Method and system for automatically measuring and analyzing impedance characteristics of electric reactor

Abstract

The invention provides a reactor impedance characteristic automatic measurement and analysis method and system, and relates to the technical field of reactors, and the method comprises the steps: collecting voltage and current signals, calculating original impedance frequency characteristic data, carrying out the frequency domain decomposition of the data, carrying out the processing of a neural network compensation model, and building an impedance frequency characteristic curve; real-time monitoring data are continuously collected and compared with theoretical features, fault feature vectors are constructed, fault judgment criteria are established through the Euclidean distance and the Mahalanobis distance, automatic detection and accurate positioning of reactor faults are achieved, and the operation reliability of a power system is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technology of reactors, and in particular to an automated measurement and analysis method and system for the impedance characteristics of reactors. Background Art

[0002] Reactors are important devices in the power system and are widely used in fields such as power transmission, power quality control, and harmonic suppression. The impedance characteristics of reactors are key performance indicators, directly affecting the stability and reliability of the power system. With the expansion of the power grid scale and the improvement of complexity, the real-time monitoring of the operating state of reactors, the accurate measurement of impedance characteristics, and the timely diagnosis of faults are particularly important. Traditional methods for measuring the impedance characteristics of reactors mainly rely on off-line detection methods, usually requiring the equipment to be shut down and professional tests to be carried out. This method is not only time-consuming and laborious but also affects the normal operation of the power system.

[0003] Traditional impedance measurement methods lack real-time and dynamic performance and cannot meet the requirements of on-line monitoring of the power system. Existing technologies mostly adopt static measurement means and are difficult to capture the impedance change characteristics of reactors under different operating conditions. Especially when the power grid parameters fluctuate greatly, the measurement results often have large errors, affecting the accuracy of fault judgment.

[0004] Existing impedance characteristic analysis methods have high requirements for the accuracy of signal processing but lack an effective compensation mechanism. In the actual measurement process, errors introduced by sampling devices, environmental noise interference, and the influence of harmonic components will cause distortion of the impedance frequency characteristic curve. Existing technologies are difficult to effectively eliminate these influencing factors, thus affecting the accuracy of impedance characteristic analysis.

[0005] Traditional reactor fault diagnosis methods mainly rely on empirical judgment or simple threshold comparison and lack a systematic fault feature extraction and intelligent fault discrimination mechanism. This method is difficult to accurately identify complex fault modes, especially has weak recognition ability for early faults and multiple faults, resulting in inaccurate diagnosis results and affecting the scientificity and timeliness of equipment maintenance decisions. Summary of the Invention

[0006] Embodiments of the present invention provide an automated measurement and analysis method and system for the impedance characteristics of reactors, which can solve the problems in the prior art.

[0007] In a first aspect of embodiments of the present invention, there is provided an automated measurement and analysis method for the impedance characteristics of reactors, including:

[0008] Collect voltage signals and current signals through a sampling device of the reactor, and calculate the original impedance frequency characteristic data of the reactor according to the voltage signals and the current signals;

[0009] Perform frequency-domain decomposition on the original impedance frequency characteristic data to obtain fundamental wave components and harmonic components, and perform Fourier transform on the fundamental wave components and the harmonic components respectively to obtain spectral data; input the spectral data into a preset neural network compensation model to obtain compensation coefficients, multiply the compensation coefficients by the spectral data to obtain compensated impedance frequency characteristic data, and the neural network compensation model uses a long short-term memory network structure to extract temporal features and perform dynamic compensation on the spectral data;

[0010] Establish an impedance frequency characteristic curve of the reactor based on the compensated impedance frequency characteristic data, and calculate real-time characteristic data at different frequency points of the impedance frequency characteristic curve; continuously collect real-time monitoring data of the reactor under different operating conditions through the sampling device, calculate the real-time impedance frequency characteristic according to the real-time monitoring data, and compare the real-time impedance frequency characteristic with the real-time characteristic data to obtain a characteristic deviation;

[0011] Construct a fault feature vector based on the characteristic deviation, establish a fault discrimination criterion by calculating the Euclidean distance and Mahalanobis distance of the fault feature vector, and determine the fault type and fault location according to the fault discrimination criterion to generate a fault diagnosis report.

[0012] Performing frequency-domain decomposition on the original impedance frequency characteristic data to obtain fundamental wave components and harmonic components, and performing Fourier transform on the fundamental wave components and the harmonic components respectively includes:

[0013] Perform empirical mode decomposition on the original impedance frequency characteristic data to obtain intrinsic mode functions and a residual term, identify the index sets of intrinsic mode functions within the fundamental wave frequency range and the index sets of intrinsic mode functions within the harmonic frequency range according to the frequency characteristics of the intrinsic mode functions, and superimpose the index sets of intrinsic mode functions within the fundamental wave frequency range and the index sets of intrinsic mode functions within the harmonic frequency range to obtain fundamental wave components and harmonic components;

[0014] Perform short-time Fourier transform on the fundamental wave components and the harmonic components respectively to obtain a time-frequency spectrum. The short-time Fourier transform uses a variable-length Kaiser window function, and the window length and shape parameters of the variable-length Kaiser window function are adaptively adjusted according to the non-stationary characteristics of the signal;

[0015] Perform feature enhancement processing on the time-frequency spectrum. The feature enhancement processing includes spectrum smoothing and noise suppression. Convolve the time-frequency spectrum in the frequency dimension through a smoothing kernel function to obtain a smoothed spectrum, perform wavelet transform on the smoothed spectrum, and remove noise components through a soft threshold function to obtain enhanced spectral data.

[0016] Based on the compensated impedance frequency characteristic data, establish the impedance frequency characteristic curve of the reactor, and calculate the real-time characteristic data of the impedance frequency characteristic curve at different frequency points, including:

[0017] Construct an immune network model based on the compensated impedance frequency characteristic data, and construct an impedance characteristic curve according to the antibody memory unit of the immune network model. The impedance characteristic curve amplifies and mutates the antibody memory unit through the clonal selection algorithm. The clonal selection algorithm includes two sub-steps: affinity calculation and clonal amplification. Optimize the fitting accuracy of the impedance characteristic curve in each frequency interval through the clonal selection algorithm;

[0018] Calculate the first derivative and the second derivative of the impedance amplitude at different frequency points of the impedance characteristic curve, and re-enter the first derivative and the second derivative into the immune network model to update the impedance change rate information stored in the antibody memory unit;

[0019] Calculate the impedance distortion degree of each frequency interval according to the updated impedance change rate information. The impedance distortion degree judges the non-linear change of the impedance characteristic curve through the clonal selection algorithm. When the impedance distortion degree exceeds the preset distortion threshold, the clonal selection algorithm triggers a high-frequency mutation operation to optimize the real-time characteristic data of the corresponding frequency interval by generating a new antibody population.

[0020] Construct an immune network model based on the compensated impedance frequency characteristic data, and construct an impedance characteristic curve according to the antibody memory unit of the immune network model. The amplification and mutation of the antibody memory unit by the clonal selection algorithm for the impedance characteristic curve includes:

[0021] The immune network model includes multiple antibody units. Each antibody unit corresponds to the impedance frequency characteristic data of a frequency point. The affinity between the antibody units is calculated through a Gaussian kernel function. The affinity characterizes the similarity degree of the impedance frequency characteristic data of adjacent frequency points;

[0022] Calculate the network activation degree of the antibody unit based on the affinity. The network activation degree is obtained by weighted accumulation of the affinity. The weight coefficient is determined by the antibody memory factor. Prioritize storing the impedance frequency characteristic data corresponding to the antibody unit with the network activation degree higher than the preset activation degree threshold into the antibody memory unit;

[0023] Clone and amplify the impedance-frequency characteristic data stored in the antibody memory unit. The number of clones is proportional to the network activation degree. The newly generated antibody units inherit the original impedance-frequency characteristic data. Perform affinity maturation operations on the new antibody units. The affinity maturation operations perturb the impedance-frequency characteristic data through Gaussian random mutation, and the mutation amplitude is inversely proportional to the network activation degree.

[0024] Re-enter the mutated new antibody units into the immune network model to calculate the updated network activation degree. Screen the antibody memory unit according to the updated network activation degree, and retain the antibody memory units with the network activation degree higher than the preset activation degree threshold to construct an impedance characteristic curve.

[0025] Calculate the real-time impedance-frequency characteristic according to the real-time monitoring data, and compare the real-time impedance-frequency characteristic with the real-time characteristic data to obtain the characteristic deviation, including:

[0026] Perform recursive least squares estimation on the real-time monitoring data, including: constructing an observation equation containing voltage observation values, current observation values, and initial impedance parameters, and calculating the prediction error based on the observation equation. The prediction error is the estimation deviation between the real-time monitoring data and the historical impedance parameters.

