A cascaded module non-intrusive monitoring method and system

Abstract

The present invention discloses a cascaded module non-invasive monitoring method and system, which acquires the voltage signal of the DC bus within a preset time window length to obtain a voltage fluctuation time-domain signal, performs signal conversion and extraction on the voltage fluctuation time-domain signal by using fast Fourier transform to obtain a plurality of main frequency components, obtains capacitance attenuation degree data according to each of the main frequency components, and then determines whether the coupling strength index is related to the capacitance attenuation degree according to the voltage fluctuation data of adjacent sub-modules of the DC bus. If they are related, a prediction neural network is used to extract features from the fluctuation coupling information to obtain a monitoring feature vector, an optimized non-invasive monitoring model is obtained by using the monitoring feature vector, and the optimized non-invasive monitoring model is used to analyze the voltage fluctuation data of the DC bus to obtain the capacitance attenuation degree. Through the above method, the accuracy and reliability of capacitance attenuation monitoring are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cascaded converter monitoring, and particularly to a non-invasive monitoring method and system for cascaded modules. Background Art

[0002] In a cascaded converter, multiple sub-modules are connected in series to form a DC bus voltage. When the electrolytic capacitor of a certain sub-module undergoes parameter attenuation, the DC-side voltage of this sub-module will fluctuate, and then the output voltage will generate a ripple with a characteristic frequency. However, this ripple will be coupled with adjacent normal sub-modules through the DC bus, causing corresponding frequency fluctuations in their voltages. This ripple coupling effect between adjacent modules will seriously interfere with the method based on single sub-module monitoring, thus affecting the monitoring accuracy. Specifically, the sub-module with parameter attenuation will generate a characteristic frequency component with a large amplitude in its output voltage. However, after the coupling effect of the DC bus, these characteristic frequency components will be absorbed to a certain extent by adjacent normal sub-modules and present similar spectral characteristics. Therefore, it is difficult to accurately judge the location and attenuation degree of the faulty module only through the spectral analysis of the voltage of a single sub-module. At the same time, as the number of cascaded sub-modules increases, the propagation and coupling process of the ripple in the DC bus becomes more complex, further reducing the monitoring accuracy. This requires that in the non-invasive monitoring method, the ripple coupling effect between sub-modules must be comprehensively considered, and through a reasonable mathematical model and algorithm, the accurate positioning and quantitative analysis of the faulty sub-module can be achieved to ensure the reliable operation of the converter. Summary of the Invention

[0003] To solve the above technical problems, an embodiment of the present invention provides a non-invasive monitoring method and system for cascaded modules to solve the technical problem of low accuracy in monitoring the capacitance attenuation degree due to the ripple coupling effect existing between sub-modules.

[0004] The first aspect of the embodiment of the present invention provides a non-invasive monitoring method for cascaded modules, and the method includes:

[0005] Obtain the voltage signal of the DC bus within a preset time window length to obtain a voltage fluctuation time-domain signal, perform signal conversion and extraction on the voltage fluctuation time-domain signal by using fast Fourier transform to obtain a plurality of main frequency components, and obtain a harmonic feature sequence according to each main frequency component;

[0006] Obtain the voltage waveform of the sub-module according to the first preset sampling frequency, perform time-frequency domain conversion on the voltage waveform by using Fourier transform to obtain a frequency-domain spectrum line, perform characteristic frequency extraction according to the frequency-domain spectrum line to obtain a plurality of characteristic frequency points, extract the capacitance attenuation characteristic frequency points based on the amplitudes corresponding to each characteristic frequency point and the capacitance attenuation characteristic library, and obtain an amplitude change trend sequence according to the amplitudes corresponding to the capacitance attenuation characteristic frequency points.

[0007] Obtain the fitting curve of the coupling strength index and the capacitance attenuation degree from the capacitance attenuation characteristic frequency points, the amplitude change trend sequence, and the voltage fluctuation data of adjacent sub-modules of the DC bus. Obtain the correlation of the coupling strength index with the capacitance attenuation degree according to the fitting curve. Based on the correlation, judge whether the coupling strength index is related to the capacitance attenuation degree. If it is related, calculate the error prediction sequence and the confidence level of the monitoring process. Analyze the voltage fluctuation data of the DC bus according to the error prediction sequence and the confidence level to obtain the monitoring result, where the monitoring result is the capacitance attenuation degree.

[0008] In a possible implementation manner of the first aspect, the voltage fluctuation time-domain signal is converted and extracted by using the fast Fourier transform to obtain multiple main frequency components. According to each main frequency component, a harmonic feature sequence is obtained, including:

[0009] Obtain the voltage signal of the DC bus in a preset time window length to obtain multiple voltage data. Calculate the fluctuation amplitude sequence according to the preset reference voltage value and each voltage data to obtain the voltage fluctuation time-domain signal;

[0010] Perform segmented windowing processing on the voltage fluctuation time-domain signal, and use a preset window function to smooth the signal to obtain the smoothed voltage fluctuation sequence;

[0011] Perform Fourier transform on the smoothed voltage fluctuation sequence to obtain the frequency-domain representation sequence. Extract spectral lines from the frequency-domain representation sequence according to the preset frequency-domain analysis range to obtain multiple main frequency components. Group according to the frequency values of each main frequency component to determine the fundamental frequency component and multiple sub-harmonic frequency components;

[0012] Calculate the ratio of the amplitude of each sub-harmonic frequency component to the fundamental frequency component respectively to obtain the harmonic component content ratio sequence. Determine the sub-harmonic frequency components greater than the preset harmonic limit value in the harmonic component content ratio sequence as the over-limit harmonic components to obtain the harmonic feature sequence.

[0013] In a possible implementation manner of the first aspect, perform characteristic frequency extraction according to the frequency-domain spectral lines to obtain multiple characteristic frequency points. Based on the amplitude corresponding to each characteristic frequency point and the capacitance attenuation characteristic library, extract the capacitance attenuation characteristic frequency points. According to the capacitance attenuation characteristic frequency points, obtain the corresponding amplitude change trend sequence, including:

[0014] Extract the frequency-domain spectral lines to obtain the harmonic frequency sequence and the harmonic amplitude sequence;

[0015] Classify the harmonic frequency sequence according to the preset frequency interval to obtain multiple harmonic frequency interval sequences, calculate the amplitude proportion of the harmonic frequency components in each harmonic frequency interval sequence to obtain the harmonic frequency distribution characteristic sequence, and select the frequency points with the amplitude proportion greater than the preset proportion in the harmonic frequency distribution characteristic sequence as the characteristic frequency points to obtain multiple characteristic frequency points;

[0016] Based on the amplitude corresponding to each characteristic frequency point, calculate the amplitude change rate using a fixed-length sliding window to obtain the amplitude change rate sequence, and take the amplitude greater than the preset change rate threshold in the amplitude change rate sequence as the amplitude mutation point;

[0017] Calculate the spectral similarity within a preset range of the amplitude mutation point, judge the similarity threshold to obtain the spectral change type, match the spectral change type with the capacitance attenuation feature library, calculate the spectral correlation degree sequence through the matching degree, extract the capacitance attenuation characteristic frequency points according to the spectral correlation degree sequence, and record the amplitude change trend sequence corresponding to the capacitance attenuation characteristic frequency points.

[0018] Combine the capacitance attenuation characteristic frequency points, the amplitude change trend sequence, and the voltage fluctuation data of the adjacent sub-modules of the DC bus to obtain the fitting curve of the coupling strength index and the capacitance attenuation data, and obtain the correlation between the coupling strength index and the capacitance attenuation degree according to the fitting curve, including:

[0019] Perform normalization processing on the capacitance attenuation characteristic frequency points and the amplitude change trend sequence to obtain the normalized feature matrix, and use the principal component dimensionality reduction method to extract features from the normalized feature matrix to obtain the dimensionality-reduced feature vector;

[0020] Input the dimensionality-reduced feature vector into the trained vector machine model for regression prediction to obtain the capacitance attenuation degree data, where the capacitance attenuation degree data includes the capacitance loss prediction value and the remaining life prediction value;

[0021] According to the voltage fluctuation data of the adjacent sub-modules of the DC bus and the capacitance attenuation degree data, obtain the fitting curve of the coupling strength index and the capacitance attenuation data, and obtain the correlation between the coupling strength index and the capacitance attenuation degree according to the fitting curve.

[0022] In a possible implementation manner of the first aspect, according to the voltage fluctuation data of the adjacent sub-modules of the DC bus and the capacitance attenuation degree data, obtain the fitting curve of the coupling strength index and the capacitance attenuation data, and obtain the correlation between the coupling strength index and the capacitance attenuation degree according to the fitting curve, including:

[0023] Collect the voltage fluctuation data of adjacent sub-modules of the DC bus, align the voltage fluctuation data using a synchronous sampling clock to obtain the aligned fluctuation data, and collect the aligned fluctuation data according to the second preset sampling frequency to obtain the complete cycle waveform data;

[0024] Segment the complete cycle waveform data using a fixed-length sliding window to obtain segmented data, perform Fourier transform on the segmented data to obtain the frequency domain representation sequence, extract multiple characteristic harmonic frequency points from the frequency domain representation sequence, calculate the frequency offset between adjacent sub-modules according to each characteristic harmonic frequency point to obtain the frequency deviation sequence, and use the characteristic harmonic frequency points with frequency offset greater than the preset frequency threshold in the frequency deviation sequence as abnormal frequency points;

[0025] According to the abnormal frequency points, calculate the characteristic harmonic amplitude ratio between adjacent sub-modules, calculate the waveform correlation coefficient using adjacent window data to obtain the coupling characteristic sequence, calculate the average value of the waveform correlation coefficient sequence to obtain the coupling strength index, perform polynomial fitting on the coupling strength index and the capacitor attenuation degree data to obtain the fitting curve, and extract the fitting curve to obtain the correlation between the coupling strength index and the capacitor attenuation degree.

