Photovoltaic inverter fault detection method and system based on multi-source data

Abstract

The invention relates to the technical field of photovoltaic inverter fault detection, and discloses a photovoltaic inverter fault detection method and system based on multi-source data, and the method comprises the steps: collecting an original current signal, and obtaining an eigenmode function sequence from the original current signal according to a preset rule; arranging the sequence according to a rule, and marking the sequence as a risk eigenmode function if an index exceeds a preset threshold value; decomposing the risk eigenmode function through a preset method to obtain a fine sub-band frequency component coefficient, extracting energy distribution characteristics of the fine sub-band frequency component coefficient, and marking a current fluctuation section if a local energy deviation value exceeds a threshold value; performing envelope demodulation on the section to obtain instantaneous amplitude and frequency, calculating a local abnormal energy value according to the instantaneous amplitude and frequency, and forming an aging trend curve; and generating a fault evolution sequence in combination with the fluctuation section and the trend curve, and outputting an early warning sequence after identifying a critical mutation point. According to the method, the fault detection accuracy of the photovoltaic inverter can be improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic inverter fault detection technology, and in particular to a photovoltaic inverter fault detection method and system based on multi-source data. Background Technology

[0002] With the rapid development of renewable energy, photovoltaic inverters, as the core equipment of photovoltaic power generation systems, have become a key link in ensuring the stability and sustainability of energy supply. Fault prediction and health management, as important means to improve equipment reliability, are increasingly needed in the operation and maintenance of photovoltaic inverters.

[0003] In one existing technology, photovoltaic inverter fault detection determines the fault by comparing real-time collected macroscopic parameters such as voltage and current with preset normal ranges and simply analyzing whether the parameter statistics deviate from the historical normal range. However, this type of method is not only difficult to capture the instantaneous and weak transient disturbances caused by early faults such as aging of switching devices, but also suffers from severe mode aliasing due to its inability to adaptively separate the multi-frequency components of non-stationary signals, making it difficult to distinguish between normal fluctuations and abnormal features. At the same time, it lacks the ability to track the evolution trend of faults from weak to severe, and the fixed model is difficult to adapt to the non-stationary characteristics of signals in complex environments, which easily leads to misjudgment.

[0004] Therefore, existing technologies suffer from inaccurate fault detection in photovoltaic inverters. Summary of the Invention

[0005] This invention provides a photovoltaic inverter fault detection method and system based on multi-source data to solve the problem of inaccurate fault detection.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a photovoltaic inverter fault detection method based on multi-source data, comprising: The original current signal is acquired, and the intrinsic mode function sequence is obtained from the original current signal according to the preset extraction rules. The intrinsic modulus function sequence is arranged according to a preset arrangement rule. If the index in the intrinsic modulus function sequence is higher than the preset index threshold, it is marked as a risk intrinsic modulus function. The risk intrinsic mode function is decomposed using a preset decomposition method to obtain the fine sub-band frequency component coefficients; Energy distribution features are extracted from the frequency component coefficients of the fine sub-band. If the local energy deviation value of the energy distribution feature exceeds a preset quantization threshold, the time period corresponding to the energy distribution feature is marked as a current fluctuation segment. Envelope demodulation is performed on the current fluctuation segment to obtain the instantaneous amplitude and instantaneous frequency; The local abnormal energy value is calculated using the instantaneous amplitude and the instantaneous frequency to form an aging trend curve; A fault evolution sequence is generated based on the current fluctuation range and the aging trend curve. Critical abrupt change points in the fault evolution sequence are identified. Based on the critical abrupt change points, an early warning sequence is obtained and output.

[0007] Secondly, the present invention provides a photovoltaic inverter fault detection system based on multi-source data, comprising: The intrinsic modulus function acquisition module is used to acquire the original current signal and obtain the intrinsic modulus function sequence from the original current signal according to the preset extraction rules. The intrinsic modulus function screening and marking module is used to arrange the intrinsic modulus function sequence according to a preset arrangement rule. If the index in the intrinsic modulus function sequence is higher than the preset index threshold, it is marked as a risk intrinsic modulus function. The risk intrinsic modulus function decomposition module is used to decompose the risk intrinsic modulus function using a preset decomposition method to obtain the fine sub-band frequency component coefficients. The energy extraction fluctuation marking module is used to extract energy distribution features from the frequency component coefficients of the fine sub-band. If the local energy deviation value of the energy distribution feature exceeds a preset quantization threshold, the time period corresponding to the energy distribution feature is marked as a current fluctuation segment. The fluctuation segment demodulation module is used to perform envelope demodulation on the current fluctuation segment to obtain the instantaneous amplitude and instantaneous frequency; An abnormal energy calculation trend module is used to calculate local abnormal energy values ​​using the instantaneous amplitude and the instantaneous frequency, thereby forming an aging trend curve; The fault sequence early warning module is used to generate a fault evolution sequence based on the current fluctuation range and the aging trend curve, identify critical mutation points in the fault evolution sequence, obtain an early warning sequence based on the critical mutation points, and output it.

[0008] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the photovoltaic inverter fault detection method based on multi-source data as described above.

[0009] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the photovoltaic inverter fault detection method based on multi-source data described above.

[0010] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a multidimensional time series matrix by collecting raw current signals through multiple sensors, obtains the intrinsic mode function sequence through interpolation fitting, mean stripping and Gaussian white noise iterative screening, and then combines instantaneous frequency sorting and preset threshold screening of risk intrinsic mode functions, thus solving the problem that the existing technology cannot separate multiple frequency components of non-stationary signals and misses weak transient disturbances of early faults.

[0011] (2) This invention obtains instantaneous amplitude and instantaneous frequency by demodulating the envelope of the current fluctuation segment, generates a local energy density sequence by weighted product, obtains local abnormal energy value by integrating with the benchmark health model, maps to the multidimensional aging feature space to calculate the degree of abnormal deviation, and forms an aging trend curve by spline interpolation fitting. This solves the problems of existing technologies relying on fixed models, being unable to adapt to non-stationary signals and lacking fault evolution tracking capabilities.

