameters and uses this to reverse-drive the corrector, enabling it to learn how to accurately compensate for the inherent computational errors of the subsequent frequency domain interpolation algorithm. This unique training mechanism forces the corrector to generate highly optimized spectra, fundamentally improving the performance of traditional algorithms and ultimately achieving measurement accuracy far exceeding that of conventional correction methods.

[0063] Furthermore, the frequency domain interpolation algorithm is a three-point parabola interpolation algorithm, which accurately estimates the frequency and amplitude of the peak by finding the fundamental wave peak and its two adjacent points in the FFT spectrum and using parabola fitting.

[0064] Furthermore, the device operating parameters include at least one or a combination of the following: device housing temperature, ambient temperature and humidity, the root mean square value of the grid voltage, and the root mean square value of the current flowing through the device. These parameters are collected at the same time as the dielectric loss factor measurement value to form a multidimensional vector.

[0065] Before combining into a multidimensional vector, all parameters are Z-score normalized. For each parameter's time series, its mean is subtracted and divided by its standard deviation, ensuring that all feature dimensions follow a standard normal distribution with a mean of 0 and a variance of 1. After normalization, the processed dielectric loss factor measurements, device housing temperature, ambient temperature, ambient humidity, RMS voltage, and RMS current are concatenated in this order to form a 6-dimensional input vector, forming a multidimensional time series for input into the health status assessment model.

[0066] By integrating multiple operating parameters such as dielectric loss factor, temperature, and voltage to build a multi-dimensional evaluation perspective, the AI ​​model can capture the complex correlations between parameters, go beyond single threshold judgment, and significantly improve the comprehensiveness and accuracy of diagnosis.

[0067] Furthermore, if Figure 3 As shown, the health status assessment model includes a time series feature encoding unit, a health pattern reconstruction unit and an abnormality score calculation unit;

[0068] The temporal feature encoding unit is used to compress and encode the input multidimensional time series into a low-dimensional latent vector that can represent its dynamic characteristics;

[0069] The health pattern reconstruction unit is used to decode and reconstruct the original multidimensional time series from the low-dimensional latent vector, wherein the health status assessment model is trained via health status data so that the health pattern reconstruction unit has a reconstruction error below a preset error threshold for the data pattern of the health condition;

[0070] The anomaly score calculation unit is used to calculate and quantify the difference between the input multidimensional time series and the time series reconstructed by the health model reconstruction unit, and use the difference as an anomaly score representing the actual insulation degradation state of the equipment.

[0071] Specifically, the health status assessment model is constructed as follows:

[0072] The input multidimensional time series is segmented into a fixed time window (e.g., a time step of 100).

[0073] The temporal feature encoding unit consists of two stacked LSTM layers, each with 128 hidden units. The encoder only returns the hidden state and cell state of the last time step as the context vector.

[0074] The healthy mode reconstruction unit consists of another two stacked LSTM layers, also with 128 hidden units. The decoder receives the encoder's context vector as its initial state and reconstructs the original input in a sequential manner. To assist in reconstruction, each decoder time step receives a fixed starting symbol or the output of the previous time step as input.

[0075] After each LSTM layer, a Dropout layer is set with a dropout rate of 0.2 to prevent the model from overfitting.

[0076] The health status assessment model uses mean squared error (MSE) as the reconstruction loss function and is trained using the Adam optimizer with a learning rate of 0.001. The training data consists of 1000 hours of continuous operating data collected from health devices, sampled at a frequency of 10 kHz, and Z-score normalized.

[0077] The health status data is collected under various typical loads and ambient temperatures within the first 1,000 operating hours of the equipment, or when offline dielectric loss testing confirms a dielectric loss factor reading of less than 0.3% at rated voltage. The collected data set should cover the range of normal operating conditions the equipment may encounter to ensure that the health patterns learned by the model have sufficient generalizability.

[0078] The anomaly score calculation unit is used to calculate the mean square error (MSE) between the input multidimensional time series and the reconstructed time series, and use the mean square error value as the anomaly score representing the actual insulation degradation state of the equipment.

[0079] Through an encoding-reconstruction architecture, intelligent and sensitive device status diagnosis is achieved. This method comprehensively evaluates multi-dimensional time series data without pre-setting complex thresholds, accurately capturing subtle precursors to faults.

[0080] Furthermore, the health status assessment model also includes an online model update unit to address the "concept drift" problem caused by natural device aging or long-term, slow changes in the operating environment. The online model update unit is configured to periodically or upon detecting that the system confidence meets preset conditions, incrementally fine-tune the parameters of the time series feature encoding unit and the health pattern reconstruction unit using newly collected high-confidence health data, thereby dynamically updating the baseline model of the device health status.