[0026] In a possible implementation manner of the first aspect, calculating the error prediction sequence and confidence level of the monitoring process includes:

[0027] Use a prediction neural network to extract features from the fluctuation coupling information to obtain a monitoring feature vector;

[0028] Input the monitoring feature vector into the constructed correlation model of coupling relationship and monitoring accuracy for processing to obtain the error prediction sequence and confidence level of the monitoring process.

[0029] In a possible implementation manner of the first aspect, using a prediction neural network to extract features from the fluctuation coupling information to obtain a monitoring feature vector includes:

[0030] Normalize the fluctuation coupling information to obtain a standardized feature matrix;

[0031] Use a processor to perform convolution operation on the standardized feature matrix to obtain a coupling feature map;

[0032] Perform a second-layer convolution operation on the coupling feature map to obtain a propagation feature map, use pooling operation to reduce the dimension of the propagation feature map to obtain a reduced-dimension feature, and use a fully connected layer to perform mapping transformation on the reduced-dimension feature to obtain a monitoring feature vector.

[0033] In a possible implementation manner of the first aspect, the prediction neural network is trained by using sample fluctuation coupling information, including:

[0034] Obtain the sample fluctuation coupling information, input the sample fluctuation coupling information into the initial prediction neural network for prediction, and obtain the initial monitoring result;

[0035] Calculate the mean square error of the initial monitoring result to obtain a monitoring error sequence, perform probability distribution fitting on the monitoring error sequence, and obtain the standard deviation and mean parameters;

[0036] Calculate the monitoring confidence interval according to the standard deviation and mean parameters to obtain the monitoring accuracy evaluation result;

[0037] Optimize the parameters of the initial prediction neural network according to the monitoring accuracy evaluation result and the monitoring error sequence, update the feature extraction parameters and prediction parameters, and obtain the prediction neural network.

[0038] In a possible implementation manner of the first aspect, analyze the voltage fluctuation data of the DC bus according to the error prediction sequence and the confidence level to obtain the monitoring result, including:

[0039] Calculate the filter center frequency and bandwidth parameters according to the error prediction sequence and the confidence level;

[0040] Use the filter center frequency parameter and bandwidth parameter to filter the bus voltage fluctuation data to obtain the filtered data, extract the filtered data, and obtain the multi-band fluctuation feature sequence;

[0041] Perform wavelet multi-scale decomposition on the multi-band fluctuation feature sequence to obtain the harmonic feature sequence, and use a support vector machine to perform classification operations on the harmonic feature sequence to obtain the attenuation feature vector;

[0042] Calculate the capacitance attenuation quantization index according to the attenuation feature vector, perform normalization processing on the capacitance attenuation index to obtain the correction result sequence, compare the correction result sequence with the original capacitance attenuation degree to obtain the correction deviation, and update the filter parameters according to the correction deviation to obtain the optimized non-intrusive monitoring model;

[0043] Use the optimized non-intrusive monitoring model to analyze the voltage fluctuation data of the DC bus to obtain the monitoring result.

[0044] To solve the same technical problem, a second aspect of the embodiments of the present invention provides a cascaded module non-intrusive monitoring system, the system includes:

[0045] An acquisition module, configured to acquire the voltage signal of the DC bus in a preset time window length to obtain a voltage fluctuation time-domain signal, perform signal conversion and extraction on the voltage fluctuation time-domain signal by using fast Fourier transform to obtain a plurality of main frequency components, and obtain a harmonic feature sequence according to each main frequency component;

[0046] The first calculation module is used to obtain the voltage waveform of the sub-module according to the first preset sampling frequency, perform time-frequency domain conversion on the voltage waveform by using Fourier transform to obtain frequency domain spectral lines, extract characteristic frequencies based on the frequency domain spectral lines to obtain multiple characteristic frequency points, extract the capacitance attenuation characteristic frequency points based on the amplitudes corresponding to each characteristic frequency point and the capacitance attenuation characteristic library, and obtain the amplitude change trend sequence according to the amplitude corresponding to the capacitance attenuation characteristic frequency point;

[0047] The second calculation module is used to obtain the fitting curve of the coupling strength index and the capacitance attenuation degree from the capacitance attenuation characteristic frequency points, the amplitude change trend sequence, and the voltage fluctuation data of the adjacent sub-modules of the DC bus, obtain the correlation of the coupling strength index with respect to the capacitance attenuation degree according to the fitting curve, and based on the correlation, determine whether the coupling strength index is related to the capacitance attenuation degree. If it is related, calculate the error prediction sequence and the confidence level of the monitoring process, and analyze the voltage fluctuation data of the DC bus according to the error prediction sequence and the confidence level to obtain the monitoring result, where the monitoring result is the capacitance attenuation degree.

[0048] The technical solution of the present invention has the following advantages:

[0049] The non-intrusive monitoring method for cascaded modules provided by the embodiment of the present invention obtains the voltage signal of the DC bus in the preset time window length to obtain the voltage fluctuation time domain signal, performs signal conversion and extraction on the voltage fluctuation time domain signal by using fast Fourier transform to obtain multiple main frequency components, obtains the capacitance attenuation degree data according to each main frequency component, and then determines whether the coupling strength index is related to the capacitance attenuation degree according to the voltage fluctuation data of the adjacent sub-modules of the DC bus. If it is related, uses a prediction neural network to extract features from the fluctuation coupling information to obtain a monitoring feature vector, uses the monitoring feature vector to obtain an optimized non-intrusive monitoring model, and analyzes the voltage fluctuation data of the DC bus by using the optimized non-intrusive monitoring model to obtain the capacitance attenuation degree. Through the above method, the accuracy and reliability of capacitance attenuation monitoring are improved. Description of the Drawings

[0050] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0051] Figure 1 It is a flowchart of the non-intrusive monitoring method for cascaded modules in the embodiment of the present invention;

[0052] Figure 2 This is the structural block diagram of the cascaded module non-intrusive monitoring system in the embodiment of the present invention. Specific implementation mode

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] In the description of the present invention, it should be noted that the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0055] The non-intrusive monitoring method for cascaded modules provided by the embodiments of the present invention, as Figure 1 shown, Figure 1 is the non-intrusive monitoring flowchart for cascaded modules, including steps S101 to S105. The specific steps are as follows:

[0056] S101: Obtain the voltage signal of the DC bus within a preset time window length to obtain the voltage fluctuation time-domain signal. Use the fast Fourier transform to perform signal conversion and extraction on the voltage fluctuation time-domain signal to obtain multiple main frequency components. According to each of the main frequency components, obtain the harmonic feature sequence.

[0057] In this embodiment, voltage sampling points on the DC bus are collected, and a complete waveform segment is intercepted from the voltage sampling points according to the preset time window length to obtain the voltage fluctuation time-domain signal. Specifically, voltage sampling points at multiple time intervals are collected for the DC bus, a complete waveform segment is intercepted from the sampling signal according to the preset time window length, and the fluctuation amplitude sequence is calculated by comparing with the preset reference voltage value to generate the voltage fluctuation time-domain signal.

[0058] In one embodiment, using the fast Fourier transform to perform signal conversion and extraction on the voltage fluctuation time-domain signal to obtain multiple main frequency components, and obtaining the harmonic feature sequence according to each main frequency component includes:

[0059] Obtain the voltage signal of the DC bus within a preset time window length to obtain multiple voltage data, calculate the fluctuation amplitude sequence according to the preset reference voltage value and each voltage data to obtain the voltage fluctuation time-domain signal;

[0060] Perform segmented windowing processing on the voltage fluctuation time-domain signal, and use a preset window function to smooth the signal to obtain the smoothed voltage fluctuation sequence;

[0061] Perform a Fourier transform on the smoothed voltage fluctuation sequence to obtain a frequency-domain representation sequence. Extract spectral lines from the frequency-domain representation sequence according to a preset frequency-domain analysis range to obtain multiple main frequency components. Group the main frequency components according to their frequency values to determine the fundamental frequency component and multiple sub-harmonic frequency components;

[0062] Calculate the ratio of the amplitude of each sub-harmonic frequency component to the fundamental frequency component respectively to obtain a harmonic component content ratio sequence. Determine the sub-harmonic frequency components in the harmonic component content ratio sequence that are greater than a preset harmonic limit value as over-limit harmonic components to obtain a harmonic feature sequence.

[0063] In this embodiment, voltage sampling points on the DC bus are collected. According to a preset time window length, a complete waveform segment is intercepted from the voltage sampling points to obtain a voltage fluctuation time-domain signal. The voltage fluctuation time-domain signal is subjected to segmented windowing processing, and a preset window function is used to smooth the signal to generate a smoothed voltage fluctuation sequence. Perform a Fourier transform on the smoothed voltage fluctuation sequence to obtain a frequency-domain representation sequence. Determine the frequency-domain analysis range according to the sampling theorem, extract spectral lines from the frequency-domain representation sequence within the frequency-domain analysis range, calculate the amplitude of each frequency point, and obtain the top percentage of main frequency components sorted by amplitude from large to small.