[0012] (3) This invention generates fine sub-band frequency component coefficients by decomposing the risk intrinsic mode function of wavelet packet, calculates sub-band energy entropy to screen abnormal sub-bands, marks current fluctuation segments by combining local energy deviation values, and identifies critical mutation points by relying on the fault state transition matrix. This solves the problem that the existing technology cannot distinguish between normal fluctuations and abnormal features and is prone to misjudgment. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the photovoltaic inverter fault detection method based on multi-source data provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a photovoltaic inverter fault detection system based on multi-source data provided in the second embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Reference Figure 1 The first embodiment of the present invention provides a photovoltaic inverter fault detection method based on multi-source data, including the following steps: S11, Acquire the original current signal and obtain the intrinsic mode function sequence from the original current signal according to the preset extraction rules; S12, the intrinsic mode function sequence is arranged according to a preset arrangement rule. If the index in the intrinsic mode function sequence is higher than the preset index threshold, it is marked as a risk intrinsic mode function. S13, decompose the risk intrinsic mode function using a preset decomposition method to obtain the fine sub-band frequency component coefficients; S14, extract energy distribution features from the frequency component coefficients of the fine sub-band. If the local energy deviation value of the energy distribution feature exceeds the preset quantization threshold, mark the time period corresponding to the energy distribution feature as the current fluctuation segment. S15, perform envelope demodulation on the current fluctuation section to obtain instantaneous amplitude and instantaneous frequency; S16, calculate the local abnormal energy value using the instantaneous amplitude and the instantaneous frequency to form an aging trend curve; S17. Generate a fault evolution sequence based on the current fluctuation range and the aging trend curve, identify critical mutation points in the fault evolution sequence, and obtain and output an early warning sequence based on the critical mutation points.

[0016] In step S11, the acquisition of the raw current signal and the extraction of the intrinsic mode function sequence from the raw current signal using a preset extraction rule include: Acquire raw current signals from multiple sensors and construct a multidimensional time series matrix; Initial screening parameters are generated based on the multidimensional time series matrix. Interpolation fitting and mean stripping are performed using the initial screening parameters to obtain the component to be determined. If the component to be determined meets the preset criterion threshold, it is confirmed as a candidate intrinsic mode function. Gaussian white noise is introduced to denoise the candidate intrinsic mode function to obtain the denoised component. The denoised components are filtered according to preset indicators to obtain the intrinsic mode function sequence.

[0017] Specifically, by deploying multiple current sensors at key monitoring points of the photovoltaic inverter, raw current signals are collected synchronously. Assuming three sensors are used, each sensor continuously collects data at a frequency of 5,000 data points per second for 10 seconds, then the final constructed multidimensional time series matrix has a structure of 3 rows and 50,000 columns, where the rows correspond to different sensors and the columns correspond to various sampling points in the time dimension. This matrix intuitively presents the current changes at different monitoring locations at different time points, providing comprehensive data support for subsequent analysis.

[0018] It should be noted that when generating initial screening parameters based on the aforementioned multidimensional time series matrix, the screening range needs to be determined in conjunction with the statistical characteristics of the signal. The initial screening parameters are set by calculating the mean and standard deviation of the signal collected by each sensor, using the mean as the center and multiples of the standard deviation as boundaries. For example, if the mean of a sensor's signal is 5.2 mA and the standard deviation is 0.8 mA, considering that signal fluctuations during normal operation are typically within twice the standard deviation of the mean, the initial screening parameters are set to 3.6-6.8 mA. This range covers the normal fluctuation range of the signal and can initially filter out obvious outliers, allowing subsequent processing to focus on effective signal characteristics.

[0019] In this invention, when performing interpolation fitting and mean stripping using initial screening parameters, linear interpolation is used to fill in any missing or discontinuous data points in the matrix. For example, if a sensor has missing data at time point 500, the filling value is calculated using valid data from adjacent time points. For instance, for sensor 2, if there is a missing data point at time point 500, linear interpolation is used to fill it. First, valid data from adjacent time points before and after the missing point are extracted. The current value at time point 499 is 5.1 mA, and the current value at time point 501 is 5.3 mA. When processing the current data collected by the sensors, if a missing data point is encountered, linear interpolation can be used to calculate the filling value for that missing point. The specific calculation logic is as follows: first, the current value of the adjacent time point before the missing point is added, and the difference between the current value of the adjacent time point after the missing point and the current value of the adjacent time point before the missing point is divided by 2 to obtain the filling value for the missing point. Substituting the data, the current value at time point 500 is calculated to be 5.2 mA.

[0020] Mean stripping involves subtracting the mean value from the current signal collected by each sensor to obtain the component to be determined, which only reflects the dynamic fluctuations of the signal. This step can effectively remove the DC bias in the signal, highlight the changing characteristics of the current signal, and lay the foundation for the subsequent determination of the intrinsic mode function.

[0021] For example, taking sensor 1 as an example, the average current data collected at 1000 time points is calculated to be 5.0 mA. The original current value of sensor 1 at the first time point is 5.2 mA, and after removing the mean, the component to be determined is 5.2 - 5.0 = 0.2 mA; the original current value at the 200th time point is 4.8 mA, and the component to be determined is 4.8 - 5.0 = -0.2 mA; the original current value at the 800th time point is 5.5 mA, and the component to be determined is 5.5 - 5.0 = 0.5 mA. Through mean stripping, the 5.0 mA DC bias in the original signal of sensor 1 is completely removed, and the component to be determined only retains the dynamic fluctuations of the current signal (such as 0.2 mA, -0.2 mA, and 0.5 mA). These fluctuations precisely reflect the current change characteristics during inverter operation, including normal load fluctuations and transient disturbances caused by potential faults. Similarly, after mean stripping, sensors 2 and 3 also yield components to be determined containing only dynamic fluctuations.

[0022] When calculating the energy change of the component to be determined, it is necessary to combine the dynamic fluctuation characteristics of the photovoltaic inverter current signal and adopt the core logic of combining time window energy statistics and adjacent window difference quantization.

[0023] For example, based on the signal disturbance period, the transient disturbance of switch aging typically lasts 50-100 milliseconds, corresponding to 250-500 sampling points at a 5000Hz sampling frequency. Therefore, a window length of 500 sampling points and a step size of 250 sampling points are set. Then, for each time window, the formula is used... Calculate energy For the first One window of energy, For the first in the window The sum of the squares of the sampled component values ​​directly reflects the energy intensity of the dynamic fluctuations; finally, the absolute value of the energy difference between adjacent windows is calculated, and this difference is the energy change of the component. If the sum of the squares of the squares of all adjacent windows is... If all values ​​are less than the preset energy convergence threshold of 0.5, it indicates that the component meets the criteria for the candidate intrinsic mode function. For example, in the components to be determined by sensor 1, the energy of the sampling window from the 1st to the 500th sampling point... Energy of sampling points 251-750 Energy changes Energy of sampling points 751-1250 ,and Energy changes And all adjacent windows If all threshold requirements are met, the component to be determined is confirmed as a candidate intrinsic modulus function.