[0081] Furthermore, the selection criteria for high-confidence health data are strictly defined as follows: at the time of collection, the "anomaly score" output by the health status assessment model must be below a preset safety threshold (e.g., 0.1), and the "measurement consistency score" output by the adversarial interpolation correction model must be above a preset reliability threshold (e.g., 0.95). This dual standard ensures that only data that the system itself confirms as "healthy" and "reliable" is used to update the health baseline, thereby ensuring the robustness and reliability of the model update process and forming a self-consistent closed-loop optimization system.

[0082] This invention effectively overcomes concept drift caused by device aging through an online model update unit. Its dual data screening mechanism, characterized by low anomaly scores and high confidence, ensures that only the most reliable data is used for model self-optimization, forming a secure closed loop. This enables the dynamic evolution of the diagnostic baseline, ensuring that the system maintains assessment accuracy over the long term.

[0083] Furthermore, the measurement result consistency score is a numerical value output by the interpolation posterior accuracy discriminator of the adversarial interpolation correction model, which is used to quantify the degree of consistency between the dielectric loss factor measurement value obtained after correction and calculation and the statistical distribution characteristics of the true health parameters in the training data set.

[0084] To generate the measurement result consistency score, we first obtain the original output score s of the interpolation posterior accuracy discriminator for the optimized spectrum. Then, we map this score to the interval [0, 1] using the Sigmoid function to obtain the quantitative measurement result consistency score.

[0085] This invention significantly improves the transparency and reliability of the assessment system by establishing a measurement result consistency score. This score, derived from the "interpolation posterior accuracy discriminator" within the adversarial interpolation correction model, assesses in real time the degree of similarity between the final calculated result and a large number of healthy samples in terms of data pattern, empowering the measurement process with self-review capabilities. This allows users to obtain not only the measurement value but also a quantitative reference indicator of data quality. This indicator helps identify and alert operators to such high-risk data, preventing misjudgments due to data quality issues and providing a solid basis for accurate maintenance decisions.

[0086] Furthermore, the step of performing real-time analysis on the complex spectrum to obtain operating condition characteristics specifically includes:

[0087] In the complex spectrum, identifying and locating a fundamental wave spectrum line and a plurality of preset harmonic wave spectrum lines, and calculating the amplitude of the fundamental wave and the amplitude of the harmonic wave respectively;

[0088] The harmonic amplitudes of the plurality of harmonic spectral lines are squared and summed, and the square root of the square sum is taken to obtain a total harmonic equivalent amplitude; the total harmonic equivalent amplitude is compared with the fundamental wave amplitude to calculate the total harmonic distortion rate; the energy of the fundamental wave spectral line is compared with the average energy within a predetermined noise frequency band to calculate the signal-to-noise ratio.

[0089] The predetermined noise frequency band is dynamically determined according to the fundamental frequency f0. For example, a frequency band from (2.5*f0) to (3.5*f0) where strong harmonics usually do not exist can be selected as the noise estimation area.

[0090] This invention specifically defines a calculation method for operating condition characteristics, ensuring the objectivity and effectiveness of feature extraction. By using standard methods to accurately calculate total harmonic distortion and signal-to-noise ratio, this method provides the most critical, standardized environmental quantitative indicators for subsequent AI models. This clear feature extraction method enables AI to understand the harmonic and noise interference level of the current power grid based on reliable data, laying a solid foundation for its precise and targeted dynamic correction, thereby significantly improving the stability and ultimate accuracy of the entire evaluation system.

[0091] Furthermore, the real-time adjustment process is specifically as follows: by performing an affine transformation on the feature map output by the intermediate layer of the adversarial interpolation correction model, using the scaling factor and offset factor to perform element-by-element multiplication and addition operations on the feature map, respectively, to dynamically change the calculation behavior and feature response.

[0092] Specifically, in the U-Net structure of the frequency-domain interpolation feedforward corrector, the scaling and offset factors are applied to the feature maps transmitted by the skip connection between the encoder (downsampling path) and the decoder (upsampling path). Before concatenating the feature map of a particular encoder layer with the upsampled output of the corresponding decoder layer, the affine transformation is performed on the feature map transmitted by the encoder. This allows the operating condition information to modulate the detailed features of the spectrum at multiple scales, thereby achieving precise adaptive correction.