[0064] Then group the main frequency components according to the frequency numerical relationship to determine the fundamental frequency component and each sub-harmonic frequency component. For each sub-harmonic frequency component, calculate the ratio of its amplitude to the amplitude of the fundamental component to obtain a harmonic component content ratio sequence. Compare the content ratio sequence with a preset harmonic limit value to mark the over-limit harmonic components and generate a harmonic feature sequence.

[0065] As an example of this embodiment, during DC bus voltage sampling, the voltage signal is discretized by selecting an appropriate sampling interval. For the original voltage values collected by the measuring device at the bus end side, 8000 data points are collected per second. On the basis of forming a fluctuation sequence, a preset reference voltage value is set to 500 volts. When the fluctuation of the measured voltage value exceeds the range of plus or minus 5% of the reference value, record the corresponding sampling points and their moments.

[0066] During the process of processing the DC bus voltage, for the measured voltage fluctuation time-domain signal, the time window width is set to 0.2 seconds, and a window function is used to achieve smoothing processing. The selection of the window function directly affects the accuracy of spectral analysis. When the set window function width is too large, the smoothed signal is distorted, while when the window function width is too small, the measurement noise interference cannot be suppressed. Apply a fast Fourier transform to the smoothed voltage fluctuation sequence. According to the sampling theorem, the frequency-domain analysis range is determined to be 4000 Hz. The time-domain signal is transformed into the frequency-domain space through Fourier transform to obtain a frequency-domain representation sequence containing amplitude spectrum and phase spectrum information.

[0067] In the frequency domain space, spectral line identification and extraction are performed on the obtained spectrum. The spectral line amplitude screening threshold is set to 3% of the fundamental wave amplitude, and the main frequency components are extracted. From the main frequency components obtained through frequency domain analysis, grouping is carried out according to the frequency magnitude relationship. In an AC power supply system, the fundamental wave frequency is 50 Hz, and the harmonic order is determined by the multiple relationship between other frequency components and the fundamental wave frequency. For example, 100 Hz is the second harmonic component, and 150 Hz is the third harmonic component. When non-integer multiple frequency components appear, they may be caused by subsynchronous oscillations or equipment failures in the power system. For the identified harmonic components, the content ratio is obtained by calculating the ratio of the harmonic amplitude to the fundamental wave amplitude. The amplitude of the fundamental wave component is 12 V, the amplitude of the second harmonic is 0.6 V, and the amplitude of the third harmonic is 0.36 V. Then the second harmonic content is 5%, and the third harmonic content is 3%. According to the power system harmonic limit standard, the second harmonic content limit is 4%, and the third harmonic content limit is 3%. By comparison, it is found that the second harmonic exceeds the limit, and marking and corresponding suppression measures need to be taken. In the spectral analysis of bus voltage fluctuations, by performing segmented windowing and frequency domain conversion on the voltage fluctuation signal, the extraction and analysis of each harmonic component are realized. From the original voltage fluctuation data obtained by sampling, the fundamental wave frequency and harmonic frequency components such as the second and third harmonics are identified, the proportion of each harmonic content is analyzed, and the degree of harmonic pollution is evaluated. The harmonic components exceeding the limit requirements are marked to provide a basis for subsequent harmonic suppression and power quality improvement.

[0068] S102: Obtain the voltage waveform of the sub-module according to the first preset sampling frequency, perform time-frequency domain conversion on the voltage waveform by using Fourier transform to obtain the frequency domain spectral line, extract the characteristic frequencies according to the frequency domain spectral line to obtain multiple characteristic frequency points, and based on the amplitudes corresponding to each characteristic frequency point and the capacitance attenuation characteristic library, extract the capacitance attenuation characteristic frequency points, and obtain the amplitude change trend sequence according to the amplitudes corresponding to the capacitance attenuation characteristic frequency points.

[0069] In this embodiment, obtain the capacitance attenuation characteristic harmonics of the sub-module, analyze the differences between the amplitudes and frequency distributions of the main harmonic components and the capacitance attenuation characteristic harmonics, and determine whether there are characteristic harmonic components related to capacitance attenuation. If so, record the characteristic frequencies and amplitude change trends of the corresponding main harmonic components. Specifically, obtain the voltage waveform from the sub-module by using the first preset sampling frequency, perform time-frequency domain conversion on the voltage waveform by using Fourier transform to obtain the frequency domain spectral line, and extract the harmonic frequency sequence and harmonic amplitude sequence from the frequency domain spectral line. Classify according to the harmonic frequency sequence within the preset frequency range, and calculate the proportion of the harmonic component amplitudes in each frequency range through the harmonic amplitude sequence to obtain the harmonic frequency distribution characteristic sequence.

[0070] Extract the characteristic frequencies from the harmonic frequency distribution characteristic sequence, sort them according to the amplitude proportion within the frequency interval, and select the frequency points with a proportion exceeding the preset threshold as the characteristic frequency points. For the amplitudes corresponding to the characteristic frequency points, calculate the amplitude change rate sequence using a fixed-length sliding window, and identify the amplitude mutation points through the preset change rate threshold.

[0071] Compare the frequency characteristics before and after the amplitude mutation point, calculate the spectral similarity before and after the mutation, and judge the spectral change type according to the similarity threshold. Match the spectral change type with the capacitance attenuation characteristic library, and obtain the spectral correlation degree sequence through the matching degree calculation. Extract the capacitance attenuation characteristic frequency points according to the spectral correlation degree sequence, and record the amplitude change trend sequence corresponding to the characteristic frequency points.

[0072] In one embodiment, extract the characteristic frequencies according to the frequency domain spectral lines to obtain multiple characteristic frequency points. Based on the amplitudes corresponding to each characteristic frequency point and the capacitance attenuation characteristic library, extract the capacitance attenuation characteristic frequency points, and according to the capacitance attenuation characteristic frequency points, obtain the corresponding amplitude change trend sequence, including:

[0073] Extract the frequency domain spectral lines to obtain the harmonic frequency sequence and the harmonic amplitude sequence;

[0074] Classify the harmonic frequency sequence according to the preset frequency interval to obtain multiple harmonic frequency interval sequences, calculate the amplitude proportion of the harmonic frequency components in each harmonic frequency interval sequence to obtain the harmonic frequency distribution characteristic sequence, and select the frequency points with an amplitude proportion greater than the preset proportion in the harmonic frequency distribution characteristic sequence as the characteristic frequency points to obtain multiple characteristic frequency points;

[0075] Based on the amplitudes corresponding to each characteristic frequency point, calculate the amplitude change rate using a fixed-length sliding window to obtain the amplitude change rate sequence, and regard the amplitudes greater than the preset change rate threshold in the amplitude change rate sequence as the amplitude mutation points;

[0076] Calculate the spectral similarity of the amplitude mutation points within the preset range, judge the similarity threshold to obtain the spectral change type, match the spectral change type with the capacitance attenuation characteristic library, obtain the spectral correlation degree sequence through the matching degree calculation, extract the capacitance attenuation characteristic frequency points according to the spectral correlation degree sequence, and record the amplitude change trend sequence corresponding to the capacitance attenuation characteristic frequency points.

[0077] In this embodiment, for the sub-module to collect voltage data, obtain the complete cycle voltage waveform through the preset sampling frequency, and then perform time-frequency domain conversion on the waveform using Fourier transform to extract the harmonic frequency sequence and the harmonic amplitude sequence from the frequency domain spectral lines.

[0078] Classify the harmonic frequency sequence using a preset frequency interval, calculate the proportion of the amplitudes of harmonic components in each frequency interval, and obtain the harmonic frequency distribution characteristic sequence. Extract the characteristic frequencies from the harmonic frequency distribution characteristic sequence, sort according to the proportion of amplitudes in the frequency interval, and select the frequency points with a proportion exceeding the preset threshold as the characteristic frequency points. For the amplitudes corresponding to the characteristic frequency points, calculate the amplitude change rate sequence using a fixed-length sliding window, and then identify the amplitude mutation points by setting a preset change rate threshold. Then, compare the frequency characteristics before and after the amplitude mutation points to obtain the spectral similarity before and after the mutation, and judge the spectral change type through the similarity threshold.

[0079] In addition, match the spectral change type with the capacitor attenuation characteristic library, and obtain the spectral correlation degree sequence through the matching degree calculation. Extract the capacitor attenuation characteristic frequency points according to the spectral correlation degree sequence, and record the amplitude change trend sequence corresponding to the capacitor attenuation characteristic frequency points.

[0080] As an example of this embodiment, when sampling the sub-module voltage, a sampling frequency of 10 kHz is selected to digitally collect the voltage waveform. The voltage data sequence collected under normal working conditions shows periodic change characteristics. The time-domain waveform is converted to the frequency-domain space through Fourier transform to obtain a harmonic frequency sequence with a frequency range from 0 to 5000 Hz. The frequency interval is divided into three frequency intervals: a low-frequency band from 0 to 500 Hz, a medium-frequency band from 500 to 2000 Hz, and a high-frequency band from 2000 to 5000 Hz according to the harmonic frequency magnitude, and the proportion of the amplitudes of harmonic components in each frequency interval is calculated respectively. Under normal working conditions, the harmonic frequencies are mainly distributed in the low-frequency band, the fundamental frequency is 50 Hz, the second harmonic is 100 Hz, and the third harmonic is 150 Hz. The amplitudes of each harmonic decrease with the increase of frequency. When the sub-module capacitor decays, new characteristic frequency components appear in the medium-frequency band. The amplitude of the characteristic frequency component is tracked through a sliding window of 500 sampling points, and the mutation points with an amplitude change rate exceeding 5% per second are recorded. Analyze the spectral characteristics before and after the amplitude mutation points, calculate the spectral similarity of 100 sampling points before and after the mutation, and judge that the spectrum has changed significantly when the similarity is lower than 0.8. The change types of the spectrum include characteristics such as abnormal increase in the amplitude of harmonic components, frequency drift, and bandwidth broadening, and these characteristics have a corresponding relationship with the deterioration of capacitor parameters.