[0024] In this embodiment, if the component to be determined meets a preset criterion threshold, it is confirmed as a candidate intrinsic mode function. Here, the preset criterion threshold is set to an energy convergence threshold of 0.5. This threshold is determined based on the energy analysis of current signals from a large number of photovoltaic inverters during normal operation. Under normal circumstances, when the energy change of the component to be determined is less than 0.5, its signal characteristics tend to stabilize. Further decomposition will lead to signal distortion. Therefore, 0.5 is used as the energy convergence threshold. If the energy change of the component to be determined meets this threshold requirement, it can be considered to have the basic conditions to be a candidate intrinsic mode function. Since the candidate intrinsic mode functions may have spectral overlap, Gaussian white noise needs to be introduced for iterative screening to obtain denoised components. The mean of the introduced Gaussian white noise is set to 0, and the standard deviation is set to 0.1. This parameter setting is based on the statistical analysis of common noises in photovoltaic inverter current signals. Gaussian white noise of this specification can effectively break the spectral overlap modes without causing excessive interference to the useful signal. Through multiple iterative screenings, the influence of residual noise on the signal is gradually weakened, making the spectral characteristics of each candidate intrinsic mode function clearer.

[0025] It is worth noting that the denoised components are selected based on a preset criterion to obtain the intrinsic mode function (IMF) sequence. Here, the preset criterion is orthogonality, specifically with a correlation coefficient threshold set to 0.1. The orthogonality criterion is chosen to ensure the independence of the final IMF sequence, avoiding redundant information from interfering with subsequent fault detection. The correlation coefficient threshold of 0.1 is determined based on extensive experimental verification. When the correlation coefficient between two denoised components is below 0.1, it indicates extremely low linear correlation between them, with no overlap in signal characteristics, and each component can reflect different frequency components of the current signal. By calculating the correlation coefficient between each denoised component, components with correlation coefficients below 0.1 are selected, ultimately forming an IMF sequence with independent characteristics that accurately reflects different frequency components of the current signal, providing a reliable signal basis for the subsequent extraction of photovoltaic inverter fault characteristics.

[0026] For example, after iterative sieving using Gaussian white noise, three denoised components are obtained. , , These correspond to signal characteristics in different frequency bands. Orthogonal components need to be selected by calculating the Pearson correlation coefficient between each pair of components. The total number of sampling points is 50,000, and the mean is the sum of the total number of sampling points for each denoised component divided by the total number of sampling points. Substituting the actual data of the three components into the calculation, the component... and The correlation coefficient is 0.07, and the components... and The correlation coefficient is 0.05, and the components... and The correlation coefficients were all 0.09, lower than the preset threshold of 0.1, indicating extremely low linear correlation among the three components, non-overlapping signal features, and low component density. Focusing on the 800-950Hz frequency band (switching) Action signal), component Focusing on the 950-1200Hz frequency band (switching) Action signal), component Focusing on the 1200-1500Hz frequency band (residual features after electromagnetic interference filtering), all three denoising components are retained to form the final intrinsic mode function sequence, providing independent and non-redundant signal support for subsequent extraction of switch aging transient disturbance features and fault location.

[0027] In step S12, arranging the intrinsic modulus function sequence according to a preset arrangement rule, and marking any index in the intrinsic modulus function sequence as a risky intrinsic modulus function, includes: Calculate the instantaneous frequency of each component of the intrinsic mode function sequence, and sort the intrinsic mode function sequence in descending order according to the instantaneous frequency to obtain the initial frequency mode components; If the instantaneous frequency of the initial frequency mode component is higher than a preset frequency threshold, then the intrinsic mode function sequence corresponding to the initial frequency mode component is marked as a risky intrinsic mode function.

[0028] Calculating the instantaneous frequency of each component of the intrinsic mode function sequence is the core prerequisite for subsequent sorting and risk labeling. The calculation of instantaneous frequency is based on the principle of signal time-frequency analysis. By extracting the frequency characteristics of each intrinsic mode function component at different time points, the dynamic change rate of the current signal corresponding to that component can be intuitively reflected.

[0029] For example, for the five intrinsic mode function sequences obtained from the photovoltaic inverter current signal, the core of calculating the instantaneous frequency of the intrinsic mode functions across the entire time axis is to analyze the time-varying characteristics of the non-stationary signal through Hilbert transform. The specific process is as follows: For the time-domain signal of a single intrinsic mode function, such as IMF3 focusing on the high-frequency band among the five intrinsic mode functions, with a total of 50,000 sampling points corresponding to a 10-second sampling duration, first perform the Hilbert transform to obtain the original real signal. Generate the corresponding imaginary part signal The formula is Construct analytic signals in complex form , The imaginary unit is used; then the phase angle of the analytic signal is calculated using the arctangent function. The phase angle fluctuates in real time with changes in signal frequency; finally, the time derivative of the phase angle is calculated (i.e., (reflecting the rate of phase change), and divided by After standardization, the instantaneous frequency formula is obtained. .

[0030] In this embodiment, the phase angle of IMF3 is 1.2 radians at the 10,000th sampling point (corresponding to 2 seconds), and the phase angle is 4.5 radians at the 10,050th sampling point (corresponding to 2.01 seconds). The time interval is... Phase change Substituting into the formula, the instantaneous frequency at that moment is approximately 52.5 Hz. After traversing all sampling points, it can be found that the peak instantaneous frequency reaches 1520 Hz at the 25000-30000th sampling point (corresponding to 5-6 seconds), and is stable at 1480-1500 Hz at other times. After the same calculation, the instantaneous frequencies of other intrinsic mode functions are all concentrated in 750-1000 Hz, thus locking the target component corresponding to the high-frequency feature.

[0031] It is worth noting that when sorting the intrinsic mode function sequence in descending order based on instantaneous frequency, the average instantaneous frequency of each component is used as the sorting criterion, with the component with the highest instantaneous frequency placed first, and so on, until the initial frequency mode component sequence is obtained. This sorting method can quickly identify the high-frequency fluctuation components in the current signal, and these high-frequency components are often directly related to transient events such as switching actions and device aging inside the inverter, and are important carriers of early fault symptoms.