[0093] The present invention achieves refined and efficient control of AI computing behavior by performing affine transformations on the model's intermediate layer feature maps. By utilizing the scaling and offset factors generated by the operating condition modulation model, the model's internal characteristic response can be directly and dynamically changed through element-by-element multiplication and addition. This sophisticated adjustment mechanism allows external operating condition information to deeply intervene in the model's computational process, allowing a single model to flexibly adapt to a changing power grid environment, greatly improving the model's expressive power and the accuracy of the final correction effect.

[0094] Furthermore, the diagnostic data object includes state parameters, operating condition characteristics, dielectric loss factor measurement values ​​and measurement result consistency scores of the capacitive device.

[0095] This diagnostic data object packages the final evaluation conclusion with key measurement values ​​and operating condition background, providing users with a complete diagnostic evidence chain, greatly facilitating in-depth tracing and comprehensive analysis, and enhancing the transparency and credibility of the evaluation results.

[0096] The online assessment method proposed in this paper utilizes an adversarial correction model dynamically modulated by real-time operating conditions to accurately compensate for errors in traditional algorithms, significantly improving the accuracy and anti-interference capabilities of dielectric loss factor calculations. A health assessment model analyzes the multidimensional time series data containing these measurements, outputting anomaly scores that represent the true degradation state, enabling automated intelligent diagnosis. The system outputs a consistency score for measurement results, providing a quantitative basis for users to assess the reliability of the results, comprehensively enhancing the accuracy, intelligence, and credibility of the assessment.

[0097] Example 2

[0098] This example uses a 220kV main transformer high-voltage side oil-impregnated paper bushing installed in a city substation as an example, combining specific application scenarios and data to further illustrate the application process and practical effects of the present method. Because the substation is located near a heavy industrial area, the power grid background harmonic interference is severe, posing a significant challenge to traditional online dielectric loss factor monitoring.

[0099] The online assessment system described in this paper was deployed on a 220 kV main transformer bushing (equipment number: TG-202501-A) that had recently undergone offline testing (test results indicated a dielectric loss factor of 0.28%, indicating excellent condition). During the first month after deployment, the system primarily collected data to train and establish a health assessment model for the equipment. During this period, the system recorded operating parameters under various loads and environments and confirmed that its anomaly score remained consistently low (less than 0.2) and its measurement consistency score remained stable at a high level (greater than 0.95).

[0100] Due to increased production capacity at a nearby factory, the grid's fifth and seventh harmonic distortion increased significantly, with total harmonic distortion (THD) rising from its usual 3% fluctuation to 6%-8%. Simultaneously, due to a minor seal defect in the bushing, trace amounts of moisture began to slowly infiltrate the insulation, causing the dielectric loss factor to gradually deteriorate.

[0101] The operating modulation model of the present invention monitors deterioration in total harmonic distortion and signal-to-noise ratio in real time. The model rapidly generates a new set of dynamic modulation parameters and passes them to the adversarial interpolation correction model. The correction model uses these parameters to adjust its internal processing of spectrum slices in real time, effectively compensating for the interference of strong harmonics on subsequent frequency-domain interpolation calculations. As a result, the system can still output accurate and stable dielectric loss factor measurements, avoiding the drastic data jumps and false alarms caused by harmonic interference in traditional methods.

[0102] The accurately measured dielectric loss factor showed a steady, slow growth trend from 0.30% to 0.45%. This multidimensional time series was input into the established health status assessment model. Because the model learns the data pattern of the equipment in a healthy state, when the input actual operating data deviates from the learned health pattern, the model's health pattern reconstruction unit produces significant reconstruction errors. The anomaly score calculation unit quantifies this error as a continuously rising anomaly score.

[0103] Table 1 shows some key data monitored by the system in the last week:

[0104] Table 1: Equipment TG-202501-A condition monitoring data fragment

[0105]

[0106] When the "anomaly score" reaches the preset alarm threshold (for example, 0.85), the system automatically triggers a high-level warning and generates a packaged diagnostic data object, which is pushed to the operation and maintenance management platform.

[0107] After receiving the diagnostic report, the O&M team noted that despite the challenging operating conditions (THD as high as 8.1%), the measurement consistency score remained high (0.96). This demonstrated that the proposed correction model successfully compensated for the strong harmonic interference, and the outputted dielectric loss factor measurement (0.45%) was a stable, well-formed value, rather than a random jump or error caused by interference. Therefore, the O&M team was confident that the high anomaly score was driven by a stable and reliable measurement of degraded dielectric loss factor, thus determining that the bushing was at risk of real internal insulation degradation and deciding to schedule a planned power outage for offline verification.