[0081] Extract typical degradation samples from the capacitance attenuation database, establish the correspondence between spectral features and capacitance parameters, set the spectral correlation degree calculation threshold to 0.75, and when the correlation degree between the measured spectrum and the sample spectrum exceeds the threshold, extract the corresponding frequency points as the capacitance attenuation characteristic frequencies. When the capacitor is operating normally, the amplitude of the characteristic frequency points remains below 3% of the fundamental wave amplitude. When the capacitor decays, the amplitude of the characteristic frequency points shows an upward trend, and the amplitude increases by more than 5% within 24 hours. The harmonic characteristics caused by capacitance attenuation are mainly reflected in the change of the spectral distribution. The health state of the capacitor is judged by monitoring the amplitude change trend of the characteristic frequency points. Under normal conditions, the harmonic amplitude in the low-frequency band is stable, and the harmonic amplitude in the medium- and high-frequency bands is small. When the capacitance parameters deteriorate, new characteristic frequency components appear in the medium-frequency band, and the amplitude shows an increasing trend. The selection of characteristic frequencies is based on the calculation of spectral similarity. The setting of the correlation degree threshold takes into account the measurement error and environmental noise effects, and the degree of deterioration is judged by the amplitude change rate.

[0082] S103: Obtain the coupling strength index and the fitting curve of the capacitance attenuation degree from the capacitance attenuation characteristic frequency points, the amplitude change trend sequence, and the voltage fluctuation data of adjacent sub-modules of the DC bus. Obtain the correlation between the coupling strength index and the capacitance attenuation degree according to the fitting curve. Based on the correlation, judge whether the coupling strength index is related to the capacitance attenuation degree. If it is related, calculate the error prediction sequence and confidence level of the monitoring process, and analyze the voltage fluctuation data of the DC bus according to the error prediction sequence and confidence level to obtain the monitoring result, where the monitoring result is the capacitance attenuation degree.

[0083] In this embodiment, according to the frequency and amplitude change trend of the characteristic harmonic components, as well as the historical data related to capacitance attenuation, a mapping relationship model between capacitance attenuation and harmonic components is established, and the support vector machine algorithm is used to train the mapping relationship model. The mapping relationship model outputs the remaining capacitance life and capacitance loss, reflecting the degree of capacitance attenuation.

[0084] Obtain the characteristic harmonic frequency sequence, amplitude change curve, capacitance loss value, and remaining life value from the capacitance attenuation sample database. The characteristic harmonic frequency sequence and amplitude change curve constitute the capacitance attenuation characteristic data set. Perform maximum-minimum normalization processing on the harmonic frequency and amplitude data in the capacitance attenuation characteristic data set to obtain the normalized characteristic matrix, which contains the harmonic frequency distribution characteristics and amplitude change trend characteristics; perform feature extraction on the normalized characteristic matrix by the principal component dimensionality reduction method to obtain the dimensionality-reduced feature vector, which contains the main feature components; use the support vector machine to perform regression prediction on the dimensionality-reduced feature vector to obtain the capacitance loss prediction value and remaining life prediction value, and the regression predictor is optimized by five-fold cross-validation.

[0085] In one embodiment, the capacitance attenuation characteristic frequency points, the amplitude change trend sequence, and the voltage fluctuation data of adjacent sub-modules of the DC bus are used to obtain a fitting curve of the coupling strength index and the capacitance attenuation data, and the correlation of the coupling strength index with the capacitance attenuation degree is obtained according to the fitting curve, including:

[0086] The capacitance attenuation characteristic frequency points and the amplitude change trend sequence are normalized to obtain a normalized feature matrix, and the principal component dimensionality reduction method is used to extract features from the normalized feature matrix to obtain a dimensionality-reduced feature vector;

[0087] The dimensionality-reduced feature vector is input into a trained vector machine model for regression prediction to obtain capacitance attenuation degree data, where the capacitance attenuation degree data includes a capacitance loss prediction value and a remaining life prediction value;

[0088] According to the voltage fluctuation data of adjacent sub-modules of the DC bus and the capacitance attenuation degree data, a fitting curve of the coupling strength index and the capacitance attenuation data is obtained, and the correlation of the coupling strength index with the capacitance attenuation degree is obtained according to the fitting curve.

[0089] In this embodiment, according to the capacitance attenuation samples recorded in the historical database, the characteristic harmonic frequency sequence, the amplitude change curve, the capacitance loss value, and the remaining life value are extracted from the sample data to generate a capacitance attenuation characteristic data set. The harmonic frequency and amplitude data in the characteristic data set are normalized by the maximum and minimum values to generate a normalized feature matrix, and the harmonic frequency distribution feature and the amplitude change trend feature are extracted from the normalized feature matrix. The principal component dimensionality reduction method is used to reduce the dimension of the feature matrix to extract the main feature components and generate a dimensionality-reduced feature vector. For the dimensionality-reduced feature vector and the corresponding capacitance loss value and remaining life value, a support vector machine is used to construct a regression predictor, and the predictor parameters are optimized by five-fold cross-validation. The prediction mean square error is calculated according to the cross-validation results, and the samples with errors exceeding the preset threshold are marked to generate abnormal sample records. Then, the optimized predictor is used to perform prediction calculations on the harmonics feature data collected in real time, and the capacitance loss value and the remaining life value are output. According to the comparison between the predicted output value and the preset threshold, the capacitance attenuation degree is judged to generate an attenuation state index sequence.

[0090] As an example of this embodiment, the capacitance attenuation sample data contains multiple-dimensional features, the harmonic frequency range is between 0 and 5000 Hz, the amplitude change curve records the amplitude change trend within 720 hours, the capacitance loss value represents the attenuation ratio of the capacitance parameter relative to the rated value, and the remaining life value reflects the remaining service life of the capacitor.

[0091] In the normalization process of characteristic data, the frequency data is divided by 5000 Hz to obtain the normalized frequency value, the amplitude data is divided by the fundamental wave amplitude to obtain the normalized amplitude, and the loss value and the life value are already in percentage form. The characteristic matrix contains the amplitude data of 50 frequency points. The 50-dimensional characteristics are reduced to 10-dimensional principal component characteristics by the principal component analysis method, and the principal component characteristics retain 85% of the information of the original data. The feature vector after dimensionality reduction is used as the input data of the support vector machine, and the corresponding capacitance loss value and remaining life value are used as output labels, and a nonlinear mapping relationship is constructed using the radial basis kernel function.

[0092] Five-fold cross-validation is adopted for cross-validation. The sample data is randomly divided into 5 parts, 4 of which are used for training and 1 for validation. After 5 cycles, the complete validation results are obtained. Taking the loss value as an example for the prediction mean square error, the prediction error of normal samples is within 3%. When the prediction error exceeds 5%, the sample is marked as an abnormal sample, and the characteristic distribution law of the abnormal sample is analyzed. In practical applications, the rated life of a certain type of capacitor is 40,000 hours. By online monitoring, the harmonic characteristics in the capacitor voltage are obtained. After feature extraction and normalization, the capacitor loss value is calculated as 12% and the remaining life is 28,000 hours by inputting into the predictor. According to the preset loss threshold of 20% and life threshold of 10,000 hours, it is judged that the capacitor is in a normal working state. Capacitor attenuation prediction adopts a data-driven method, and a mapping relationship between features and attenuation degree is established through historical data accumulation. The feature extraction process pays attention to data denoising and dimensionality reduction to avoid the influence of redundant features on the prediction accuracy. The construction of the predictor is based on the nonlinear regression principle of the support vector machine, and cross-validation ensures the rationality of model parameter selection.

[0093] In practical applications, the online assessment of the capacitor state is realized by combining the preset threshold, providing data support for preventive maintenance. When the capacitor is in a normal working state, the harmonic characteristics remain stable and the error of the prediction result is small. When the capacitor parameters change significantly, the harmonic characteristics change accordingly, and the prediction result can timely reflect the change trend of the capacitor state.

[0094] In one embodiment, according to the voltage fluctuation data of adjacent sub-modules of the DC bus and the capacitor attenuation degree data, a fitting curve of the coupling strength index and the capacitor attenuation data is obtained, and the correlation between the coupling strength index and the capacitor attenuation degree is obtained according to the fitting curve, including:

[0095] Collect the voltage fluctuation data of adjacent sub-modules of the DC bus, align the voltage fluctuation data using a synchronous sampling clock to obtain the aligned fluctuation data, and collect the aligned fluctuation data according to the second preset sampling frequency to obtain the complete cycle waveform data;

[0096] The complete cycle waveform data is segmented using a fixed-length sliding window to obtain segmented data. The segmented data is subjected to Fourier transform to obtain a frequency-domain representation sequence. Multiple characteristic harmonic frequency points are extracted from the frequency-domain representation sequence. According to each characteristic harmonic frequency point, the frequency offset between adjacent sub-modules is calculated to obtain a frequency deviation sequence. The characteristic harmonic frequency points with a frequency offset greater than a preset frequency threshold in the frequency deviation sequence are taken as abnormal frequency points;

[0097] According to the abnormal frequency points, the characteristic harmonic amplitude ratio of adjacent sub-modules is calculated. The waveform correlation coefficient is calculated using adjacent window data to obtain a coupling characteristic sequence. The average value of the waveform correlation coefficient sequence is calculated to obtain a coupling strength index. Polynomial fitting is performed on the coupling strength index and the capacitor attenuation degree data to obtain a fitting curve. The fitting curve is extracted to obtain the correlation between the coupling strength index and the capacitor attenuation degree.