[0032] If the instantaneous frequency of the initial frequency mode component is higher than the preset frequency threshold, its corresponding intrinsic mode function sequence is marked as a risky intrinsic mode function. Here, the preset frequency threshold is set to 1200Hz. This threshold is determined based on the working principle of the photovoltaic inverter and statistical analysis of actual fault data. During normal operation, the switching frequency of the inverter's power switching devices is typically stable between 750-1000Hz, and the instantaneous frequency of the corresponding high-frequency components of the current signal is mostly lower than 1200Hz. However, when the switching devices experience slight aging, the synchronicity of the switching action decreases, resulting in transient current fluctuations with frequencies higher than 1200Hz. Through extensive collection of current signal data under normal inverter operation and switch aging conditions, it has been verified that 1200Hz as the frequency threshold can effectively distinguish between normal high-frequency components and fault-related high-frequency components, with a false positive rate of less than 5%. Therefore, 1200Hz is set as the preset frequency threshold. When the instantaneous average frequency of a certain initial frequency mode component exceeds 1200Hz, it indicates that the component is very likely to contain transient disturbance signals caused by switch aging. It needs to be marked as a risk intrinsic mode function to provide a clear analysis object for further extraction of fault features and location of aging fluctuation sections.

[0033] In step S13, the step of decomposing the risk intrinsic mode function using a preset decomposition method to obtain the fine sub-band frequency component coefficients includes: The frequency model component corresponding to the risk intrinsic mode function is subjected to multi-level wavelet packet decomposition to generate a wavelet packet decomposition tree structure. Extract the initial subband frequency component coefficients of the wavelet packet decomposition tree structure and construct the subband coefficient matrix; Calculate the energy entropy value of the sub-band coefficient matrix. If the energy entropy value exceeds a preset energy entropy threshold, mark the initial sub-band frequency component coefficients as abnormal sub-band frequency component coefficients. The frequency component coefficients of the anomalous subband are reconstructed by inverse wavelet packet transform to obtain the frequency component coefficients of the fine subband.

[0034] Performing multi-level wavelet packet decomposition on the frequency mode components corresponding to the risk intrinsic mode function to generate a wavelet packet decomposition tree structure is a key step in achieving refined signal processing. Compared with traditional wavelet decomposition, wavelet packet decomposition can equally divide the high-frequency and low-frequency components of the signal. Because it usually has a wide frequency distribution characteristic, it is more suitable for transient disturbance signals that may be contained in the risk intrinsic mode function.

[0035] Specifically, considering the frequency range of the photovoltaic inverter current signal, the high-frequency components during normal operation are mostly between 500-1000Hz, while the transient signal during a fault may extend to above 1500Hz. Therefore, a four-layer wavelet packet decomposition is chosen. The number of layers is determined based on a balance between frequency resolution requirements and computational efficiency. The four-layer decomposition divides the original signal into 16 terminal nodes (2...). 4 Each node corresponds to a frequency subband width of approximately 156.25Hz (assuming a sampling frequency of 5000Hz, and according to the Nyquist criterion, the effective frequency range is 0-2500Hz; after four levels of decomposition, the subband width is approximately 2500 / 16≈156.25Hz). This approach accurately captures transient characteristics of different frequency bands while avoiding a surge in computation due to excessive decomposition levels. The db4 wavelet basis function is used during the decomposition process. This wavelet basis has excellent time-frequency localization characteristics and strong ability to capture abrupt signals. Experiments have verified that its decomposition error for millisecond-level current fluctuations caused by aging of photovoltaic inverter switches is less than 3%, thus it is used as the preset wavelet basis function. Through multi-level decomposition, a wavelet packet decomposition tree structure covering the entire frequency band is finally generated, with each terminal node of the tree corresponding to the decomposition result of a specific frequency subband.

[0036] It is worth noting that extracting the initial sub-band frequency component coefficients of the wavelet packet decomposition tree structure and constructing the sub-band coefficient matrix requires traversing all terminal nodes of the decomposition tree and arranging the coefficients corresponding to each node according to the time dimension to form the sub-band coefficient matrix. For example, for 16 terminal nodes after four-level decomposition, assuming the original signal acquisition time is 10 seconds and the sampling frequency is 5000Hz, with each node corresponding to 50,000 time points, the sub-band coefficient matrix has a structure of 16 rows and 50,000 columns. The rows represent different frequency sub-bands, and the columns represent the time series. The amplitude of each element in the matrix reflects the signal strength of the corresponding sub-band at the corresponding time point. This matrix form can intuitively present the localization characteristics of the signal in the time and frequency domain. For example, the coefficient matrix of a certain terminal node (corresponding to the frequency sub-band 781.25-937.5Hz) shows a significant amplitude peak at the 10000-12500th time point (corresponding to 2-2.5 seconds), indicating that there is abnormal signal fluctuation in this frequency band during this time period, providing an intuitive basis for subsequent abnormal sub-band screening.

[0037] In this invention, the energy entropy value of the sub-band coefficient matrix is ​​calculated to determine whether the sub-band is abnormal. The calculation of energy entropy is based on the information entropy theory. The complexity of the signal is reflected by quantifying the uniformity of energy distribution within the sub-band. The energy of a normal, stable signal is more uniformly distributed within the sub-band, and the energy entropy value is higher. However, the energy of a signal containing transient disturbances will be concentrated in a specific time or frequency range, and the energy entropy value is lower.

[0038] In this embodiment, the energy distribution probability, i.e., the proportion of energy in the total energy of the sub-band at a certain moment, is first calculated for the coefficient sequence of each sub-band. In the sub-band energy entropy calculation for photovoltaic inverter fault detection, taking the 781.25-937.5Hz fine sub-band as an example, this sub-band contains real coefficients at 10 time points, corresponding to a sampling duration of 0.002 seconds and a sampling frequency of 5000Hz. The coefficient values ​​are 0.02, 0.03, 0.02, 0.4, 0.38, 0.42, 0.03, 0.02, 0.03, and 0.02 mA, respectively. The 10 points not only fully cover all steps, ensuring that the calculation logic of each step (accumulation, proportional conversion, logarithmic operation) can be intuitively reflected through specific values, but also avoid the cumbersome formula derivation and numerical calculation process caused by excessive data volume (such as hundreds or thousands of points), thus preventing excessive redundant data from obscuring the core principle of how energy distribution affects entropy values. First, according to the formula... The energy at each time point was calculated to be 0.0004, 0.0009, 0.0004, 0.16, 0.1444, 0.1764, 0.0009, 0.0004, 0.0009, and 0.0004 mA. 2 Then, these energies are summed to obtain the total energy of the subband. Press again The energy distribution probabilities at each time point were calculated to be approximately 0.0008, 0.0019, 0.0008, 0.3295, 0.2974, 0.3633, 0.0019, 0.0008, 0.0019, and 0.0008, respectively; finally, these probabilities were substituted into the energy entropy formula. The calculated energy entropy of the sub-band is approximately 1.52 bits, which is lower than the preset threshold of 1.8 bits. Therefore, it can be determined that there is a fault transient disturbance in the sub-band, and the corresponding time period should be marked as the current fluctuation segment.