[0027] Calculate the gain matrix according to the prediction error. The gain matrix is used to adjust the parameter update step size. Multiply the prediction error by the gain matrix to obtain the parameter update amount, and add the parameter update amount to the historical impedance parameters to obtain the impedance parameters at the current moment.

[0028] Perform Kalman filtering on the impedance parameters, calculate the state prediction value of the impedance parameters, and perform smoothing processing based on the deviation between the state prediction value and the real-time monitoring data to obtain the filtered impedance state estimation value.

[0029] Calculate the real-time impedance-frequency characteristic based on the impedance state estimation value. The real-time impedance-frequency characteristic includes amplitude characteristic, phase characteristic, and quality factor. Compare the real-time impedance-frequency characteristic with the real-time characteristic data to obtain the final characteristic deviation.

[0030] Construct a fault feature vector according to the characteristic deviation, establish a fault discrimination criterion by calculating the Euclidean distance and Mahalanobis distance of the fault feature vector, and determine the fault type and fault location according to the fault discrimination criterion to generate a fault diagnosis report, including:

[0031] Fuse the fault feature vector with the expert experience feature vector, weight the fault feature vector through a cognitive weight matrix to obtain an empirical fusion coefficient; weight the expert experience feature vector through the empirical fusion coefficient to obtain a cognitive enhanced feature vector;

[0032] Calculate the Euclidean distance and Mahalanobis distance of the cognitive enhanced feature vector. The Euclidean distance represents the direct distance between feature vectors, and the Mahalanobis distance takes into account the covariance relationship of feature distributions. Based on the Euclidean distance and the Mahalanobis distance, construct a fault discrimination criterion;

[0033] Perform cognitive weight optimization on the fault discrimination criterion. Input the Euclidean distance and the Mahalanobis distance into a loss function, introduce a cognitive rule constraint term, and obtain the optimized cognitive weight by minimizing the weighted sum of the loss function and the cognitive rule constraint term;

[0034] Calculate the fault probability distribution based on the optimized cognitive weight. The fault probability distribution adjusts the discrimination degree of fault types through a discrimination sensitivity parameter, and combine the fault probability distribution with the fault space distribution function to obtain a fault location probability mapping;

[0035] Use the reinforcement learning value function to update the expert experience knowledge. The expert experience knowledge is weighted between historical experience and the reinforcement learning value function through a learning rate, and optimize the fault location probability mapping based on the updated expert experience knowledge to generate a fault diagnosis report.

[0036] Calculating the fault probability distribution based on the optimized cognitive weight, and adjusting the discrimination degree of fault types through a discrimination sensitivity parameter, the combination of the fault probability distribution and the fault space distribution function to obtain a fault location probability mapping includes:

[0037] Calculate the fault probability distribution based on the optimized cognitive weight, and adjust the discrimination degree of fault types through a discrimination sensitivity parameter, and the discrimination sensitivity parameter is adaptively adjusted according to the degree of difference of fault features;

[0038] Perform weighted combination of the fault probability distribution and the fault space distribution function to obtain a fault location probability mapping. The fault space distribution function describes the spatial distribution characteristics of fault types, and the fault location probability mapping reflects the spatial probability distribution of the fault occurrence location.

[0039] In the second aspect of the embodiments of the present invention, an automatic measurement and analysis system for the impedance characteristics of a reactor is provided, including:

[0040] The first unit is used to collect voltage signals and current signals through a sampling device of the reactor, and calculate the original impedance frequency characteristic data of the reactor according to the voltage signals and the current signals;

[0041] A second unit, configured to perform frequency-domain decomposition on the original impedance-frequency characteristic data to obtain a fundamental wave component and a harmonic component, perform Fourier transform on the fundamental wave component and the harmonic component respectively to obtain spectrum data; input the spectrum data into a preset neural network compensation model to obtain a compensation coefficient, multiply the compensation coefficient by the spectrum data to obtain compensated impedance-frequency characteristic data, and the neural network compensation model uses a long short-term memory network structure to extract temporal features and perform dynamic compensation on the spectrum data;

[0042] A third unit, configured to establish an impedance-frequency characteristic curve of the reactor according to the compensated impedance-frequency characteristic data, and calculate real-time characteristic data at different frequency points of the impedance-frequency characteristic curve; continuously collect real-time monitoring data of the reactor under different operating conditions through the sampling device, calculate the real-time impedance-frequency characteristic according to the real-time monitoring data, and compare the real-time impedance-frequency characteristic with the real-time characteristic data to obtain a characteristic deviation;

[0043] A fourth unit, configured to construct a fault feature vector according to the characteristic deviation, establish a fault discrimination criterion by calculating the Euclidean distance and Mahalanobis distance of the fault feature vector, determine the fault type and fault location according to the fault discrimination criterion, and generate a fault diagnosis report.

[0044] In a third aspect of the embodiments of the present invention, an electronic device is provided, including:

[0045] A processor;

[0046] A memory for storing instructions executable by the processor;

[0047] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0048] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0049] The beneficial effects of the present application are as follows:

[0050] The present invention obtains the voltage and current signals of the reactor through a sampling device, calculates the original impedance characteristic and then performs frequency-domain decomposition and Fourier transform, and uses a neural network compensation model with a long short-term memory network structure to perform dynamic compensation on the spectrum data, effectively improving the accuracy and stability of the measurement of the impedance characteristic of the reactor, and solving the problem that the traditional measurement method is easily interfered under complex working conditions.

[0051] The present invention establishes a complete comparative analysis mechanism for impedance-frequency characteristic curves and real-time characteristic data. By continuously monitoring the changes in real-time impedance characteristics under different operating conditions and calculating the characteristic deviations, it realizes the dynamic tracking of the reactor operating state and early warning of abnormalities, improving the real-time monitoring ability of the system and the accuracy of fault prediction.

[0052] The present invention constructs a fault diagnosis method based on feature vectors. By calculating the Euclidean distance and Mahalanobis distance to establish a fault discrimination criterion, it can accurately identify the fault type and locate the fault position, realizing the intelligent diagnosis and analysis of reactor faults, providing a scientific basis for the maintenance and management of power grid equipment, and having significant engineering application value. Brief Description of the Drawings

[0053] Figure 1 It is a schematic flow chart of the automatic measurement and analysis method for the impedance characteristics of the reactor in the embodiment of the present invention;

[0054] Figure 2 It is a schematic comparison diagram of the time-frequency resolution of different decomposition methods in the embodiment of the present invention;

[0055] Figure 3 It is a flow chart for establishing the impedance-frequency characteristic curve and calculating characteristic data in the embodiment of the present invention;

[0056] Figure 4 It is a schematic comparison diagram of the estimation stability under different environmental interferences in the embodiment of the present invention;

[0057] Figure 5 It is a bar chart for the performance comparative analysis of the cognitive enhanced fault diagnosis method in the embodiment of the present invention. Detailed Embodiments

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0060] Figure 1 It is a schematic flow chart of the automatic measurement and analysis method for the impedance characteristics of the reactor in the embodiment of the present invention, as Figure 1 shown, the method includes:

[0061] The sampling device for the reactor collects voltage signals and current signals, and calculates the original impedance frequency characteristic data of the reactor according to the voltage signals and the current signals;

[0062] The original impedance frequency characteristic data is decomposed in the frequency domain to obtain a fundamental wave component and a harmonic component, and the Fourier transform is respectively performed on the fundamental wave component and the harmonic component to obtain spectrum data; the spectrum data is input into a preset neural network compensation model to obtain a compensation coefficient, and the compensation coefficient is multiplied by the spectrum data to obtain compensated impedance frequency characteristic data. The neural network compensation model uses a long short-term memory network structure to extract time series features and perform dynamic compensation on the spectrum data;

[0063] An impedance frequency characteristic curve of the reactor is established according to the compensated impedance frequency characteristic data, and real-time characteristic data at different frequency points of the impedance frequency characteristic curve is calculated; the sampling device continuously collects real-time monitoring data of the reactor under different operating conditions, calculates the real-time impedance frequency characteristic according to the real-time monitoring data, and compares the real-time impedance frequency characteristic with the real-time characteristic data to obtain a characteristic deviation;

[0064] A fault feature vector is constructed according to the characteristic deviation, a fault discrimination criterion is established by calculating the Euclidean distance and the Mahalanobis distance of the fault feature vector, and the fault type and the fault location are determined according to the fault discrimination criterion, and a fault diagnosis report is generated.