[0098] In this embodiment, the voltage fluctuation data of adjacent sub-modules of the DC bus is obtained to obtain the fluctuation coupling information between the adjacent sub-modules and the DC bus. By analyzing the propagation of characteristic harmonic components in adjacent sub-modules, the correlation between the coupling strength and the capacitor attenuation degree is judged.

[0099] Voltage fluctuation data is collected for adjacent sub-modules of the DC bus. The fluctuation data is aligned using a synchronous sampling clock, and a complete cycle waveform is obtained through a preset sampling frequency. The synchronous sampling data is segmented using a fixed-length sliding window, and the segmented data is subjected to Fourier transform to obtain a frequency-domain representation sequence. Characteristic harmonic frequency points are extracted from the frequency-domain representation sequence, the frequency offset between adjacent sub-modules is calculated, and a frequency deviation sequence is generated. The frequency deviation sequence is compared with a preset frequency threshold, and the frequency deviation over-limit frequency points are marked to generate a frequency anomaly record. Then, the characteristic harmonic amplitude ratio of adjacent sub-modules is calculated, and the waveform correlation coefficient is calculated using adjacent window data to generate a coupling characteristic sequence. The propagation loss between adjacent sub-modules is calculated according to the coupling characteristic sequence, the coupling strength index is obtained from the propagation loss data, and then polynomial fitting is performed on the coupling strength index and the sub-module capacitor attenuation data. The variation law of the coupling strength index with the capacitor attenuation degree is extracted from the fitting curve to obtain the correlation between the coupling strength index and the capacitor attenuation degree.

[0100] As an example of this embodiment, there is a coupled propagation characteristic in the voltage fluctuations of adjacent sub-modules on the DC bus. The adjacent sub-modules are synchronously sampled at a sampling frequency of 10 kHz, the time interval between adjacent sampling points is 0.1 ms, and the sampling clock deviation is controlled within 1 μs to ensure the synchronism of waveform data. By setting a data segment with a sampling length of 1 s, the complete cycle waveforms of adjacent sub-modules are recorded. In waveform analysis, a sliding window with a length of 0.2 s is used to segment the data, and the overlapping rate of adjacent windows is set to 50%, and the Fourier transform is performed on the data of each window. In the frequency domain space, the fundamental frequency is 50 Hz, and the main characteristic harmonics are distributed in the frequency band below 500 Hz. The frequency points with an amplitude exceeding 3% of the fundamental wave amplitude in this frequency band are recorded as characteristic harmonic frequency points. For the characteristic harmonic frequency points, the frequency offset between adjacent sub-modules is calculated. Under normal operating conditions, the frequency offset is within 1 Hz, and when the frequency offset exceeds 2 Hz, it is marked as an abnormal frequency point.

[0101] The harmonic amplitude ratio of adjacent sub-modules reflects the attenuation characteristic during the fluctuation propagation process. An amplitude ratio between 0.9 and 1.1 indicates a small propagation loss and a strong coupling effect. The waveform correlation coefficient is calculated using the data of adjacent windows with a length of 0.1 s, and the calculation interval is 0.02 s to obtain a waveform correlation coefficient sequence. A correlation coefficient greater than 0.8 indicates good waveform synchronism and high coupling strength, and a correlation coefficient less than 0.5 indicates serious waveform distortion and low coupling strength. The average value is calculated from the correlation coefficient sequence as the coupling strength index.

[0102] In practical applications, when the capacitance attenuation of a certain type of sub-module reaches 15%, the coupling strength index between adjacent modules decreases from 0.85 to 0.65, and the propagation loss increases. The corresponding relationship between the coupling strength and the capacitance attenuation is obtained through polynomial fitting. The fitting curve shows that the coupling strength decreases with the increase of the capacitance attenuation degree. When the capacitance attenuation exceeds 20%, the coupling strength index is lower than 0.5, and the fluctuation propagation characteristic changes significantly. The coupling effect between sub-modules is closely related to the capacitance parameters. By analyzing the fluctuation propagation characteristic, the change of the capacitance state can be reflected. The coupling strength index comprehensively considers characteristics such as frequency offset, amplitude attenuation, and waveform correlation to evaluate the capacitance attenuation. Within the normal range of capacitance parameters, the coupling effect between adjacent sub-modules remains stable, and the fluctuation propagation loss is small. As the capacitance parameters deteriorate, the coupling effect weakens, the propagation loss increases, and the waveform distortion intensifies.

[0103] In one embodiment, calculating the error prediction sequence and confidence level of the monitoring process includes:

[0104] Using a prediction neural network to extract features from the fluctuation coupling information to obtain a monitoring feature vector;

[0105] The monitored feature vector is input into the constructed correlation model between the coupling relationship and the monitoring accuracy for processing, and an error prediction sequence and a confidence level of the monitoring process are obtained.

[0106] In this embodiment, it is determined whether the coupling strength index is related to the degree of capacitance attenuation according to the correlation. If they are related, a prediction neural network is used to extract features from the fluctuating coupling information. The features include the coupling strength, propagation speed, etc. Then, in combination with the capacitance attenuation quantization index, a correlation model between the coupling relationship and the monitoring accuracy is constructed, and an error prediction sequence and a confidence level of the monitoring process are obtained.

[0107] A data set is constructed according to the fluctuating coupling information matrix, and the maximum and minimum normalization is performed on the coupling strength sequence, the propagation speed sequence, and the attenuation quantization sequence to generate a standardized feature matrix. A multi-layer convolution structure is constructed. In the first layer, a processor performs a convolution operation on the standardized feature matrix to generate a coupling feature map. A second convolution operation is performed on the coupling feature map to extract a propagation feature map, and a pooling operation is used to reduce the feature dimension. Through a fully connected layer, the dimension-reduced features are mapped and transformed to generate a monitored feature vector.

[0108] In one embodiment, a prediction neural network is used to extract features from the fluctuating coupling information to obtain a monitored feature vector, including:

[0109] The fluctuating coupling information is normalized to obtain a standardized feature matrix;

[0110] A processor performs a convolution operation on the standardized feature matrix to obtain a coupling feature map;

[0111] A second convolution operation is performed on the coupling feature map to obtain a propagation feature map, a pooling operation is used to reduce the dimension of the propagation feature map to obtain dimension-reduced features, and a fully connected layer is used to map and transform the dimension-reduced features to obtain a monitored feature vector.

[0112] In this embodiment, the fluctuation coupling information includes three types of characteristic data: coupling strength, propagation speed, and attenuation quantization. The numerical range of the coupling strength is between 0 and 1, the numerical value of the propagation speed is between 0 and 1000 meters per second, and the numerical value of the attenuation quantization is between 0 and 100%. All the characteristic data are mapped to the interval of 0 to 1 through min-max normalization to ensure the comparability of characteristics with different dimensions. The convolutional structure adopts a two-layer design. The first layer uses 16 3×3 convolutional kernels to process the input features, with a stride set to 1 and the padding method using zero padding, and outputs 16 feature maps. The feature maps reflect the local characteristics and spatial correlation of the coupling data, and the feature expression ability is enhanced through a non-linear activation function. The second layer uses 32 3×3 convolutional kernels for feature extraction, with a stride set to 2 to achieve feature dimensionality reduction. The pooling layer adopts a 2×2 maximum pooling window with a stride of 2 to extract the most significant features in the local area, reduce the data dimension, and improve the robustness of the feature expression. The fully connected layer contains 128 neurons, and a non-linear activation function is used to perform mapping transformation on the features, and the monitoring feature vector is output.

[0113] Then, for the waveform data of adjacent sub-modules, the characteristic harmonic frequency points are extracted, the amplitude ratio between adjacent sub-modules is calculated, and the harmonic propagation attenuation sequence is obtained. The harmonic propagation attenuation sequence is segmented by time window, the time displacement between adjacent windows is calculated, and the propagation speed characteristic sequence is obtained according to the physical distance between sub-modules. The AC voltage waveform and current waveform at both ends of the capacitor are collected, the peak-to-valley value of the voltage is calculated to obtain the ripple coefficient, and the loss angle parameter is obtained from the voltage-current phase difference. A normalization processor is used to standardize the propagation speed characteristic sequence, the ripple coefficient, and the loss angle parameter to generate a characteristic data matrix. A three-layer neural network structure is constructed, that is, the correlation model between the constructed coupling relationship and the monitoring accuracy. The input layer receives the characteristic data matrix, the hidden layer extracts the feature mapping relationship, and the output layer generates an error prediction sequence. Probability statistics are performed on the error prediction sequence, the mean and standard deviation parameters are calculated, and an error distribution function is generated. The cumulative probability value is calculated according to the error distribution function, and the monitoring error range is obtained through a preset confidence threshold. The monitoring error range is compared and verified with the measured data, and the neural network parameters are updated to optimize the feature extraction method.