[0039] It should be noted that the preset energy entropy threshold here is set to 1.8 bits. This threshold is determined based on the statistical analysis of sub-band energy entropy under a large number of photovoltaic inverters in normal and fault states. Under normal operation, the energy entropy values ​​of each sub-band are mostly above 2.0 bits, and the energy distribution is uniform. When aging of switching devices causes transient disturbances, the energy of the sub-band where the disturbance occurs will be concentrated in a short period of time, and the energy entropy value will drop below 1.8 bits. Through verification of 500 sets of normal data and 300 sets of fault data, the anomaly identification accuracy of 1.8 bits as the threshold can reach 92%, so it is set as the preset energy entropy threshold. If the energy entropy value of a certain sub-band coefficient matrix is ​​lower than 1.8 bits, the initial sub-band frequency component coefficients corresponding to that sub-band are marked as abnormal sub-band frequency component coefficients, locking in the sub-band that may contain fault transient signals.

[0040] The frequency component coefficients of the anomalous subband are reconstructed by inverse wavelet packet transform to obtain the frequency component coefficients of the fine subband. Inverse wavelet packet transform is the inverse process of wavelet packet decomposition. By recombining the coefficients of the anomalous subband, the time domain signal corresponding to the subband is restored.

[0041] For example, taking the reconstruction of the anomalous subband node (4,3) (corresponding to 781.25-937.5Hz) as an example, first take the coefficient sequence of this node, such as [0.42,0.38,0.51,0.45], and the scaling function coefficients of the db4 wavelet basis, approximately [0.48296,0.83652,0.22414,-0.12941]. Perform inner product operation, multiply each digit and sum them, which is approximately equal to 0.5768, to obtain one coefficient of the parent node (3,1). Traversing all subband coefficients can generate the complete coefficient sequence of the parent node (3,1). Then, proceed to the reverse recursive step. Since the coefficients of other non-anomalous subbands have been set to zero, only the parent node (3,1) is used. The coefficients are then multiplied by the scaling function coefficients to obtain the coefficients of the grandfather node (2,0). The logic for generating the coefficients of the upper-level nodes is calculated according to this inner product, and the process is repeated layer by layer upwards to the root node. Finally, the coefficient sequence output by the root node is the reconstructed fine sub-band frequency component coefficients. The final reconstructed signal is the fine sub-band frequency component coefficient sequence of 781.25-937.5Hz. This sequence exhibits obvious pulse characteristics in the 2.05-2.15 second time period, with the amplitude rising sharply from 0.08 mA to 0.32 mA and then falling back. It accurately preserves the transient current fluctuation information caused by switch aging and can be directly used for subsequent energy distribution feature extraction and current fluctuation segment marking.

[0042] In step S14, extracting energy distribution features from the fine sub-band frequency component coefficients, and marking the time period corresponding to the energy distribution features as a current fluctuation segment if the local energy deviation value of the energy distribution features exceeds a preset quantization threshold, includes: Calculate the energy values ​​of the frequency component coefficients of the fine sub-band to form energy distribution characteristics; The local energy deviation value is calculated by comparing the energy distribution characteristics with the preset benchmark energy distribution. If the local energy deviation exceeds a preset quantization threshold, the time period corresponding to the energy distribution feature is marked as a current fluctuation segment.

[0043] In this invention, calculating the energy values ​​of the fine sub-band frequency component coefficients to form energy distribution characteristics requires energy statistics for each fine sub-band frequency component coefficient based on the principle of signal energy calculation, using a time window. Specifically, a sliding time window is used, with a window length of 0.1 seconds and a step size of 0.05 seconds. This parameter is determined based on the continuous characteristics of transient disturbances in photovoltaic inverters. Transient fluctuations caused by switch aging typically last from 50 to 100 milliseconds. A 0.1-second window can fully cover a single disturbance while avoiding the superposition of multiple disturbance signals due to an excessively long window. A 0.05-second step size enables continuous monitoring of the signal, avoiding the omission of disturbance details.

[0044] It is worth noting that the energy value is calculated by summing the squares of the fine subband coefficients within each window using the following formula: ,in For the fine subband coefficients within the window, This represents the number of coefficients within the window. For example, for the fine subband frequency component coefficients, corresponding to the frequency subband 781.25-937.5 Hz, the sum of squares of the coefficient sequence within a 2-2.1 second window is 0.9 mA. 2 The sum of squares within the 2.05-2.15 second window is 1.2 mA. 2 The energy values ​​of all windows are arranged chronologically to form an energy distribution characteristic reflecting the energy change of the fine sub-band over time. This characteristic visually presents the time periods of concentrated energy, providing a basis for subsequent anomaly detection. The energy distribution characteristic is compared with a preset benchmark energy distribution to calculate the local energy deviation value. The preset benchmark energy distribution needs to be constructed based on historical data under normal operating conditions of the photovoltaic inverter. Energy distribution data of the same fine sub-band of the inverter under fault-free and stable load conditions are collected, and energy distribution characteristics of 100 normal operating cycles are selected. The average energy value corresponding to each time window is calculated to form a benchmark energy vector. For example, the benchmark energy for the 2-2.1 second window is 0.5 mA. 2 The reference energy for the 2.05-2.15 second window is 0.6 mA. 2 .

[0045] In this photovoltaic inverter fault detection method, milliampere 2 It is not a standard unit of energy in physics, but rather a relative reference unit designed in signal processing to simplify the comparison of the strength of current signals. The energy distribution characteristics in the method essentially quantify the relative strength of the signal by the square of the current signal amplitude. For example, when the current amplitude is 0.3 mA, the relative signal strength is 0.09 mA. 2 During the fault, the amplitude increases to 0.8 mA, and the relative intensity is 0.64 mA. 2 Through milliampere 2 Numerical differences can intuitively determine changes in signal strength, i.e., whether abnormal fluctuations exist. Physical energy units are not used because fault detection only needs to focus on the instantaneous fluctuations of the current signal itself; there is no need to introduce additional parameters such as voltage or time. Milliamperes (mA) are used. 2 As an organization, it can directly and conveniently compare the signal differences between normal and abnormal states, which is more suitable for the needs of fault feature extraction.