[0065] In an optional implementation manner, decomposing the original impedance frequency characteristic data in the frequency domain to obtain a fundamental wave component and a harmonic component, and respectively performing the Fourier transform on the fundamental wave component and the harmonic component to obtain spectrum data includes:

[0066] The original impedance frequency characteristic data is decomposed by empirical mode decomposition to obtain intrinsic mode functions and a residue term, the index sets of the intrinsic mode functions in the fundamental wave frequency range and the index sets of the intrinsic mode functions in the harmonic frequency range are identified according to the frequency characteristics of the intrinsic mode functions, and the index sets of the intrinsic mode functions in the fundamental wave frequency range and the index sets of the intrinsic mode functions in the harmonic frequency range are superimposed to obtain a fundamental wave component and a harmonic component;

[0067] The short-time Fourier transform is respectively performed on the fundamental wave component and the harmonic component to obtain a time-frequency spectrum. The short-time Fourier transform uses a variable-length Kaiser window function, and the window length and shape parameters of the variable-length Kaiser window function are adaptively adjusted according to the non-stationary characteristics of the signal;

[0068] Perform feature enhancement processing on the time-frequency spectrum. The feature enhancement processing includes spectrum smoothing and noise suppression. The time-frequency spectrum is convolved in the frequency dimension through a smoothing kernel function to obtain a smoothed spectrum. Wavelet transform is performed on the smoothed spectrum, and noise components are removed through a soft threshold function to obtain enhanced spectrum data.

[0069] For the frequency-domain decomposition of the original impedance frequency characteristic data, the empirical mode decomposition (EMD) method is adopted in this embodiment. Suppose there is a set of collected original impedance frequency characteristic data with 10,000 data points and a sampling frequency of 1 kHz. The EMD method decomposes the signal into multiple intrinsic mode functions (IMFs) and a residual term. In actual operation, the EMD process finds the local extreme points of the signal, connects all the maximum and minimum points to form the upper and lower envelope lines respectively. Through iterative calculation, a series of IMF components are obtained. For the 10,000-point original impedance data, 10 IMF components and 1 residual term are obtained after EMD decomposition.

[0070] For the obtained IMF components, it is necessary to identify which ones belong to the fundamental frequency range and which ones belong to the harmonic frequency range. This step is achieved by analyzing the frequency characteristics of each IMF component. Specifically, calculate the main frequency component of each IMF. If the main frequency falls within the predefined fundamental frequency range (e.g., 45 - 55 Hz), add the index of this IMF to the index set of the intrinsic mode functions in the fundamental frequency range; if the main frequency falls within the harmonic frequency range (e.g., 95 - 105 Hz, 145 - 155 Hz, etc.), add the index of this IMF to the index set of the intrinsic mode functions in the harmonic frequency range.

[0071] By analyzing the main frequency characteristics of the 10 IMFs, it is determined that the main frequencies of IMF1, IMF2, and IMF3 are 150 Hz, 100 Hz, and 50 Hz respectively. Therefore, the index set of the intrinsic mode functions in the fundamental frequency range is {3} (corresponding to IMF3), and the index set of the intrinsic mode functions in the harmonic frequency range is {1, 2} (corresponding to IMF1 and IMF2). The main frequencies of the remaining IMFs are not within the concerned frequency range and are not included in these two index sets.

[0072] Superimpose the IMFs within the fundamental frequency range (i.e., IMF3) to obtain the fundamental component, and superimpose the IMFs within the harmonic frequency range (i.e., IMF1 and IMF2) to obtain the harmonic component. In this way, the original impedance frequency characteristic data is decomposed into two parts: the fundamental component and the harmonic component.

[0073] The short-time Fourier transform (STFT) is performed on the fundamental wave component and harmonic components respectively to obtain time-frequency spectrum information. In STFT, a variable-length Kaiser window function is used to segment the signal. The Kaiser window function has good frequency-domain characteristics and can reduce the spectrum leakage phenomenon. The window length and shape parameter of the variable-length Kaiser window function are adaptively adjusted according to the non-stationary characteristics of the signal.

[0074] Specifically, the selection of the window length depends on the local stationarity of the signal. For relatively stationary signal segments, a longer window (e.g., 512 points) is used to obtain higher frequency resolution; for signal segments with faster changes, a shorter window (e.g., 128 points) is used to obtain higher time resolution. The shape parameter β controls the sidelobe attenuation degree of the window function. The larger the β value, the smaller the spectrum leakage, but the main lobe width increases and the frequency resolution decreases. In the example, for the fundamental wave component, β = 6.0 is selected as the shape parameter of the Kaiser window; for the harmonic components, β = 8.0 is selected to obtain better frequency resolution ability.

[0075] Through STFT processing, the time-frequency spectra of the fundamental wave component and harmonic components are obtained. The time-frequency spectrum is a three-dimensional data. The abscissa represents time, the ordinate represents frequency, and the color depth represents the energy magnitude. In this example, the time-frequency spectrum of the fundamental wave component shows an obvious energy concentration region near 50 Hz, and the time-frequency spectrum of the harmonic components has energy distributions at 100 Hz and 150 Hz.

[0076] To improve the quality of the time-frequency spectrum, feature enhancement processing is performed on it, including two steps: spectrum smoothing and noise suppression. Spectrum smoothing is achieved by convolving with a smoothing kernel function in the frequency dimension. In this embodiment, a Gaussian kernel function is selected as the smoothing kernel, and the kernel width is set to 5 frequency points to balance the smoothing effect and frequency resolution. Through the convolution operation, the original time-frequency spectrum is converted into a smoothed spectrum, eliminating local mutations and spikes in the spectrum.

[0077] The noise suppression processing is implemented based on wavelet transform and soft threshold function. The smoothed spectrum data is wavelet decomposed, and the Daubechies4 (db4) wavelet basis is selected for 5-level decomposition. The soft threshold function is applied to the wavelet coefficients, and the threshold is set to 3 times the standard deviation of the wavelet coefficients. The soft threshold function sets the coefficients smaller than the threshold to zero, and shrinks the coefficients larger than the threshold, thereby effectively removing the random noise components in the spectrum while retaining the main features of the signal.

[0078] Through the inverse wavelet transform, the processed wavelet coefficients are recombined into enhanced spectrum data. After the feature enhancement processing, the signal-to-noise ratio of the time-frequency spectrum is significantly improved, and the frequency characteristics of the fundamental wave and harmonic components are more clearly distinguishable, providing a high-quality data basis for subsequent feature extraction and pattern recognition.

[0079] In practical applications, through decomposition and reconstruction, this method can effectively separate the fundamental wave and harmonic components in the power system, reveal the variation law of impedance characteristics over time through time-frequency analysis, and further improve the data quality through feature enhancement processing, providing accurate and reliable technical support for power system impedance modeling and harmonic assessment.

[0080] Figure 2 Schematic diagram for comparing the time-frequency resolution of different decomposition methods in the embodiments of the present invention:

[0081] This figure shows the comparison of the analysis accuracy of three different analysis methods in the frequency range of 0 - 500 Hz. The technical solution of the present invention (circular marker) exhibits excellent performance in the low-frequency band, reaching an accuracy of 0.92 at 0 Hz and maintaining a high accuracy level of 0.93 - 0.94 in the range of 50 - 100 Hz. As the frequency increases, its accuracy shows a slow downward trend, but still maintains a relatively high accuracy of 0.77 at 500 Hz. The overall performance of the fixed-window STFT method (diamond marker) is at a medium level, about 0.68 - 0.70 in the low-frequency band, but gradually decreases to 0.55 - 0.57 as the frequency increases. The performance of the standard Fourier transform method (square marker) is relatively the worst, only 0.39 - 0.42 in the low-frequency band. Although it slightly increases to 0.53 - 0.55 in the range of 200 - 250 Hz, it continues to decline to about 0.40 in the high-frequency band. From the trend, the technical solution of the present invention maintains an obvious performance advantage in the entire frequency band, especially in the low-frequency band, with an average advantage of 20 - 40 percentage points higher than the other two methods, demonstrating stronger frequency analysis capabilities.

[0082] In an alternative embodiment, an impedance-frequency characteristic curve of the reactor is established based on the compensated impedance-frequency characteristic data, and calculating the real-time characteristic data of the impedance-frequency characteristic curve at different frequency points includes:

[0083] Construct an immune network model based on the compensated impedance-frequency characteristic data, construct an impedance characteristic curve according to the antibody memory unit of the immune network model, the impedance characteristic curve amplifies and mutates the antibody memory unit through the clonal selection algorithm, the clonal selection algorithm includes two sub-steps of affinity calculation and clonal amplification, and optimizes the fitting accuracy of the impedance characteristic curve in each frequency interval through the clonal selection algorithm;

[0084] Calculate the first derivative and the second derivative of the impedance amplitude at different frequency points of the impedance characteristic curve, and re-input the first derivative and the second derivative into the immune network model to update the impedance change rate information stored in the antibody memory unit;

[0085] Calculate the impedance distortion degree of each frequency interval according to the updated impedance change rate information. The impedance distortion degree judges the non-linear change of the impedance characteristic curve through the clonal selection algorithm. When the impedance distortion degree exceeds the preset distortion threshold, the clonal selection algorithm triggers a high-frequency mutation operation to optimize the real-time characteristic data of the corresponding frequency interval by generating a new antibody population.