[0108] The method of the present invention can achieve accurate measurement of the dielectric loss factor of capacitive equipment and intelligent diagnosis of its status in complex field environments with strong interference such as harmonics. The consistency score of the measurement results provides strong evidence for the reliability of the diagnostic conclusion, providing strong technical support for the realization of condition-based maintenance and fault warning.

[0109] Example 3

[0110] This example uses a 35kV collector cable in a large wind farm as an example to illustrate the application of the present invention's method in renewable energy power generation and transmission and distribution systems. Wind farm operating conditions are characterized by significant volatility and randomness, and the wind turbine inverters generate complex harmonics, placing extremely high demands on online monitoring of cable insulation conditions.

[0111] In March 2024, the online assessment system described in this invention was deployed on a key collection line within the wind farm (line number: WL-C04). This line aggregates the power output of eight 2MW wind turbines and transmits it to a step-up substation via underground XLPE (cross-linked polyethylene) cable. The line is 3 kilometers long and includes four intermediate joints, with the cable intermediate joints being a potential weak link.

[0112] During the first three months after deployment, the system experienced a variety of wind conditions, ranging from low-power operation in light winds to high-power output under full load and even overspeed control. During this period, the "Health Assessment Model" continuously learned the multi-dimensional data characteristics of the cable line under various dynamic operating conditions, thereby establishing a dynamic health baseline that can adapt to large power fluctuations. During this period, despite significant changes in operating conditions (such as output power and harmonic content), the system's anomaly score remained stable and safe below 0.2.

[0113] In September 2024, due to abundant autumn wind resources, the line was operating under high load for an extended period. Due to a minor flaw in construction, the insulation performance of one of the cable joints, numbered WL-C04-J3, began to slowly deteriorate under the repeated thermal and electrical stresses generated by high-load operation.

[0114] At high power output, the high-order harmonics generated by the turbine inverter increase significantly; the operating condition modulation model of the present invention captures this change and adjusts the internal parameters of the adversarial interpolation correction model in real time; the system monitors that the dielectric loss factor measurement value of the line shows a new strong positive correlation with the output power that has never appeared in the healthy baseline.

[0115] Table 2 shows the data continuously monitored by the system in mid-September 2024, reflecting the degradation process.

[0116] Table 2: Line WL-C04 status monitoring data fragment

[0117]

[0118] When the anomaly score first exceeded the alarm threshold of 0.85 under high load, the system immediately sent an alarm to the wind farm control center, attaching a diagnostic data object that clearly indicated that the line had signs of load-related insulation degradation.

[0119] While analyzing the diagnostic report, the operations team discovered two key points: First, the system's anomaly score showed a strong positive correlation with the line's average output power, exceeding the alarm threshold only under high load. Second, throughout the entire process, regardless of whether the load was high or low, the measurement consistency score remained at an extremely high level, above 0.97.

[0120] Based on this diagnostic report, the operations and maintenance team avoided blindly inspecting the entire 3-kilometer cable. Instead, they used a cable fault locator to conduct a targeted offline inspection of the line during a low wind speed window. The results quickly pinpointed the fault to the intermediate joint WL-C04-J3. On-site excavation revealed signs of partial discharge and damp insulation at the joint, verifying the accuracy of the online assessment system.

[0121] Through timely warnings and precise diagnostic guidance, the operations team successfully performed preventative maintenance, avoiding the complete shutdown of the entire collection line and significant power generation losses caused by a complete breakdown of the cable connector. This example demonstrates that the present invention can effectively address the complex operating conditions of new energy power generation systems and achieve intelligent and reliable diagnosis of early-stage, subtle faults in critical electrical equipment.

[0122] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for online assessment of capacitive device status based on artificial intelligence and frequency domain interpolation, comprising: Acquiring voltage and current signals of a capacitive device, performing a fast Fourier transform to obtain a complex spectrum, and performing real-time analysis on the complex spectrum to obtain operating condition characteristics; Inputting the operating condition characteristics into an operating condition modulation model to generate a set of dynamic modulation parameters; Inputting the complex spectrum slices into an adversarial interpolation correction model, and using the dynamic modulation parameters to adjust the internal data processing flow of the adversarial interpolation correction model in real time, converting the input spectrum slices into optimized spectrum, and then inputting the optimized spectrum into a frequency domain interpolation algorithm to calculate the dielectric loss factor measurement value; Combining the dielectric loss factor measurement value with the equipment operating parameters into a multi-dimensional time series; Inputting the multidimensional time series into a health status assessment model; and using the abnormality score output by the health status assessment model to determine the actual insulation degradation state of the capacitive device; Generate and output a diagnostic data object, the diagnostic data object being used to encapsulate the state parameters of the capacitive device and a measurement result consistency score; the measurement result consistency score is a numerical value output by the interpolation posterior accuracy discriminator of the adversarial interpolation correction model, and is used to quantify the degree of consistency between the dielectric loss factor measurement value obtained after correction and calculation and the statistical distribution characteristics of the true health parameters in the training data set.