[0046] In this embodiment, the local energy deviation is calculated using the absolute value of the difference. Specifically, for each time window, the absolute value of the difference between the energy value of that window in the current energy distribution characteristics and the energy value of the corresponding window in the baseline energy vector is calculated. For example, the current energy in the 2-2.1 second window is 0.9 mA. 2 The reference energy is 0.5 mA.2 The local energy deviation is 0.4 mA. 2 .

[0047] It is worth noting that this calculation can accurately quantify the degree of deviation between the energy level and the normal state in each time period. If the local energy deviation exceeds a preset quantization threshold, the corresponding time period is marked as a current fluctuation segment. Here, the preset quantization threshold is set to 0.5 mA. 2 The threshold was determined based on statistical analysis of energy deviations under a large number of normal and fault conditions. During normal operation, affected by minor load fluctuations and electromagnetic interference, the local energy deviation is mostly within 0.3 mA. 2 The following applies: When aging of switching devices causes transient disturbances, the energy will increase sharply in a short period of time, and the deviation from the quantization value will exceed 0.5 mA. 2 Through verification using 300 sets of normal data and 200 sets of switch aging fault data, 0.5 mA 2 The accuracy rate of identifying abnormal segments using this threshold is as high as 93%, with a false positive rate of less than 4% (avoiding misclassifying normal, minor fluctuations as abnormal). Therefore, it is set as the preset quantization threshold. For example, a deviation of 0.4 mA from the quantization value within a 2-2.1 second window is considered normal. 2 If the threshold is not exceeded, it is not marked as a fluctuation range; the deviation of the quantized value from the 2.05-2.15 second window is 0.6 mA. 2 If the threshold is exceeded, the time period of 2.05-2.15 seconds corresponding to the window is marked as the current fluctuation segment. At the same time, the fine sub-band signal characteristics within this time period are associated to provide a clear time location for further analysis of the cumulative trend of aging symptoms and fault evolution.

[0048] In step S15, the envelope demodulation of the current fluctuation segment to obtain the instantaneous amplitude and instantaneous frequency includes: Obtain the sub-band signal sequence corresponding to the current fluctuation segment, perform Hilbert transform operation on the sub-band signal sequence, and construct an analytical signal sequence; Calculate the magnitude of the analytic signal sequence to obtain the instantaneous amplitude of the sub-band signal sequence; The instantaneous frequency of the sub-band signal sequence is obtained by taking the time derivative of the phase of the analytical signal sequence.

[0049] It should be noted that obtaining the sub-band signal sequence corresponding to the current fluctuation segment is the basis of envelope demodulation. Assume that this sub-band signal sequence originates from the current fluctuation segment of 2.05-2.15 seconds marked in step S14, and corresponds to the fine sub-band frequency component of 781.25-937.5Hz obtained from the reconstruction in step S13. Since the current fluctuation segment has been locked to include the time period containing transient disturbances from switch aging, extracting the sub-band signal sequence within this time period allows us to focus on fault-related signal components and eliminate interference from irrelevant time intervals. For example, for the current fluctuation segment of 2.05-2.15 seconds, extracting the signal sequence of the 781.25-937.5Hz sub-band within this time period reveals that its original waveform exhibits short-time pulse characteristics, with amplitude rapidly fluctuating between 0.08-0.3 mA. This sequence is the target for subsequent Hilbert transform processing.

[0050] It is worth noting that the Hilbert transform operation is performed on the sub-band signal sequence to construct an analytic signal sequence. The core function of the Hilbert transform is to generate the corresponding imaginary part signal for the real signal, forming an analytic signal in complex form, which makes it easier to extract the instantaneous amplitude and instantaneous frequency of the signal.

[0051] In this invention, the magnitude of the analytical signal sequence is calculated to obtain the instantaneous amplitude of the sub-band signal sequence. The magnitude of the analytical signal is essentially the length of the vector formed by the real and imaginary parts of the signal. This magnitude represents the instantaneous amplitude of the signal at each time point, directly reflecting the changes in the signal's energy strength. For example, for an analytical signal sequence of 2.05-2.15 seconds, the calculated instantaneous amplitude is 0.08 mA at 2.05 seconds, reaches a peak of 0.32 mA at 2.08 seconds, rapidly drops to 0.1 mA at 2.12 seconds, and recovers to 0.09 mA at 2.15 seconds. This trend perfectly matches the transient current fluctuations caused by switch aging. The contact resistance of the aging switch increases instantaneously, leading to a short-term peak in current, which then gradually decreases as the device heats up and stabilizes. Extracting the instantaneous amplitude transforms the signal's energy fluctuations into intuitive numerical changes, providing a quantitative basis for subsequently judging the severity of aging symptoms.

[0052] It is worth noting that the instantaneous frequency of the subband signal sequence is obtained by taking the time derivative of the phase of the analytical signal sequence, as described in S12. The physical significance of this calculation process is to capture the dynamic change of the signal frequency over time. During normal operation, the operating frequency of the photovoltaic inverter switching devices is stable (e.g., 750-1000Hz), and the instantaneous frequency fluctuation range is extremely small. However, when the switches age, the switching speed of the devices decreases, which can lead to a shift or fluctuation in the instantaneous frequency. For example, within the current fluctuation range of 2.05-2.15 seconds, the time derivative calculation results of the analytical signal phase show that the instantaneous frequency is 1200Hz at 2.05 seconds, drops to 1150Hz at 2.08 seconds (peak amplitude), rises back to 1180Hz at 2.12 seconds, and recovers to 1200Hz at 2.15 seconds. This frequency fluctuation is due to the increased turn-on and turn-off delay of the aging switch, which causes a momentary shift in the frequency of the high-frequency sub-band signal. The extraction of the instantaneous frequency precisely captures this subtle change, which, together with the instantaneous amplitude, constitutes a key feature reflecting the aging state of the switch, providing data support for subsequent tracking of fault evolution trends.

[0053] In step S16, the calculation of local abnormal energy values ​​using the instantaneous amplitude and the instantaneous frequency to form an aging trend curve includes: A weighted product calculation is performed on the instantaneous amplitude and the instantaneous frequency to generate a local energy density sequence; The local energy density sequence is input into a preset benchmark health model, and an integral operation is performed to obtain the local abnormal energy value. The local abnormal energy value is mapped to a multidimensional aging feature space, and the Euclidean distance with the preset health state center vector is calculated to determine the degree of abnormal deviation. Spline interpolation fitting is performed on the degree of abnormal deviation to construct an aging trend curve.