[0086] As Figure 3 shown, the method includes:

[0087] Obtain the original impedance frequency characteristic measurement data of the reactor, and perform compensation processing on it to eliminate the influence of measurement errors and environmental factors. The compensation processing adopts measurement system error correction technology to identify and correct the system error and random error in the original data. For example, for a set of measured impedance data [120.5Ω, 125.3Ω, 132.1Ω, 140.8Ω, 152.4Ω] at the frequency points [50Hz, 100Hz, 500Hz, 1kHz, 5kHz], the corrected values [120.2Ω, 125.0Ω, 132.5Ω, 141.0Ω, 152.0Ω] can be obtained through compensation processing.

[0088] Based on the compensated impedance frequency characteristic data, construct an immune network model. This model consists of an antigen recognition layer, an antibody network layer, and a memory unit layer. The antigen recognition layer receives the compensated impedance data as input. The antibody network layer contains multiple antibody units, and each unit corresponds to the impedance characteristic of a frequency point. The memory unit layer stores the optimized impedance characteristic curve parameters. In actual implementation, for a dry-type air-core reactor with a rated voltage of 35kV, the number of units in the antibody network layer can be set to 50, and each unit contains three parameters: impedance amplitude, phase angle, and impedance change rate.

[0089] When constructing the impedance characteristic curve, it is carried out according to the content of the antibody memory unit of the immune network model. In specific implementation, the frequency range is divided into multiple intervals. For example, the range from 50Hz to 5kHz is divided into 20 intervals, and 5 characteristic frequency points are selected for each interval, and a total of 100 frequency points form a characteristic frequency set. For each frequency point, the corresponding impedance characteristic parameters are extracted from the antibody memory unit to form an initial impedance characteristic curve.

[0090] The antibody memory units are amplified and mutated using the clonal selection algorithm to optimize the fitting accuracy of the impedance characteristic curve. The clonal selection algorithm consists of two sub-steps: affinity calculation and clonal amplification. In the affinity calculation stage, an affinity function between the antibody and the antigen is defined to evaluate the degree of agreement between the current impedance characteristic curve and the actual measurement data. For example, at the frequency point f1 = 100Hz, if the actual measured impedance value is 125.0Ω and the predicted value of the current curve is 124.8Ω, the calculated affinity is 0.998. In the clonal amplification stage, the antibodies with high affinity are cloned and amplified, and the number of clones is proportional to the affinity. For example, an antibody with an affinity of 0.998 can produce 20 clones, while an antibody with an affinity of 0.85 can only produce 5 clones.

[0091] Mutation operations are performed on the amplified antibody population. The mutation probability is inversely proportional to the affinity to increase the population diversity. There are two mutation methods: Gaussian mutation and uniform mutation. Different mutation strategies are selected for different frequency ranges. For the low-frequency region (50Hz - 500Hz), small-amplitude Gaussian mutation is used, and the mutation amplitude is controlled within the range of ±2% of the original value; for the high-frequency region (1kHz - 5kHz), larger-amplitude uniform mutation is used, and the mutation amplitude can reach ±5% of the original value. After mutation, the affinity is recalculated, and the antibody population with the highest affinity is selected to update the memory unit.

[0092] For the optimized impedance characteristic curve, the first derivative and the second derivative of the impedance amplitude are calculated at each frequency point. The first derivative represents the rate of change of the impedance with respect to frequency, and the second derivative represents the trend of the rate of change. For example, in the frequency range from 100Hz to 500Hz, the calculated first derivative is 0.025Ω / Hz, indicating that for every 1Hz increase, the impedance increases by 0.025Ω; the second derivative is 0.00002Ω / Hz 2 , indicating that the rate of change of the impedance shows a weak accelerating trend.

[0093] The calculated first derivative and second derivative are re-input into the immune network model to update the impedance rate of change information stored in the antibody memory unit. The update process uses an adaptive weight method to perform weighted fusion on the newly calculated rate of change and the existing data in the memory unit. The weight coefficient is automatically adjusted according to the credibility of the current data. For example, for a measurement interval with a high signal-to-noise ratio, the weight of the new data can be set to 0.8, while for an interval with a large amount of noise, the weight can be reduced to 0.4.

[0094] According to the updated impedance change rate information, calculate the impedance distortion degree of each frequency interval. The impedance distortion degree is defined as the deviation degree between the impedance curve and the ideal linear change. During the calculation, the actual impedance curve is compared with the theoretical linear model, and the distortion degree is quantified by the cumulative sum of squared errors. For example, in the frequency interval from 500 Hz to 1 kHz, if the calculated impedance distortion degree is 0.07, it indicates that the impedance characteristics in this interval show a slight non-linear change.

[0095] When the impedance distortion degree of a certain frequency interval exceeds the preset distortion threshold (such as 0.15), trigger the high-frequency mutation operation of the clonal selection algorithm. The high-frequency mutation operation generates a more diverse antibody population by increasing the mutation probability and mutation amplitude, and searches for a better combination of characteristic parameters. For example, for the 1 kHz to 2 kHz interval with a distortion degree reaching 0.18, the mutation probability can be increased from the standard 0.1 to 0.3, and the mutation amplitude can be expanded from ±5% to ±10%. Generate 100 new antibody candidate solutions, and select the 10 with the highest affinity from these candidate solutions to update the memory unit, and optimize the real-time characteristic data of this frequency interval.

[0096] Through the above method, the impedance frequency characteristic curve of the reactor can be accurately constructed, and the real-time characteristic data at different frequency points can be calculated, providing reliable data support for the condition monitoring and fault diagnosis of the reactor.

[0097] In an alternative embodiment, an immune network model is constructed based on the compensated impedance frequency characteristic data, and an impedance characteristic curve is constructed according to the antibody memory unit of the immune network model. The impedance characteristic curve amplifies and mutates the antibody memory unit through the clonal selection algorithm, including:

[0098] The immune network model includes multiple antibody units. Each antibody unit corresponds to the impedance frequency characteristic data of a frequency point. The affinity between the antibody units is calculated by the Gaussian kernel function, and the affinity characterizes the similarity degree of the impedance frequency characteristic data of adjacent frequency points;

[0099] Based on the affinity, calculate the network activation degree of the antibody unit. The network activation degree is obtained by weighted accumulation of the affinity, and the weight coefficient is determined by the antibody memory factor. The impedance frequency characteristic data corresponding to the antibody unit with a network activation degree higher than the preset activation degree threshold is preferentially stored in the antibody memory unit;

[0100] Clone and amplify the impedance frequency characteristic data stored in the antibody memory unit. The number of clones is proportional to the network activation degree. The newly generated antibody units inherit the original impedance frequency characteristic data. Perform affinity maturation operations on the new antibody units. The affinity maturation operation perturbs the impedance frequency characteristic data through Gaussian random mutation, and the mutation amplitude is inversely proportional to the network activation degree.

[0101] Re-enter the mutated new antibody units into the immune network model to calculate the updated network activation degree. Screen the antibody memory unit according to the updated network activation degree, and retain the antibody memory unit whose network activation degree is higher than the preset activation degree threshold to construct an impedance characteristic curve.

[0102] When constructing an immune network model based on the compensated impedance frequency characteristic data, the system first represents the impedance data corresponding to each frequency point as an antibody unit. For example, for the impedance data of 100 measured frequency points, the system creates 100 antibody units, and each antibody unit contains the impedance amplitude, phase information, and frequency identifier corresponding to the frequency point. In practical applications, for example, when performing EIS testing on a lithium battery, impedance data of 64 frequency points in the range of 1 Hz to 10 kHz are obtained, and each of these 64 measured values forms an antibody unit.

[0103] The affinity between antibody units is calculated by a Gaussian kernel function, which reflects the similarity of impedance data at adjacent frequency points. For any two antibody units A i and A j , calculate the Euclidean distance between them, including the differences in the real and imaginary parts of the impedance. Substitute this distance into the Gaussian kernel function to obtain the affinity value. The bandwidth parameter of the kernel function is set to 0.75 to ensure that the impedance data at adjacent frequency points have a high affinity, and the affinity of points far away decays rapidly. In practice, for example, for the antibody units at two frequency points of 1 Hz and 1.5 Hz, if the impedance data are similar, the calculated affinity can reach 0.92; while the affinity between the antibody units at 1 Hz and 100 Hz is only 0.13.

[0104] When calculating the network activation degree, the system performs weighted accumulation on the affinity between each antibody unit and all other units. The weight coefficient is determined by the antibody memory factor. Initially, all antibody memory factors are set to 1.0. As the learning process progresses, the memory factors of important frequency points will increase. For example, after calculating the affinity between an antibody unit and 64 other units respectively, multiply by the corresponding memory factors (initially all 1.0) and sum to obtain the network activation degree of this unit. In the lithium battery test case, the antibody units corresponding to the inflection point frequencies of the impedance spectrum usually obtain a higher activation degree. For example, the activation degree of the 30 Hz frequency point can reach 6.8, while the activation degree of a point far from the characteristic frequency, such as 9 kHz, is only 2.3.