[0114] The harmonic propagation characteristics between adjacent sub-modules reflect the capacitance parameter status. By measuring the attenuation of characteristic harmonics between adjacent modules, the amplitude ratios of harmonic frequencies at 50 Hz, 100 Hz, and 150 Hz are extracted. Under normal operating conditions, the harmonic propagation attenuation ratio is between 0.9 and 1.1. When the capacitance parameters deteriorate, the propagation attenuation intensifies and the ratio drops below 0.7. In the time window analysis, a 0.1-second data segment is selected to calculate the propagation time difference. The physical distance between adjacent sub-modules is 0.5 meters, and the propagation speed is calculated from the time difference and the distance. Under normal conditions, the propagation speed is about 800 m / s. When the capacitance parameters deteriorate, the propagation speed drops below 500 m / s, indicating a change in the harmonic propagation characteristics. In the measurement of the AC characteristics of the capacitor, the peak value of the voltage waveform is 550 V, the valley value is 450 V, and the calculated value of the ripple coefficient is 0.1, reflecting the voltage smoothness. The voltage-current phase difference is 15 degrees, and the tangent value of the loss angle is 0.27, characterizing the loss characteristics of the capacitor.

[0115] Through normalization, the ripple coefficient and loss angle data are mapped to the range of 0 to 1 to ensure the comparability of data with different dimensions. The neural network adopts a three-layer structure. The input layer contains 12 neurons corresponding to three types of features: propagation speed, ripple coefficient, and loss angle. The hidden layer is set with 24 neurons, and the hyperbolic tangent activation function is used to enhance the non-linear expression ability. The output layer is set with 1 neuron to generate the error prediction value. The network training uses 1000 sets of historical data, of which 800 sets are used for training and 200 sets are used for verification. The error statistical analysis shows that the mean of the prediction sequence is 0.03 and the standard deviation is 0.015. The normal distribution test shows that the error basically conforms to the normal distribution. At the 95% confidence level, the monitoring error range is plus or minus 2 times the standard deviation of the mean, that is, 0 to 0.06. The verification with the measured data shows that when the capacitance parameters are normal, the prediction error is within 0.02 and the confidence level reaches 98%. As the capacitance parameters deteriorate, the prediction error increases to 0.05 and the confidence level drops to 92%. By comparing the verification results, the neural network structure is optimized, the number of neurons in the hidden layer is adjusted to 32, the weight parameters are updated, and the feature extraction method is improved. The average error of the optimized predictor on the test data is reduced to 0.02, the standard deviation is reduced to 0.01, and the 95% confidence interval is narrowed to 0 to 0.04. In particular, the prediction performance in the capacitance deterioration state is improved, providing a more accurate evaluation basis for capacitance status monitoring.

[0116] In one embodiment, the prediction neural network is trained by using the sample fluctuation coupling information, including:

[0117] Obtain the sample fluctuation coupling information, input the sample fluctuation coupling information into the initial prediction neural network for prediction, and obtain the initial monitoring result;

[0118] Calculate the mean square error of the initial monitoring results to obtain a monitoring error sequence, perform probability distribution fitting on the monitoring error sequence, and obtain the standard deviation and mean parameters;

[0119] Calculate the monitoring confidence interval based on the standard deviation and mean parameters to obtain the monitoring accuracy evaluation result;

[0120] Optimize the parameters of the initial prediction neural network according to the monitoring accuracy evaluation result and the monitoring error sequence, update the feature extraction parameters and prediction parameters, and obtain a prediction neural network.

[0121] In this embodiment, the convolutional neural network is trained using sample fluctuation coupling information, and the sample fluctuation coupling information is input into the initial prediction neural network for prediction to obtain the initial monitoring results. A mean square error calculator is used to calculate the error of the initial monitoring results to generate a monitoring error sequence. Probability distribution fitting is performed on the monitoring error sequence to calculate the standard deviation and mean parameters. The monitoring confidence interval is calculated based on the standard deviation and mean parameters to generate the monitoring accuracy evaluation result. The predictor parameters are optimized using the monitoring accuracy evaluation result, and the feature extraction parameters and prediction parameters are updated.

[0122] The mean square error criterion is used for error calculation, and 1000 groups of test data are verified to calculate the error sequence of the predicted value and the true value. The mean of the error sequence is 0.05, and the standard deviation is 0.02. The normal test shows that the error basically conforms to the normal distribution. At the 95% confidence level, the confidence interval is plus or minus 2 times the standard deviation of the mean, that is, from 0.01 to 0.09. The monitoring accuracy evaluation result shows that for data with a coupling strength greater than 0.8, the prediction error is within 0.03, and the confidence level reaches 98%. For data with a coupling strength between 0.5 and 0.8, the prediction error is within 0.05, and the confidence level is 95%. When the coupling strength is lower than 0.5, the prediction error increases to 0.08, and the confidence level decreases to 90%. According to the evaluation result, the predictor parameters are optimized, the number and size of the convolutional kernels are adjusted, the pooling strategy is updated, and the number of neurons in the fully connected layer is optimized. The average error of the optimized predictor on the test data is reduced from 0.05 to 0.03, the standard deviation is reduced from 0.02 to 0.015, and the 95% confidence interval is narrowed to 0.01 to 0.07. The improvement in monitoring accuracy is reflected in the reduction of the prediction error and the convergence of the confidence interval, especially the improvement of the prediction performance at low coupling strengths.

[0123] In one embodiment, according to the error prediction sequence and the confidence level, the voltage fluctuation data of the DC bus is analyzed to obtain the monitoring results, including:

[0124] Calculate the filter center frequency and bandwidth parameters according to the error prediction sequence and the confidence level;

[0125] Filter the bus voltage fluctuation data using the filter center frequency parameter and the bandwidth parameter to obtain the filtered data, extract the filtered data to obtain a multi-band fluctuation feature sequence;

[0126] Perform wavelet multi-scale decomposition on the multi-band fluctuation feature sequence to obtain a harmonic feature sequence, and use a support vector machine to perform classification operations on the harmonic feature sequence to obtain an attenuation feature vector;

[0127] Calculate the capacitance attenuation quantization index according to the attenuation feature vector, perform normalization processing on the capacitance attenuation index to obtain a correction result sequence, compare the correction result sequence with the original capacitance attenuation degree to obtain a correction deviation, and update the filter parameters according to the correction deviation to obtain an optimized non-invasive monitoring model;

[0128] Use the optimized non-invasive monitoring model to analyze the voltage fluctuation data of the DC bus to obtain a monitoring result.

[0129] In this embodiment, according to the error rate and the confidence level, adjust the filter parameters and the sampling frequency of the non-invasive monitoring algorithm, optimize the extraction accuracy of the characteristic harmonic components, re-analyze the DC bus voltage fluctuation data, extract the corrected characteristic harmonic components, and determine whether the capacitance attenuation degree is consistent with the correction result. If the correction result is consistent with the capacitance attenuation degree, output the final monitoring result. If not, re-adjust the model parameters and iteratively optimize the monitoring algorithm until the correction result is consistent with the capacitance attenuation degree.

[0130] First, calculate the filter center frequency and bandwidth parameters using the error prediction sequence and the confidence level, extract the passband fluctuation and stopband attenuation characteristics from the frequency response curve, and generate a filter optimization parameter sequence. Construct a multi-level filter bank, set the passband characteristics of each level of filter according to the optimization parameter sequence, subdivide the filter passband range, and generate a frequency-divided filter bank.

[0131] Preprocess the bus voltage fluctuation data using the frequency-divided filter bank, extract the voltage fluctuation characteristics from each frequency band to obtain a multi-band fluctuation feature sequence. Perform wavelet multi-scale decomposition on the multi-band fluctuation feature sequence, set the wavelet basis function and the decomposition layer number, and extract the frequency characteristics from each scale coefficient. Set the amplitude threshold according to the frequency characteristics to screen the characteristic harmonic components and generate a harmonic feature sequence. Use a support vector machine to classify the harmonic feature sequence, set the kernel function parameter and the penalty factor, and generate an attenuation feature vector.

[0132] Calculate the capacitance attenuation quantization index according to the attenuation eigenvector, normalize the capacitance attenuation index, and generate a corrected result sequence. Compare and verify the corrected result sequence with the original attenuation degree, calculate the correction deviation, and update the filter optimization parameters. The filter optimization design is based on the error rate and confidence index. When the error rate is 0.05 and the confidence is 95%, the center frequency is set at 500 Hz, the passband bandwidth is 200 Hz, and the stopband attenuation is greater than 40 dB. The filter optimization parameters include four dimensions: center frequency, bandwidth, passband ripple, and stopband attenuation. The optimal parameter combination is determined through the analysis of the frequency response curve. The multi-stage filter bank adopts a 4-stage structure, and the frequency range is divided into four sub-bands: 0 to 250 Hz, 250 to 500 Hz, 500 to 750 Hz, and 750 to 1000 Hz. The passband ripple of each stage of the filter is controlled within 1 dB, and the stopband attenuation increases gradually with frequency. The fine division of the frequency band is realized through multi-stage filtering, and the frequency resolution of feature extraction is improved. In the extraction of voltage fluctuation features, the sampling frequency is set at 10 kHz, and the wavelet decomposition is performed on the fluctuation data of each frequency band. The db4 wavelet basis function is selected for 4-layer decomposition. Frequency features are extracted from each scale coefficient. In the frequency band of 0 to 250 Hz, the fundamental component and low-order harmonics are extracted; in the frequency band of 250 to 500 Hz, the intermediate-frequency harmonic features are concerned; in the frequency band of 500 to 1000 Hz, the changes in high-frequency components are monitored. The amplitude threshold for harmonic feature discrimination is set at 3% of the fundamental amplitude, and the characteristic harmonic frequency points are screened. The fundamental frequency is 50 Hz, the second harmonic is 100 Hz, and the third harmonic is 150 Hz. The significance of each harmonic is judged by the amplitude ratio. The support vector machine adopts the radial basis kernel function, the kernel parameter is set at 0.1, and the penalty factor is 100 to establish the mapping relationship of the feature space. The calculation of the capacitance attenuation quantization index shows that the capacitance parameter attenuation is within 10% under normal working conditions, and the deviation between the corrected result and the original judgment is within 3%. When the capacitance parameter deteriorates to 20%, the deviation of the corrected result increases to 5%, indicating that the feature extraction accuracy decreases with the increase of the attenuation degree. The filter parameters are updated through comparison and verification, the passband characteristics and stopband attenuation are adjusted, and the optimized correction deviation is controlled within 2%. The filter optimization design runs through the entire monitoring process, from the optimization of the frequency response characteristics to the design of the multi-stage filter structure, and then to the improvement of the feature extraction method, forming a complete optimization chain. Through the feedback of the error rate and confidence index, the filter parameters are continuously adjusted to improve the feature extraction accuracy. In practical applications, the selection of filter parameters needs to balance the frequency resolution and computational complexity, ensuring both the accuracy of feature extraction and the real-time requirements of online monitoring.