[0054] It is worth noting that when performing a weighted product calculation on instantaneous amplitude and instantaneous frequency to generate a local energy density sequence, the weights must be set based on the contribution of both to the fault characteristics. Instantaneous amplitude directly reflects the strength of signal energy and is the core characteristic of aging transient disturbances.

[0055] In this invention, instantaneous frequency reflects the signal frequency offset and helps to judge the stability of the disturbance. Based on the experimental data statistics of photovoltaic inverter fault detection, the instantaneous amplitude accounts for about 60% of the contribution of abnormal energy, and the instantaneous frequency accounts for about 40%. Therefore, the weight of instantaneous amplitude is set to 0.6 and the weight of instantaneous frequency is set to 0.4.

[0056] It should be noted that, to ensure the rationality and universality of normalization, the historical extreme values ​​of the photovoltaic inverter under all operating conditions should be selected as the normalization benchmark. This avoids distortion of the normalization results due to excessively small peak values ​​in local sections. First, historical monitoring data from long-term inverter operation is statistically analyzed to determine the maximum instantaneous amplitude under all operating conditions as 1.2 mA and the maximum instantaneous frequency under all operating conditions as 2000 Hz, covering all operating conditions including normal operation, slight aging, and severe aging, ensuring that the benchmark values ​​are compatible with all fluctuation ranges. For each time point within the current fluctuation range, the instantaneous amplitude... and instantaneous frequency Divide each value by the corresponding maximum value under all operating conditions to obtain the dimensionless normalized value; then calculate the weighted sum with an amplitude weight of 0.6 and a frequency weight of 0.4.

[0057] For example, at the peak instantaneous amplitude of 2.08 seconds, the instantaneous amplitude is 0.32 mA and the instantaneous frequency is 850 Hz. The calculated local energy density is 0.33. Arranging the local energy densities of all time points in sequence forms a local energy density sequence, which can quantify the anomalous energy intensity at each moment.

[0058] In this embodiment, the local energy density sequence is input into a preset benchmark health model and integral calculation is performed to obtain the local abnormal energy value. The preset benchmark health model needs to be constructed based on the normal operating state of the photovoltaic inverter. 100 sets of local energy density data of the same fine sub-band under fault-free operating conditions are collected, the average energy density of each time window is calculated, and a benchmark energy density curve is formed. This curve represents the energy distribution pattern under normal conditions.

[0059] It is worth noting that when the current local energy density sequence is input, the model first calculates the difference between the current energy density and the reference energy density. Positive residuals represent anomalous energy contributions, while negative residuals are considered normal fluctuations and treated as zero. Subsequently, the residual sequence is integrated over the current fluctuation range (e.g., 2.05-2.15 seconds), and the integration result is the local anomalous energy value, mathematically expressed as: , , The start and end times of the fluctuation range. Given the current energy density, This is the baseline energy density. For example, the average baseline energy density of the baseline health model in the 2.05-2.15 second window is 0.35, and the integral result of the current residual sequence in this interval is 0.08. This is the local abnormal energy value. The larger this value is, the more concentrated the abnormal energy is in the fluctuation range, and the more serious the aging of the switch is.

[0060] In this invention, local abnormal energy values ​​are mapped to a multidimensional aging feature space and the Euclidean distance between the value and the preset health state center vector is calculated to determine the degree of abnormal deviation. The multidimensional aging feature space needs to include multiple feature dimensions related to switch aging. In addition to local abnormal energy values, features such as instantaneous amplitude peak value and instantaneous frequency offset need to be included. Based on fault mechanism analysis, these features are strongly correlated with the degree of aging, forming a three-dimensional feature space. Feature 1: local abnormal energy value, Feature 2: instantaneous amplitude peak value, Feature 3: instantaneous frequency maximum offset.

[0061] It should be noted that the preset health state center vector is obtained based on a large amount of statistical data from normal operating conditions. For example, under normal conditions, the local abnormal energy value is close to 0, the instantaneous amplitude peak is about 0.1 mA, and the maximum instantaneous frequency offset is ≤5 Hz. Therefore, the center vector is set to [0, 0.1, 5]. The mapping process involves taking the characteristic values ​​of the current fluctuation segment, the local abnormal energy value of 0.08, the instantaneous amplitude peak of 0.32, and the maximum instantaneous frequency offset of 30 Hz, as a point in space, and calculating the Euclidean distance between this point and the center vector, which is d = 25.002. This distance represents the degree of abnormal deviation. The larger the distance, the more significant the difference between the current state and the healthy state, and the more obvious the signs of aging.

[0062] In this embodiment, spline interpolation fitting is performed on the abnormal deviation degree to construct the aging trend curve, which requires abnormal deviation degree data from multiple consecutive monitoring periods. Photovoltaic inverter fault detection is typically performed at fixed intervals, such as once per hour, with each analysis yielding an abnormal deviation degree value. If no fluctuation segment is detected, it is counted as 0. The deviation degree values ​​from multiple periods are arranged in chronological order to form discrete time-deviation degree data points. The spline interpolation fitting uses a cubic spline function. This function can pass through all discrete data points and form a smooth curve between adjacent points, avoiding the broken line distortion caused by linear interpolation and more accurately reflecting the gradual characteristics of aging. For example, the abnormal deviation degree values ​​for six consecutive monitoring periods (1-6 hours) are 5.2, 8.7, 12.3, 18.5, 25.0, and 31.2, respectively. The cubic spline interpolation function is composed of five interval cubic polynomial segments, with the time variable denoted as... The unit is h, and the degree of abnormal deviation is Satisfying the core constraints, that is, the constraints of each interval. All are cubic polynomials; The values ​​at all data points are completely consistent with the discrete values; the first and second derivatives of the polynomials in adjacent intervals are continuous at the connection points, ensuring a smooth curve; natural cubic spline boundary conditions are usually used, i.e., the first and last ends... and The second derivative is 0.

[0063] Specifically, on the interval [1,2], On [2,3], ...... On [5,6], The coefficients of each interval , , , It needs to be solved by solving a system of linear equations consisting of interpolation conditions for data points, continuity conditions for first and second derivatives of adjacent intervals, and natural boundary conditions.