[0105] The system sets the activation threshold to 4.0, and preferentially stores the impedance frequency characteristic data corresponding to the antibody units with network activation higher than this threshold into the antibody memory unit. In an actual case, through this threshold screening, about 20 out of 64 frequency points are stored in the antibody memory unit, and these points are mostly concentrated in the characteristic frequency region of the impedance spectrum, such as the frequency points corresponding to the SEI film impedance and charge transfer impedance during the charge and discharge process.

[0106] Clone amplification is performed on the impedance data already stored in the antibody memory unit, and the number of clones is positively correlated with the network activation. For example, an antibody unit with a network activation of 6.8 clones to produce 6 new antibodies, while a unit with an activation of 4.2 only produces 4 new antibodies. The newly generated antibody units inherit the characteristics of the original impedance data, but need to perform affinity maturation operations.

[0107] The affinity maturation operation perturbs the impedance characteristic data through Gaussian random mutation. The mutation amplitude is inversely proportional to the network activation. Antibodies with high activation have a small mutation amplitude to maintain stability; antibodies with low activation have a large mutation amplitude to increase exploration. When actually executed, for an antibody unit with an activation of 6.8, its mutation standard deviation is set to 1.5% of the original impedance value; while for a unit with an activation of 4.2, the mutation standard deviation is set to 3.5%. Through this mechanism, the system can fully explore the edge region while maintaining the stability of important feature points.

[0108] The newly generated antibody units after mutation are re-input into the immune network model, and the network activation is recalculated. For example, a point with an original antibody unit activation of 6.8 generates 6 new antibody units after cloning and mutation, and their new activations are 6.9, 6.7, 6.8, 6.5, 7.0, and 6.6 respectively. The system retains the units with activation higher than the preset threshold of 4.0. In this case, all 6 new antibodies are retained.

[0109] After completing multiple rounds of clone selection iteration (usually set to 5 rounds), the system integrates all the retained antibody memory units to construct a complete impedance characteristic curve. The final result is an optimized impedance curve based on the original measurement data, highlighting the characteristics of key frequency points, and at the same time filling the data sparse region through the mutation mechanism. In the case of the battery impedance spectrum, after the immune network optimization of the original 64 frequency points, a smooth impedance curve containing more than 120 frequency points can be obtained. This curve has a higher point density and accuracy in the key frequency region (such as 0.5 Hz to 10 Hz), and can more accurately reflect the impedance characteristics of the electrochemical system.

[0110] In system implementation, the antibody memory unit is stored using a priority queue data structure, which facilitates sorting by network activation degree. The clonal selection process is executed in parallel on a multi-core processor to improve computational efficiency. For large-scale impedance data containing 1000 frequency points, the complete processing time is controlled within 2 seconds, meeting the requirements of real-time analysis.

[0111] In an alternative implementation, calculating the real-time impedance frequency characteristics from the real-time monitoring data and comparing the real-time impedance frequency characteristics with the real-time feature data to obtain a feature deviation includes:

[0112] Performing recursive least squares estimation on the real-time monitoring data, including: constructing an observation equation containing voltage observation values, current observation values, and initial impedance parameters, and calculating a prediction error based on the observation equation, where the prediction error is the estimation deviation between the real-time monitoring data and historical impedance parameters;

[0113] Calculating a gain matrix based on the prediction error, where the gain matrix is used to adjust the parameter update step size, multiplying the prediction error by the gain matrix to obtain a parameter update amount, and adding the parameter update amount to the historical impedance parameters to obtain the impedance parameters at the current moment;

[0114] Performing Kalman filtering on the impedance parameters, calculating a state prediction value of the impedance parameters, and performing smoothing processing based on the deviation between the state prediction value and the real-time monitoring data to obtain a filtered impedance state estimation value;

[0115] Calculating the real-time impedance frequency characteristics based on the impedance state estimation value, where the real-time impedance frequency characteristics include amplitude characteristics, phase characteristics, and quality factor; comparing the real-time impedance frequency characteristics with the real-time feature data to obtain the final feature deviation.

[0116] Performing recursive least squares estimation on the collected real-time monitoring data of the power grid, then performing Kalman filtering, and finally calculating the real-time impedance frequency characteristics and comparing and analyzing them with the standard feature data.

[0117] The real-time monitoring data of the power grid includes information such as sampling time points, voltage amplitudes, current amplitudes, phase angles, etc. To implement recursive least squares estimation, the system first constructs an observation equation that integrates voltage observation values, current observation values, and initial impedance parameters. The initial impedance parameters can be determined based on historical operation data or theoretical calculation values. For example, at a certain 500 kV power grid node, the initial resistance parameter can be set to 0.25 ohms, and the initial inductance parameter can be set to 0.0012 henries.

[0118] After the observation equation is constructed, the system calculates the prediction error, which is the deviation between the real-time monitoring data and the predicted value based on historical impedance parameters. For example, when the system detects that the voltage at a certain moment is 510 kV and the predicted value based on historical impedance parameters is 505 kV, the prediction error is 5 kV. This prediction error reflects the difference between the current power grid state and the historical model and is an important basis for updating the impedance parameters.

[0119] After the prediction error is calculated, the system calculates the gain matrix, which is used to adjust the parameter update step size. The calculation of the gain matrix takes into account the influence weight of historical observation data on the current estimate and usually decreases as the amount of observation data increases to ensure the stability of the estimation process. In practical applications, the forgetting factor can be set to adjust the weight ratio of new and old data. For example, when the forgetting factor is set to 0.98, it means that the weight of new data is slightly higher than that of old data, which helps the impedance parameters to adapt to the changes in the power grid state faster.

[0120] Multiply the prediction error by the gain matrix to obtain the parameter update amount. For example, the update amount of the resistance parameter at a certain moment is 0.02 ohms, and the update amount of the inductance parameter is 0.0001 henry. Then add these update amounts to the historical impedance parameters to obtain the estimated value of the impedance parameters at the current moment. For example, the updated resistance is 0.27 ohms and the inductance is 0.0013 henry.

[0121] To further improve the accuracy of impedance parameter estimation, the system performs Kalman filtering on the impedance parameters obtained by recursive least squares. Kalman filtering first calculates the state prediction value of the impedance parameters, which is obtained based on the impedance state at the previous moment and the system model. In this embodiment, the state prediction model assumes that the impedance parameters change little in a short time and uses the estimated value at the previous moment as the prediction value at the current moment, while considering a certain process noise. For example, the standard deviation of the process noise of the resistance parameter is set to 0.01 ohms, and the standard deviation of the process noise of the inductance parameter is set to 0.0005 henry.

[0122] After the state prediction is completed, the system compares the deviation between the predicted value and the real-time monitoring data and performs smoothing processing based on this deviation. The key parameter for smoothing processing is the Kalman gain, which determines the degree of trust of the system in new measurement data. When the measurement noise is small, the Kalman gain is large, and the system is more inclined to trust the new measurement value; when the measurement noise is large, the Kalman gain is small, and the system is more inclined to trust the prior estimate. For example, when the standard deviation of the measurement noise of the voltage is 0.5 kV and the standard deviation of the measurement noise of the current is 20 A, the Kalman gain is approximately 0.7, indicating that the system trusts the new measurement data more.

[0123] Through Kalman filtering processing, the system obtains the filtered impedance state estimation value, which is smoother and has stronger anti-interference ability than the result obtained by simply using the recursive least squares method. For example, the filtered resistance parameter is 0.265 ohms and the inductance parameter is 0.00128 henries, which has lower volatility than the original estimation value.

[0124] Based on the filtered impedance state estimation value, the system calculates the real-time impedance frequency characteristics, including the amplitude characteristics, phase characteristics, and quality factor at different frequency points. The amplitude characteristics represent the magnitude relationship of the impedance changing with frequency, the phase characteristics represent the phase angle relationship of the impedance changing with frequency, and the quality factor reflects the quality of the impedance characteristics. Under normal operating conditions, the standard impedance amplitude at 50 Hz of a certain power grid node is 0.3 ohms, the phase is 25 degrees, and the quality factor is 10.5.

[0125] Compare the calculated real-time impedance frequency characteristics with the stored real-time characteristic data to obtain the final characteristic deviation. The characteristic deviation reflects the difference between the current power grid impedance characteristics and the normal state or expected state, and is an important indicator for evaluating the health status of the power grid. For example, if the impedance amplitude at 50 Hz calculated in real time is 0.35 ohms, the phase is 28 degrees, and the quality factor is 9.8, then the amplitude deviation is 0.05 ohms, the phase deviation is 3 degrees, and the quality factor deviation is -0.7.