2. The method for online assessment of capacitive device status based on artificial intelligence and frequency domain interpolation according to claim 1 is characterized in that: The working condition modulation model includes a feature extraction unit, a working condition encoding unit and a parameter generation unit; The feature extraction unit is used to calculate and quantify at least one preset operating condition characteristic index from the complex spectrum, wherein the operating condition characteristic index includes a total harmonic distortion rate, a signal-to-noise ratio, and an amplitude of a specified subharmonic component; The working condition encoding unit is used to perform nonlinear mapping and fusion on the working condition characteristic indicators to generate a low-dimensional potential vector that can represent the current comprehensive working condition; The parameter generation unit is used to decode the low-dimensional latent vector into n groups of dynamic modulation parameters, where the dynamic modulation parameters include a scaling factor and an offset factor.

3. The method for online assessment of capacitive device status based on artificial intelligence and frequency domain interpolation according to claim 1 is characterized in that: The adversarial interpolation correction model includes a frequency domain interpolation feedforward corrector and an interpolation posterior accuracy discriminator; The frequency domain interpolation feedforward corrector is used to receive slices of the complex spectrum and convert them into an optimized spectrum; During the adversarial training stage, the interpolation posterior accuracy discriminator is used to judge the accuracy of the parameters finally calculated after processing by the frequency domain interpolation feedforward corrector and the frequency domain interpolation algorithm, and thereby drive the frequency domain interpolation feedforward corrector to learn to compensate for the inherent calculation error of the frequency domain interpolation algorithm.

4. The method for online assessment of capacitive device status based on artificial intelligence and frequency domain interpolation according to claim 1 is characterized in that: The health status assessment model includes a time series feature encoding unit, a health pattern reconstruction unit and an anomaly score calculation unit; The temporal feature encoding unit is used to compress and encode the input multidimensional time series into a low-dimensional latent vector that can represent its dynamic characteristics; The health pattern reconstruction unit is used to decode and reconstruct the original multidimensional time series from the low-dimensional latent vector, wherein the health status assessment model is trained via health status data so that the health pattern reconstruction unit has a reconstruction error below a preset error threshold for the data pattern of the health condition; The anomaly score calculation unit is used to calculate and quantify the difference between the input multidimensional time series and the time series reconstructed by the health model reconstruction unit, and use the difference as an anomaly score representing the actual insulation degradation state of the equipment.

5. The online evaluation method for capacitive device status based on artificial intelligence and frequency domain interpolation according to claim 1 is characterized in that: The step of performing real-time analysis on the complex spectrum to obtain operating condition characteristics specifically includes: In the complex spectrum, identifying and locating a fundamental wave spectrum line and a plurality of preset harmonic wave spectrum lines, and calculating the amplitude of the fundamental wave and the amplitude of the harmonic wave respectively; The harmonic amplitudes of the plurality of harmonic spectral lines are squared and summed, and the square root of the square sum is taken to obtain a total harmonic equivalent amplitude; the total harmonic equivalent amplitude is compared with the fundamental wave amplitude to calculate the total harmonic distortion rate; the energy of the fundamental wave spectral line is compared with the average energy within a predetermined noise frequency band to calculate the signal-to-noise ratio.

6. The method for online assessment of capacitive device status based on artificial intelligence and frequency domain interpolation according to claim 2, characterized in that: The real-time adjustment process is specifically as follows: by performing an affine transformation on the feature map output by the intermediate layer of the adversarial interpolation correction model, using the scaling factor and offset factor to perform element-by-element multiplication and addition operations on the feature map, the calculation behavior and feature response are dynamically changed.

7. The method for online assessment of capacitive device status based on artificial intelligence and frequency domain interpolation according to claim 1, characterized in that: The diagnostic data objects include state parameters, operating condition characteristics, dielectric loss factor measurement values ​​and measurement result consistency scores of the capacitive device.

Images