[0133] Calculate the difference sequence based on the correction result and the degree of capacitance attenuation, normalize the difference sequence, and generate a correction deviation index. Calculate the harmonic feature extraction parameters according to the correction deviation index, extract the optimization direction from the parameter update curve, and generate a parameter adjustment sequence. Set an update step size for the parameter adjustment sequence, and update the neural network parameters according to the step size to obtain an optimized network parameter set.

[0134] Re-extract the voltage fluctuation features using the optimized network parameters, calculate the characteristic harmonic frequency and amplitude sequence, and generate a corrected characteristic sequence. Calculate the degree of capacitance attenuation based on the corrected characteristic sequence, quantify the attenuation degree value, and generate a corrected attenuation index.

[0135] Calculate the deviation value between the corrected attenuation index and the original attenuation degree, and determine whether the deviation value is less than the preset threshold. If the deviation value is less than the preset threshold, output the monitoring result; if the deviation value is greater than the preset threshold, return to the parameter optimization link. Judge the convergence situation according to the optimization iteration times, and stop the optimization when the iteration times exceed the preset value, and output the latest monitoring result.

[0136] The difference calculation between the correction result and the degree of capacitance attenuation uses the relative error method. When the correction result is 15% and the actual attenuation degree is 12%, the relative error is 25%. By normalizing the processing, all differences are mapped to the interval from 0 to 1, and the correction deviation index 0.25 is obtained. The correction deviation index reflects the accuracy of the monitoring result, and the smaller the index value, the more accurate the monitoring.

[0137] During the parameter optimization process, the neural network includes 12 neurons in the input layer, 24 neurons in the hidden layer, and 1 neuron in the output layer. The adaptive learning rate method is used for parameter adjustment, and the initial learning rate is set to 0.01. When the correction deviation is greater than 0.2, the learning rate is increased to 0.02 to accelerate parameter convergence; when the correction deviation is less than 0.1, the learning rate is decreased to 0.005 to improve the optimization accuracy. The feature extraction parameters include three dimensions: filter bandwidth, sampling frequency, and window length. The bandwidth is adjusted from 200 Hz to 300 Hz, the sampling frequency is increased from 8 kHz to 12 kHz, and the window length is extended from 0.1 s to 0.15 s. The parameter adjustment direction is determined by the gradient of the correction deviation, and the parameter update step size is proportional to the deviation magnitude. The larger the deviation, the larger the step size. The optimized parameters are used to re-extract the voltage fluctuation features, and the characteristic harmonic components are identified in the frequency band from 0 to 1000 Hz, and the fundamental wave of 50 Hz and its harmonic components are extracted. The corrected characteristic harmonic amplitudes show that the fundamental wave amplitude is 10 V, the second harmonic is 0.8 V, and the third harmonic is 0.5 V. The harmonic distribution more accurately reflects the capacitor parameter state. The quantification of the capacitor attenuation degree uses the piecewise linear method. When the harmonic content is less than 5%, the attenuation degree is within 10%; when the harmonic content reaches 8%, the attenuation degree increases to 15%; when the harmonic content exceeds 10%, the attenuation degree exceeds 20%. Through the correspondence between the harmonic features and the attenuation degree, an accurate state evaluation standard is established. The maximum number of iterations in the optimization iteration process is set to 100, and convergence is determined when the change in the correction deviation for 10 consecutive iterations is less than 0.01. In actual optimization, convergence conditions are generally achieved after 20 to 30 iterations, and the correction deviation is reduced from the initial 0.25 to less than 0.05.

[0138] For data under different working conditions, there are differences in the convergence speed, but the expected optimization goals can ultimately be achieved. The optimization process of capacitor state monitoring reflects the adaptive adjustment ability, continuously optimizing parameters through the feedback of the correction deviation to improve the monitoring accuracy. Parameter optimization, feature extraction, and state evaluation constitute a complete optimization chain to ensure the reliability of the monitoring results. During the optimization process, it is necessary to balance the optimization accuracy and the convergence speed, ensuring both accurate monitoring results and meeting the real-time requirements of on-line monitoring.

[0139] The cascaded module non-intrusive monitoring system provided by the embodiment of the present invention, as Figure 2 shown, Figure 2 is a block diagram of the cascaded module non-intrusive monitoring system 200, including:

[0140] An acquisition module 201, configured to acquire the voltage signal of the DC bus in a preset time window length to obtain a voltage fluctuation time-domain signal, perform signal conversion and extraction on the voltage fluctuation time-domain signal by using the fast Fourier transform to obtain a plurality of main frequency components, and obtain a harmonic feature sequence according to each main frequency component;

[0141] The first calculation module 202 is configured to obtain the voltage waveform of the sub-module according to the first preset sampling frequency, perform time-frequency domain conversion on the voltage waveform by using Fourier transform to obtain frequency domain spectral lines, extract characteristic frequencies according to the frequency domain spectral lines to obtain a plurality of characteristic frequency points, extract the capacitance attenuation characteristic frequency points based on the amplitudes corresponding to the respective characteristic frequency points and the capacitance attenuation characteristic library, and obtain an amplitude change trend sequence according to the amplitudes corresponding to the capacitance attenuation characteristic frequency points;

[0142] The second calculation module 203 is configured to obtain a fitting curve of the coupling strength index and the capacitance attenuation degree by using the capacitance attenuation characteristic frequency points, the amplitude change trend sequence, and the voltage fluctuation data of adjacent sub-modules of the DC bus, obtain the correlation of the coupling strength index with respect to the capacitance attenuation degree according to the fitting curve, determine whether the coupling strength index is related to the capacitance attenuation degree based on the correlation. If they are related, calculate the error prediction sequence and the confidence level of the monitoring process, and analyze the voltage fluctuation data of the DC bus according to the error prediction sequence and the confidence level to obtain a monitoring result, where the monitoring result is the capacitance attenuation degree.

[0143] The specific implementation manner of the non-intrusive monitoring system for cascaded modules is basically the same as the specific embodiments of the above non-intrusive monitoring method for cascaded modules, and will not be described in detail here.

[0144] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered that the scope recorded in this specification.

[0145] The above specific embodiments have further elaborated the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. In particular, for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A non-intrusive monitoring method for cascade modules, characterized in that: include: Acquire a voltage signal of a DC bus in a preset time window length to obtain a voltage fluctuation time domain signal, perform signal conversion and extraction on the voltage fluctuation time domain signal using a fast Fourier transform to obtain a plurality of main frequency components, and obtain a harmonic characteristic sequence according to each of the main frequency components; Acquire a voltage waveform of the submodule according to a first preset sampling frequency, perform time-frequency domain conversion on the voltage waveform using Fourier transform to obtain a frequency domain spectrum, perform characteristic frequency extraction according to the frequency domain spectrum to obtain a plurality of characteristic frequency points, extract capacitance attenuation characteristic frequency points based on the amplitude corresponding to each of the characteristic frequency points and a capacitance attenuation characteristic library, and obtain an amplitude change trend sequence according to the amplitude corresponding to the capacitance attenuation characteristic frequency point; The capacitance attenuation characteristic frequency point, the amplitude change trend sequence and the voltage fluctuation data of the adjacent submodules of the DC bus are used to obtain a fitting curve of the coupling strength index and the capacitance attenuation degree. The correlation between the coupling strength index and the capacitance attenuation degree is obtained according to the fitting curve. Whether the coupling strength index is correlated with the capacitance attenuation degree is determined based on the correlation. If so, the error prediction sequence and the confidence level of the monitoring process are calculated. According to the error prediction sequence and the confidence level, the voltage fluctuation data of the DC bus is analyzed to obtain a monitoring result, wherein the monitoring result is the capacitance attenuation degree.

2. The non-intrusive monitoring method of cascade modules according to claim 1, characterized in that: The voltage fluctuation time domain signal is converted and extracted by fast Fourier transform to obtain a plurality of main frequency components, and a harmonic characteristic sequence is obtained according to each of the main frequency components, including: Acquire a voltage signal of the DC bus in a preset time window length to obtain a plurality of voltage data, calculate a fluctuation amplitude sequence according to a preset reference voltage value and each of the voltage data, and obtain a voltage fluctuation time domain signal; Performing segmented windowing processing on the voltage fluctuation time domain signal, smoothing the signal using a preset window function, and obtaining a smoothed voltage fluctuation sequence; Performing Fourier transform on the smoothed voltage fluctuation sequence to obtain a frequency domain representation sequence, performing spectrum line extraction on the frequency domain representation sequence according to a preset frequency domain analysis range to obtain a plurality of main frequency components, grouping the main frequency components according to their frequency values, and determining a fundamental frequency component and a plurality of subharmonic frequency components; The ratios of the amplitudes of the subharmonic frequency components to the fundamental frequency component are calculated to obtain a harmonic component content ratio sequence, and the subharmonic frequency components in the harmonic component content ratio sequence that are greater than a preset harmonic limit are determined to be out-of-limit harmonic components to obtain a harmonic characteristic sequence.