[0126] Through the above technical means, this method can accurately and real-time monitor the changes of power grid impedance parameters, timely detect abnormal situations, and provide strong support for the safe and stable operation of the power grid. Practice shows that this method can maintain an impedance estimation accuracy of more than 95% in a complex power grid environment, and the average detection delay of the characteristic deviation is less than 50 milliseconds, meeting the requirements of real-time monitoring of the power grid.

[0127] Figure 4 The following is a schematic diagram for comparing the estimation stability under different environmental interferences in the embodiments of the present invention:

[0128] The box plot shows the comparison of the measurement error distributions of three different methods under five working conditions. Under normal circumstances, the measurement error of this technical solution is the smallest, with a median of only 0.5%, while the median errors of the traditional method and the comparative method are 1.2% and 1.1% respectively. As the complexity of the working conditions increases, the measurement errors of all three methods show an upward trend: under the temperature fluctuation condition, the median errors of the three methods increase to 0.9%, 2.0% and 1.5% respectively; in the electromagnetic interference environment, the errors further increase to 1.5%, 3.1% and 2.5%; under the load change condition, the errors reach 2.0%, 3.7% and 3.1% respectively; in the most complex harmonic interference condition, the errors are the largest, being 2.5%, 4.7% and 4.1% respectively. From the perspective of the error distribution range, this technical solution shows a smaller error fluctuation range under various working conditions, demonstrating better stability. Especially under complex working conditions, the upper limit of the error of this technical solution is significantly lower than that of the other two methods, showing stronger anti-interference ability and adaptability.

[0129] In an alternative embodiment, a fault feature vector is constructed according to the characteristic deviation, a fault discrimination criterion is established by calculating the Euclidean distance and Mahalanobis distance of the fault feature vector, and the fault type and fault location are determined according to the fault discrimination criterion, and a fault diagnosis report is generated including:

[0130] The fault feature vector is fused with the expert experience feature vector, the fault feature vector is weighted by a cognitive weight matrix to obtain an empirical fusion coefficient; the expert experience feature vector is weighted by the empirical fusion coefficient to obtain a cognitive enhanced feature vector;

[0131] Calculate the Euclidean distance and Mahalanobis distance of the cognitive enhanced feature vector. The Euclidean distance represents the direct distance between feature vectors, and the Mahalanobis distance takes into account the covariance relationship of the feature distribution. A fault discrimination criterion is constructed based on the Euclidean distance and the Mahalanobis distance;

[0132] Perform cognitive weight optimization on the fault discrimination criterion, input the Euclidean distance and the Mahalanobis distance into a loss function, introduce a cognitive rule constraint term, and obtain the optimized cognitive weight by minimizing the weighted sum of the loss function and the cognitive rule constraint term;

[0133] Calculate the fault probability distribution based on the optimized cognitive weight. The fault probability distribution adjusts the discrimination degree of the fault type through a discrimination sensitivity parameter, and combines the fault probability distribution with the fault space distribution function to obtain a fault location probability mapping;

[0134] Update the expert experience knowledge using the reinforcement learning value function. The expert experience knowledge is weighted between historical experience and the reinforcement learning value function through a learning rate. Optimize the fault location probability mapping based on the updated expert experience knowledge to generate a fault diagnosis report.

[0135] Accurately determine the fault type and location by constructing a fault feature vector and combining the Euclidean distance and Mahalanobis distance to establish a fault discrimination criterion. The system first collects the device operation status data, including key parameters such as vibration signals, temperature changes, and current fluctuations. Preprocess the collected data to eliminate noise interference and extract effective information. Then calculate the feature deviation between the normal operation state and the current state to form an initial fault feature vector.

[0136] This method performs a fusion process on the fault feature vector and the expert experience feature vector. Assume the initial fault feature vector is [0.58, 0.23, 0.67, 0.41], representing the deviation values of four key feature parameters, and the expert experience feature vector is [0.60, 0.20, 0.65, 0.45]. Weight the fault feature vector through a cognitive weight matrix, which is obtained based on historical case analysis. For this type of fault, it can be set as [0.4, 0.2, 0.3, 0.1], representing the importance of the four feature parameters. Multiply the fault feature vector by the cognitive weight matrix to obtain an empirical fusion coefficient of 0.484. Subsequently, use this coefficient to weight the expert experience feature vector to obtain a cognitively enhanced feature vector [0.290, 0.097, 0.315, 0.218].

[0137] Calculate the Euclidean distance and Mahalanobis distance of the cognitively enhanced feature vector to accurately judge the fault type. The Euclidean distance represents the direct distance between feature vectors and intuitively reflects the degree of deviation of the fault from the normal state. The Euclidean distance between the sample feature vector [0.290, 0.097, 0.315, 0.218] and the standard fault type A feature vector [0.300, 0.100, 0.320, 0.220] is calculated to be 0.013, indicating that they are very close. The Mahalanobis distance takes into account the covariance relationship of the feature distribution and can better handle the correlation between features. For the same set of data, the calculated Mahalanobis distance is 0.867, which is lower than the fault discrimination threshold of 1.0, further confirming that the fault belongs to type A. Based on these two distance metrics, construct a comprehensive fault discrimination criterion: when the Euclidean distance is less than 0.05 and the Mahalanobis distance is less than 1.0, it is determined as the corresponding fault type.

[0138] Perform cognitive weight optimization on the fault discrimination criteria to improve the discrimination accuracy. Take the Euclidean distance and Mahalanobis distance as input parameters, and introduce a loss function to calculate the discrimination error. The loss function is designed as a weighted combination of the Euclidean distance and Mahalanobis distance, with the initial weights being 0.6 and 0.4 respectively. At the same time, introduce a cognitive rule constraint term, and set the characteristic distribution rules of different fault types according to expert knowledge. Minimize the weighted sum of the loss function and the cognitive rule constraint term through an iterative optimization method. After 50 iterations, the optimized cognitive weights are [0.45, 0.15, 0.32, 0.08], which can more accurately reflect the actual contributions of each characteristic parameter to fault judgment than the initial weights.

[0139] Calculate the fault probability distribution based on the optimized cognitive weights, which helps to determine the fault type and location. For the identified fault type A, the probability distributions of each fault subtype calculated according to the optimized cognitive weights are [0.65, 0.20, 0.10, 0.05], indicating that the probability of belonging to subtype A1 is the highest. The fault probability distribution is adjusted by the discrimination sensitivity parameter, which is set to 1.2, improving the discrimination between fault types and making the main fault characteristics more prominent. Combine the fault probability distribution with the fault space distribution function of the equipment. The fault space distribution function is predefined based on the structural characteristics of the equipment and describes the distribution law of faults in the physical space. After combination, a fault location probability map is obtained, indicating that the fault occurs in the rotor bearing part of the equipment with a probability of 0.78.

[0140] This method also uses the reinforcement learning value function to dynamically update the expert experience knowledge base. After each fault diagnosis is completed, calculate the diagnosis accuracy rate according to the actual fault confirmation situation as the reward signal for reinforcement learning. For this diagnosis, the coincidence degree between the actual fault confirmation and the diagnosis result is 92%, generating a positive reward of 0.92. The expert experience knowledge is weighted between the historical experience and the reinforcement learning value function through a learning rate of 0.3, keeping 70% of the historical stability of the knowledge base while absorbing 30% of the new experience. Based on the updated expert experience knowledge, the system fine-tunes the fault location probability map, and the bearing fault location probability is increased to 0.82, while the probabilities of other locations are correspondingly reduced.

[0141] Generate a fault diagnosis report, including the fault type (bearing fault type A1), fault location (front rotor bearing), fault probability (82%), fault cause analysis (caused by insufficient bearing lubrication), and maintenance suggestions (replace the bearing and check the lubrication system). The maintenance personnel successfully eliminated the fault according to the report guidance, verifying the effectiveness and accuracy of this method.

[0142] Figure 5 This is a bar chart for the performance comparison and analysis of the cognitive enhanced fault diagnosis method in the embodiments of the present invention:

[0143] The chart compares and analyzes the recognition accuracy performance of three different analysis methods in three test scenarios. Specifically, in the aspect of identifying internal faults in transformers, the accuracy rate of the traditional analysis method is 64.5%, the Euclidean distance method is improved to 76.3%, while the cognitive enhancement method reaches the highest accuracy rate of 91.8%. In the scenario of evaluating the status of coil groups, the accuracy rate of the traditional analysis method is 68.2%, the Euclidean distance method is increased to 78.6%, and the cognitive enhancement method achieves an excellent performance of 89.5%. In the test of detecting abnormal resonance frequencies, the accuracy rate of the traditional analysis method is 71.9%, the Euclidean distance method is significantly improved to 82.2%, and the cognitive enhancement method even reaches the best effect of 96.2%. From the overall trend, the accuracy rate ranking of the three methods in all test scenarios is cognitive enhancement method > Euclidean distance method > traditional analysis method, and the cognitive enhancement method has a significant performance advantage compared with the other two methods, with an average improvement of more than 20 percentage points, demonstrating a powerful analysis ability.

[0144] In an optional implementation manner, a fault probability distribution is calculated based on the optimized cognitive weights. The fault probability distribution adjusts the discrimination degree of fault types through a discrimination sensitivity parameter. Combining the fault probability distribution with a fault space distribution function to obtain a fault location probability mapping includes:

[0145] Calculate a fault probability distribution based on the optimized cognitive weights, and adjust the discrimination degree of fault types through a discrimination sensitivity parameter. The discrimination sensitivity parameter is adaptively adjusted according to the degree of difference in fault characteristics;

[0146] Perform a weighted combination of the fault probability distribution and the fault space distribution function to obtain a fault location probability mapping. The fault space distribution function describes the spatial distribution characteristics of fault types, and the fault location probability mapping reflects the spatial probability distribution of the fault occurrence location.

[0147] Obtain optimized cognitive weights. These weights are usually optimized based on historical fault data and expert experience, reflecting the importance of different fault characteristics. For example, for a certain power equipment monitoring system, the cognitive weight of abnormal temperature is 0.35, the cognitive weight of abnormal vibration is 0.28, the cognitive weight of abnormal sound is 0.22, and the cognitive weight of abnormal current is 0.15. These weights reflect the relative importance of each characteristic for fault judgment.

[0148] Based on these optimized cognitive weights, the system calculates the fault probability distribution. The fault probability distribution represents the probabilities of various fault types occurring under given sample features. Specifically, the system calculates the probability value Pk for each fault type k. For example, for a certain set of equipment operation data collected, the system calculates that the probability of bearing fault is 0.62, the probability of gear fault is 0.25, the probability of lubrication system fault is 0.08, and the probability of other faults is 0.05. The calculation of these probability values fully considers the cognitive weights of each fault feature.

[0149] During the process of calculating the fault probability distribution, the system introduces a discrimination sensitivity parameter to adjust the discrimination degree of fault types. The role of this parameter is to adaptively adjust according to the degree of difference in fault features, making the probability distribution gap between significantly different fault types larger, thereby improving the accuracy of fault diagnosis. The discrimination sensitivity parameter can be dynamically adjusted according to the variance or entropy value of the input features. In practical applications, the system can set a basic sensitivity value, such as 1.5, and then adjust it according to the degree of difference in fault features. When the detected feature differences are obvious, the sensitivity parameter increases to 2.2; while when the feature differences are small, the sensitivity parameter decreases to 1.2.

[0150] Taking equipment vibration monitoring as an example, when the vibration spectra at different positions of the equipment show obvious differences (such as the frequency peak difference exceeds 30%), the system increases the discrimination sensitivity parameter to 2.0; while when the vibration spectra differences are small (such as the frequency peak difference is lower than 10%), the discrimination sensitivity parameter decreases to 1.3. This adaptive adjustment ensures reasonable discrimination of fault types under different fault feature difference conditions.

[0151] Based on the adjusted fault probability distribution, the system needs to further consider the distribution characteristics of faults in terms of spatial location. The system introduces a fault spatial distribution function, which describes the distribution characteristics of various fault types in the equipment space. For example, for a large industrial equipment, bearing faults mostly occur at the connection points of rotating parts of the equipment, gear faults are mainly concentrated in the transmission box, and lubrication system faults are distributed throughout the lubrication pipeline network.

[0152] The fault spatial distribution function can be obtained through historical fault data statistics, represented as a correspondence between spatial location and fault probability. For example, for 10 key monitoring points of a certain equipment, the system obtains that the probability of bearing fault at location 1 is 0.35, the probability at location 2 is 0.28, and so on. This spatial distribution feature is crucial for accurately locating the fault position.

[0153] The failure probability distribution and the failure space distribution function are weighted and combined to generate a failure location probability map. In this step, the system assigns appropriate weights to the failure probability distribution and the space distribution function, such as 0.6 and 0.4, and then obtains the final failure location probability map through weighted summation.

[0154] Taking an industrial device as an example, assume that the system calculates the bearing failure probability to be 0.62, and the space distribution probability of this failure at position 3 is 0.45. Then the joint probability of a bearing failure occurring at position 3 is 0.62×0.6 + 0.45×0.4 = 0.552. The system performs this calculation for all combinations of failure types and positions, and finally generates a comprehensive failure location probability map.

[0155] This failure location probability map intuitively reflects the spatial probability distribution of the failure occurrence location, providing accurate failure location information for maintenance personnel. For example, the map shows that there is a 55.2% probability of a failure occurring at position 3 of the device, a 23.8% probability of a failure occurring at position 7, and the failure probabilities at other positions are relatively low. Maintenance personnel can prioritize inspections of high-risk positions based on this, improving maintenance efficiency.

[0156] In practical applications, this method has been verified in the equipment predictive maintenance system of a large manufacturing enterprise. Through this method, the failure location accuracy rate of the system has been improved from the original 78% to 91%, and the average failure diagnosis time has been reduced from 4 hours to 1.5 hours, significantly improving maintenance efficiency and reducing equipment downtime.

[0157] In the second aspect of the embodiments of the present invention, an automated measurement and analysis system for the impedance characteristics of a reactor is provided, including:

[0158] A first unit for collecting voltage signals and current signals through a sampling device of the reactor, and calculating the original impedance frequency characteristic data of the reactor according to the voltage signals and the current signals;

[0159] A second unit for performing frequency-domain decomposition on the original impedance frequency characteristic data to obtain fundamental wave components and harmonic components, respectively performing Fourier transform on the fundamental wave components and the harmonic components to obtain spectrum data; inputting the spectrum data into a preset neural network compensation model to obtain compensation coefficients, and multiplying the compensation coefficients by the spectrum data to obtain compensated impedance frequency characteristic data, where the neural network compensation model uses a long short-term memory network structure to extract temporal features and perform dynamic compensation on the spectrum data;

[0160] A third unit, configured to establish an impedance-frequency characteristic curve of a reactor according to the compensated impedance-frequency characteristic data, and calculate real-time characteristic data of the impedance-frequency characteristic curve at different frequency points; continuously collect real-time monitoring data of the reactor under different operating conditions through the sampling device, calculate a real-time impedance-frequency characteristic according to the real-time monitoring data, and compare the real-time impedance-frequency characteristic with the real-time characteristic data to obtain a characteristic deviation;

[0161] A fourth unit, configured to construct a fault characteristic vector according to the characteristic deviation, establish a fault discrimination criterion by calculating the Euclidean distance and Mahalanobis distance of the fault characteristic vector, determine the fault type and fault location according to the fault discrimination criterion, and generate a fault diagnosis report.

[0162] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including:

[0163] A processor;

[0164] A memory for storing instructions executable by the processor;

[0165] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0166] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0167] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are carried.

[0168] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic measurement and analysis method for the impedance characteristics of a reactor, characterized in that, Including: Sampling devices collect voltage signals and current signals through reactors, and calculate the original impedance frequency characteristic data of the reactors according to the voltage signals and the current signals; Perform frequency-domain decomposition on the original impedance frequency characteristic data to obtain fundamental wave components and harmonic components, perform Fourier transform on the fundamental wave components and the harmonic components respectively to obtain spectrum data; input the spectrum data into a preset neural network compensation model to obtain compensation coefficients, multiply the compensation coefficients by the spectrum data to obtain compensated impedance frequency characteristic data, and the neural network compensation model uses a long short-term memory network structure to extract time series features and perform dynamic compensation on the spectrum data; Establish an impedance frequency characteristic curve of the reactor according to the compensated impedance frequency characteristic data, and calculate real-time characteristic data at different frequency points of the impedance frequency characteristic curve; continuously collect real-time monitoring data of the reactor under different operating conditions through the sampling device, calculate the real-time impedance frequency characteristic according to the real-time monitoring data, and compare the real-time impedance frequency characteristic with the real-time characteristic data to obtain a characteristic deviation; Construct a fault characteristic vector according to the characteristic deviation, establish a fault discrimination criterion by calculating the Euclidean distance and Mahalanobis distance of the fault characteristic vector, determine the fault type and fault location according to the fault discrimination criterion, and generate a fault diagnosis report.