3. The non-intrusive monitoring method of cascade modules according to claim 1, characterized in that: The characteristic frequency is extracted according to the frequency domain spectrum to obtain a plurality of characteristic frequency points, and the capacitance attenuation characteristic frequency points are extracted based on the amplitude corresponding to each of the characteristic frequency points and the capacitance attenuation characteristic library, and the corresponding amplitude change trend sequence is obtained according to the capacitance attenuation characteristic frequency points, including: Extract the frequency domain spectrum lines to obtain a harmonic frequency sequence and a harmonic amplitude sequence; Classifying the harmonic frequency sequence according to a preset frequency interval to obtain a plurality of harmonic frequency interval sequences, calculating the amplitude proportion of the harmonic frequency components in each of the harmonic frequency interval sequences to obtain a harmonic frequency distribution characteristic sequence, and selecting a frequency point whose amplitude proportion is greater than a preset proportion in the harmonic frequency distribution characteristic sequence as a characteristic frequency point to obtain a plurality of characteristic frequency points; Based on the amplitude corresponding to each of the characteristic frequency points, a fixed-length sliding window is used to calculate the amplitude change rate to obtain an amplitude change rate sequence, and an amplitude in the amplitude change rate sequence that is greater than a preset change rate threshold is used as an amplitude mutation point; Calculate the spectrum similarity of the amplitude mutation point within a preset range, judge the similarity threshold to obtain the spectrum change type, match the spectrum change type with the capacitance attenuation feature library, obtain the spectrum correlation sequence through matching calculation, extract the capacitance attenuation characteristic frequency point according to the spectrum correlation sequence, and record the amplitude change trend sequence corresponding to the capacitance attenuation characteristic frequency point.

4. The non-intrusive monitoring method of cascade modules according to claim 1, characterized in that: The method of obtaining a fitting curve between a coupling strength index and capacitance attenuation data by using the capacitance attenuation characteristic frequency point, the amplitude change trend sequence and the voltage fluctuation data of the adjacent submodules of the DC bus, and obtaining the correlation between the coupling strength index and the capacitance attenuation degree according to the fitting curve includes: Normalizing the capacitance attenuation characteristic frequency points and the amplitude change trend sequence to obtain a normalized characteristic matrix, and extracting features from the normalized characteristic matrix using a principal component dimensionality reduction method to obtain a reduced dimensionality characteristic vector; Inputting the dimension-reduced feature vector into a trained vector machine model for regression prediction to obtain capacitance attenuation degree data, wherein the capacitance attenuation degree data includes a capacitance loss prediction value and a remaining life prediction value; A fitting curve between a coupling strength index and the capacitance attenuation data is obtained according to the voltage fluctuation data of adjacent submodules of the DC bus and the capacitance attenuation degree data, and a correlation between the coupling strength index and the capacitance attenuation degree is obtained according to the fitting curve.

5. The non-intrusive monitoring method of cascade modules according to claim 4, characterized in that: The method of obtaining a fitting curve between a coupling strength index and capacitance attenuation data according to the voltage fluctuation data of adjacent submodules of the DC bus and the capacitance attenuation degree data, and obtaining a correlation between the coupling strength index and the capacitance attenuation degree according to the fitting curve, includes: Collecting voltage fluctuation data of adjacent sub-modules of the DC bus, aligning the voltage fluctuation data using a synchronous sampling clock to obtain aligned fluctuation data, and collecting the aligned fluctuation data according to a second preset sampling frequency to obtain complete cycle waveform data; The complete cycle waveform data is segmented using a fixed-length sliding window to obtain segmented data, the segmented data is Fourier transformed to obtain a frequency domain representation sequence, a plurality of characteristic harmonic frequency points are extracted from the frequency domain representation sequence, and the frequency offset between adjacent submodules is calculated according to each of the characteristic harmonic frequency points to obtain a frequency deviation sequence, and the characteristic harmonic frequency point in the frequency deviation sequence whose frequency offset is greater than a preset frequency threshold is taken as an abnormal frequency point; According to the abnormal frequency point, the characteristic harmonic amplitude ratio of adjacent submodules is calculated, the waveform correlation coefficient is calculated using adjacent window data to obtain a coupling characteristic sequence, the average value of the waveform correlation coefficient sequence is calculated to obtain a coupling strength index, a polynomial fitting is performed on the coupling strength index and the capacitance attenuation degree data to obtain a fitting curve, the fitting curve is extracted to obtain the correlation between the coupling strength index and the capacitance attenuation degree.

6. The non-intrusive monitoring method of cascade modules according to claim 1, characterized in that: The error prediction sequence and confidence level of the computing monitoring process include: The prediction neural network is used to extract the characteristics of the wave coupling information and obtain the monitoring characteristic vector; The monitoring feature vector is input into the constructed correlation model of coupling relationship and monitoring accuracy for processing, so as to obtain the error prediction sequence and confidence level of the monitoring process.

7. The non-intrusive monitoring method of cascade modules according to claim 6, characterized in that: The method of extracting features from the fluctuation coupling information using a prediction neural network to obtain a monitoring feature vector includes: Normalizing the wave coupling information to obtain a standardized feature matrix; A processor is used to perform convolution operation on the standardized feature matrix to obtain a coupling feature map; A second layer of convolution operation is performed on the coupling feature map to obtain a propagation feature map, a pooling operation is used to reduce the dimension of the propagation feature map to obtain a reduced-dimensional feature, and a fully connected layer is used to map and transform the reduced-dimensional feature to obtain a monitoring feature vector.

8. The non-intrusive monitoring method of cascade modules according to claim 6, characterized in that: The prediction neural network is trained by using sample fluctuation coupling information, including: Obtain sample fluctuation coupling information, input the sample fluctuation coupling information into the initial prediction neural network for prediction, and obtain the initial monitoring result; Calculating the mean square error of the initial monitoring result to obtain a monitoring error sequence, performing probability distribution fitting on the monitoring error sequence to obtain standard deviation and mean parameters; Calculate the monitoring confidence interval according to the standard deviation and the mean parameter to obtain a monitoring accuracy evaluation result; The parameters of the initial prediction neural network are optimized according to the monitoring accuracy evaluation result and the monitoring error sequence, and the feature extraction parameters and the prediction parameters are updated to obtain the prediction neural network.

9. The non-intrusive monitoring method of cascade modules according to claim 1, characterized in that: The step of analyzing the voltage fluctuation data of the DC bus according to the error prediction sequence and the confidence level to obtain a monitoring result includes: Calculating filter center frequency and bandwidth parameters according to the error prediction sequence and the confidence level; The bus voltage fluctuation data is filtered using the filter center frequency parameter and the bandwidth parameter to obtain filtered data, and the filtered data is extracted to obtain a multi-band fluctuation feature sequence; Performing wavelet multi-scale decomposition on the multi-band fluctuation characteristic sequence to obtain a harmonic characteristic sequence, and using a support vector machine to perform classification operation on the harmonic characteristic sequence to obtain an attenuation characteristic vector; Calculating a capacitance attenuation quantification index according to the attenuation characteristic vector, normalizing the capacitance attenuation index to obtain a correction result sequence, comparing the correction result sequence with the original capacitance attenuation degree to obtain a correction deviation, updating the filter parameters according to the correction deviation, and obtaining an optimized non-intrusive monitoring model; The optimized non-intrusive monitoring model is used to analyze the voltage fluctuation data of the DC bus to obtain a monitoring result.

10. A cascade module non-intrusive monitoring system, characterized in that: include: An acquisition module is used to acquire a voltage signal of a DC bus in a preset time window length, obtain a voltage fluctuation time domain signal, perform signal conversion and extraction on the voltage fluctuation time domain signal using a fast Fourier transform, obtain a plurality of main frequency components, and obtain a harmonic characteristic sequence according to each of the main frequency components; A first calculation module is used to obtain a voltage waveform of a submodule according to a first preset sampling frequency, perform a time-frequency domain conversion on the voltage waveform using Fourier transform to obtain a frequency domain spectrum, perform characteristic frequency extraction according to the frequency domain spectrum to obtain a plurality of characteristic frequency points, extract a capacitance attenuation characteristic frequency point based on the amplitude corresponding to each of the characteristic frequency points and a capacitance attenuation characteristic library, and obtain an amplitude change trend sequence according to the amplitude corresponding to the capacitance attenuation characteristic frequency point; The second calculation module is used to obtain a fitting curve of a coupling strength index and a capacitance attenuation degree by using the capacitance attenuation characteristic frequency point, the amplitude change trend sequence and the voltage fluctuation data of adjacent sub-modules of the DC bus, obtain the correlation between the coupling strength index and the capacitance attenuation degree according to the fitting curve, judge whether the coupling strength index is correlated with the capacitance attenuation degree based on the correlation, and if so, calculate the error prediction sequence and the confidence level of the monitoring process, analyze the voltage fluctuation data of the DC bus according to the error prediction sequence and the confidence level, and obtain a monitoring result, wherein the monitoring result is the capacitance attenuation degree.

Images