2. The method according to claim 1, wherein Performing frequency-domain decomposition on the original impedance frequency characteristic data to obtain fundamental wave components and harmonic components, and performing Fourier transform on the fundamental wave components and the harmonic components respectively to obtain spectrum data includes: Perform empirical mode decomposition on the original impedance frequency characteristic data to obtain intrinsic mode functions and a residual term, identify the index set of intrinsic mode functions within the fundamental wave frequency range and the index set of intrinsic mode functions within the harmonic frequency range according to the frequency characteristics of the intrinsic mode functions, and superimpose the index set of intrinsic mode functions within the fundamental wave frequency range and the index set of intrinsic mode functions within the harmonic frequency range to obtain fundamental wave components and harmonic components; Perform short-time Fourier transform on the fundamental wave components and the harmonic components respectively to obtain a time-frequency spectrum, and the short-time Fourier transform uses a variable-length Kaiser window function, and the window length and shape parameters of the variable-length Kaiser window function are adaptively adjusted according to the non-stationary characteristics of the signal; Perform feature enhancement processing on the time-frequency spectrum, and the feature enhancement processing includes spectrum smoothing and noise suppression. Convolve the time-frequency spectrum in the frequency dimension through a smoothing kernel function to obtain a smoothed spectrum, perform wavelet transform on the smoothed spectrum, and remove noise components through a soft threshold function to obtain enhanced spectrum data.

3. The method according to claim 1, characterized in that, Establishing an impedance frequency characteristic curve of the reactor according to the compensated impedance frequency characteristic data, and calculating the real-time characteristic data at different frequency points of the impedance frequency characteristic curve includes: Construct an immune network model based on the compensated impedance frequency characteristic data, construct an impedance characteristic curve according to the antibody memory unit of the immune network model, and amplify and mutate the antibody memory unit through the clone selection algorithm for the impedance characteristic curve. The clone selection algorithm includes two sub-steps: affinity calculation and clone amplification, and optimize the fitting accuracy of the impedance characteristic curve in each frequency interval through the clone selection algorithm; Calculate the first derivative and the second derivative of the impedance amplitude at different frequency points of the impedance characteristic curve, and re-input the first derivative and the second derivative into the immune network model to update the impedance change rate information stored in the antibody memory unit; Calculate the impedance distortion degree of each frequency interval according to the updated impedance change rate information. The impedance distortion degree judges the non-linear change of the impedance characteristic curve through the clone selection algorithm. When the impedance distortion degree exceeds the preset distortion threshold, the clone selection algorithm triggers a high-frequency mutation operation to optimize the real-time characteristic data of the corresponding frequency interval by generating a new antibody population.

4. The method according to claim 3, wherein Construct an immune network model based on the compensated impedance frequency characteristic data, and constructing an impedance characteristic curve according to the antibody memory unit of the immune network model. The amplification and mutation of the antibody memory unit by the clone selection algorithm for the impedance characteristic curve include: The immune network model includes multiple antibody units, each antibody unit corresponding to the impedance frequency characteristic data of a frequency point. The affinity between the antibody units is calculated by a Gaussian kernel function, and the affinity characterizes the similarity of the impedance frequency characteristic data at adjacent frequency points; Calculate the network activation degree of the antibody unit based on the affinity. The network activation degree is obtained by weighted accumulation of the affinity, and the weight coefficient is determined by the antibody memory factor. The impedance frequency characteristic data corresponding to the antibody unit with the network activation degree higher than the preset activation degree threshold is preferentially stored in the antibody memory unit; Perform clone amplification on the impedance frequency characteristic data stored in the antibody memory unit. The number of clones is proportional to the network activation degree. The newly generated antibody units by cloning inherit the original impedance frequency characteristic data, and perform an affinity maturation operation on the new antibody units. The affinity maturation operation perturbs the impedance frequency characteristic data through Gaussian random mutation, and the mutation amplitude is inversely proportional to the network activation degree; Re-input the mutated new antibody units into the immune network model to calculate the updated network activation degree, screen the antibody memory unit according to the updated network activation degree, retain the antibody memory unit with the network activation degree higher than the preset activation degree threshold, and construct an impedance characteristic curve.

5. The method according to claim 1, wherein Calculate the real-time impedance frequency characteristic according to the real-time monitoring data, and compare the real-time impedance frequency characteristic with the real-time characteristic data to obtain the characteristic deviation, including: Perform recursive least squares estimation on the real-time monitoring data, including: constructing an observation equation containing voltage observation values, current observation values, and initial impedance parameters, and calculating a prediction error based on the observation equation, where the prediction error is the estimation deviation between the real-time monitoring data and historical impedance parameters; Calculate a gain matrix according to the prediction error, where the gain matrix is used to adjust the parameter update step size, multiply the prediction error by the gain matrix to obtain a parameter update amount, and add the parameter update amount to the historical impedance parameters to obtain the impedance parameters at the current moment; Perform Kalman filtering on the impedance parameters, calculate the state prediction value of the impedance parameters, and perform smoothing processing based on the deviation between the state prediction value and the real-time monitoring data to obtain a filtered impedance state estimation value; Calculate the real-time impedance frequency characteristics based on the impedance state estimation value, where the real-time impedance frequency characteristics include amplitude characteristics, phase characteristics, and quality factor; compare the real-time impedance frequency characteristics with the real-time characteristic data to obtain the final characteristic deviation.

6. The method according to claim 1, wherein Construct a fault feature vector according to the characteristic deviation, establish a fault discrimination criterion by calculating the Euclidean distance and Mahalanobis distance of the fault feature vector, and determine the fault type and fault location according to the fault discrimination criterion, and generate a fault diagnosis report including: Fuse the fault feature vector with the expert experience feature vector, weight the fault feature vector through a cognitive weight matrix to obtain an empirical fusion coefficient; weight the expert experience feature vector through the empirical fusion coefficient to obtain a cognitive enhanced feature vector; Calculate the Euclidean distance and Mahalanobis distance of the cognitive enhanced feature vector, where the Euclidean distance represents the direct distance between feature vectors, and the Mahalanobis distance considers the covariance relationship of feature distributions, and construct a fault discrimination criterion based on the Euclidean distance and the Mahalanobis distance; Perform cognitive weight optimization on the fault discrimination criterion, input the Euclidean distance and the Mahalanobis distance into a loss function, introduce a cognitive rule constraint term, and obtain the optimized cognitive weight by minimizing the weighted sum of the loss function and the cognitive rule constraint term; Calculate the fault probability distribution based on the optimized cognitive weight, where the fault probability distribution adjusts the discrimination degree of fault types through a discrimination sensitivity parameter, and combine the fault probability distribution with the fault space distribution function to obtain a fault location probability mapping; Update the expert experience knowledge using the reinforcement learning value function, where the expert experience knowledge is balanced between historical experience and the reinforcement learning value function through a learning rate, and optimize the fault location probability mapping based on the updated expert experience knowledge to generate a fault diagnosis report.

7. The method according to claim 6, wherein Calculate the fault probability distribution based on the optimized cognitive weight, where the fault probability distribution adjusts the discrimination degree of fault types through a discrimination sensitivity parameter, and combine the fault probability distribution with the fault space distribution function to obtain a fault location probability mapping including: Calculate the fault probability distribution based on the optimized cognitive weight, adjust the discrimination degree of fault types through a discrimination sensitivity parameter, and the discrimination sensitivity parameter is adaptively adjusted according to the degree of fault feature difference; The failure probability distribution is weighted and combined with the failure space distribution function to obtain a failure location probability mapping. The failure space distribution function describes the spatial distribution characteristics of the failure type, and the failure location probability mapping reflects the spatial probability distribution of the failure occurrence location.

8. A reactor impedance characteristic automatic measurement and analysis system for implementing the method according to any one of the preceding claims 1-7, characterized in that, It includes: The first unit is configured to collect voltage signals and current signals through a sampling device of a reactor, and calculate the original impedance frequency characteristic data of the reactor according to the voltage signals and the current signals; The second unit is configured to perform frequency-domain decomposition on the original impedance frequency characteristic data to obtain fundamental wave components and harmonic components, and perform Fourier transforms on the fundamental wave components and the harmonic components respectively to obtain spectrum data; Input the spectrum data into a preset neural network compensation model to obtain compensation coefficients, and multiply the compensation coefficients by the spectrum data to obtain compensated impedance frequency characteristic data. The neural network compensation model uses a long short-term memory network structure to extract temporal features and perform dynamic compensation on the spectrum data; The third unit is configured to establish an impedance frequency characteristic curve of the reactor according to the compensated impedance frequency characteristic data, and calculate real-time characteristic data at different frequency points of the impedance frequency characteristic curve; continuously collect real-time monitoring data of the reactor under different operating conditions through the sampling device, calculate the real-time impedance frequency characteristic according to the real-time monitoring data, and compare the real-time impedance frequency characteristic with the real-time characteristic data to obtain a characteristic deviation; The fourth unit is configured to construct a failure feature vector according to the characteristic deviation, establish a failure discrimination criterion by calculating the Euclidean distance and Mahalanobis distance of the failure feature vector, determine the failure type and failure location according to the failure discrimination criterion, and generate a failure diagnosis report.

9. An electronic device, characterized in that, It includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images