[0064] It is worth noting that the aging trend curve obtained after cubic spline interpolation shows a non-linear upward trend. The deviation increases slowly in the first 3 hours and accelerates in the next 3 hours. This is consistent with the actual evolution of switch aging: the initial aging is mild and abnormal energy accumulates slowly; as aging intensifies, the switch contact resistance continues to increase, and the rate of abnormal energy accumulation accelerates. This trend curve can intuitively show the development direction and speed of aging symptoms, providing a dynamic basis for subsequent fault early warning.

[0065] It should be noted that the aging of core components of an inverter, such as switching devices and capacitors, is a gradual process and will not show significant abnormal deviations within minutes. A 1-hour period can capture the slow increase in deviation caused by aging (such as the low-speed changes in the initial 3 hours) and also cover the rapid increase in the later stage when aging intensifies, avoiding data redundancy due to too short a period and missing key trend inflection points due to too long a period.

[0066] If the cycle is too short, instantaneous fluctuations in equipment operation, such as power grid voltage fluctuations and sudden changes in light intensity, will be misjudged as aging signals; if the cycle is too long, the accelerated increase in the degree of deviation cannot be detected in time, and the timeliness of fault warning will be lost.

[0067] In step S17, generating a fault evolution sequence based on the current fluctuation range and the aging trend curve, identifying critical abrupt change points in the fault evolution sequence, and obtaining and outputting a warning sequence based on the critical abrupt change points include: Map the current fluctuation segment to the aging trend curve and extract a local evolution trajectory segment; Gradient calculation is performed on the local evolution trajectory segment to obtain the evolution rate vector. If the evolution rate vector exceeds a preset steady-state threshold, a feature vector sequence is extracted. The feature vector sequence is input into a preset fault state transition matrix to generate a fault evolution sequence, and critical mutation points in the fault evolution sequence are identified. Extract the timestamps of the critical mutation points, construct early warning entries and sort them by time to obtain the early warning sequence and output it.

[0068] In this invention, the current fluctuation segment is mapped to the aging trend curve to extract local evolution trajectory segments. First, the time correlation between the two needs to be established. The current fluctuation segment marks the abnormal signal within a specific time period (e.g., 2.05-2.15 seconds of the 3rd hour), while the aging trend curve records the degree of abnormal deviation in each period with the monitoring cycle (e.g., 1 hour / cycle) as the horizontal axis.

[0069] It is worth noting that during mapping, based on the time allocation of the current fluctuation segment, such as the 3rd hour monitoring cycle, the abnormal deviation data point corresponding to that cycle is located on the aging trend curve. Then, combined with the data from the 2nd and 4th hours of the adjacent cycle, a continuous segment centered on that cycle is extracted, such as the 1st to 5th hours, as a local evolution trajectory segment.

[0070] In this embodiment, the abnormal deviations for hours 1-5 are 5.2, 8.7, 12.3, 18.5, and 25.0, respectively. The corresponding local evolution trajectory segment is the curve segment formed by these five data points. This segment focuses on the aging trend of the current fluctuation section and its surrounding area, allowing for a direct observation of the changing pattern of the abnormal deviation, thus laying the foundation for subsequent evolution rate calculation. Gradient calculation is performed on the local evolution trajectory segment to obtain the evolution rate vector. The core of gradient calculation is to quantify the rate of change of the abnormal deviation during adjacent monitoring weeks, and the mathematical expression is: , For the first The degree of abnormal deviation in the cycle, The period interval is fixed at 1 hour here, and the evolution rate vector is composed of the rate values ​​of each adjacent period. For example, for a local trajectory segment from hour 1 to hour 5, the calculated evolution rate vector is [3.5, 3.6, 6.2, 6.5], which reflects the rate of change of the aging trend.

[0071] It is worth noting that the preset steady-state threshold is set to 5.0. This threshold is determined based on the steady-state evolution data of photovoltaic inverter switch aging. During the normal aging stage, the evolution rate is mostly stable at 3-4. When the rate exceeds 5.0, it indicates that aging has entered an accelerated stage, and the increase in the degree of abnormal deviation exceeds the steady-state range. If there are elements in the evolution rate vector that exceed the threshold, such as 6.2 and 6.5 in the above vector, the periodic features corresponding to these elements are extracted, such as the degree of abnormal deviation, local abnormal energy value, and instantaneous amplitude peak value in the 3rd-4th hour and the 4th-5th hour, forming a feature vector sequence. For example, the feature vector sequence is [[12.3,0.08,0.32],[18.5,0.12,0.45],[25.0,0.15,0.58]], which correspond to the degree of deviation, local abnormal energy value, and instantaneous amplitude peak value, respectively. This sequence contains the key features of the accelerated aging stage and provides input for the generation of the fault evolution sequence.

[0072] In this invention, a feature vector sequence is input into a preset fault state transition matrix to generate a fault evolution sequence. This preset fault state transition matrix is ​​constructed based on the fault mechanism and historical fault data of photovoltaic inverter switch aging. First, fault state levels are classified, such as normal state S0, slight aging S1, moderate aging S2, and severe aging S3. Then, the transition probabilities between each state are calculated; for example, the probability of transitioning from S1 to S2 is 70%, and the probability of transitioning from S2 to S3 is 85%. These probabilities are determined based on the statistical results of 500 sets of switch aging fault cases to ensure that the matrix reflects the actual fault evolution pattern. When the feature vector sequence is input into the matrix, the matrix matches the corresponding fault state according to each feature value, such as an abnormal deviation of 25.0 corresponding to S3 and 18.5 corresponding to S2, and concatenates them in chronological order to form a fault evolution sequence.

[0073] It is worth noting that the fault states corresponding to the feature vector sequence are S2, S3, and S3 respectively, and the fault evolution sequence is [S2→S3→S3]. To identify the critical mutation point in this sequence, we need to pay attention to the moment when the state first transitions from a low level to a high level. For example, the transition from S2 to S3 in the sequence occurs in the 4th hour monitoring period, and the evolution rate increases from 3.6 to 6.2, exceeding the steady-state threshold of 5.0. Therefore, the 4th hour monitoring period is marked as the critical mutation point. This point marks the transition of the fault from moderate aging to severe aging, which is the core basis for subsequent early warning.

[0074] In this embodiment, the timestamps of critical mutation points are extracted to construct early warning entries and sorted by time to obtain an early warning sequence, which is then output. The timestamps of critical mutation points need to be determined in conjunction with the monitoring period and the specific time of the current fluctuation segment. For example, the current fluctuation segment within the 4th hour monitoring period is 4 hours 1.10-1.20 seconds, so the timestamp is recorded as 4 hours 1.10-1.20 seconds. Each early warning entry must include a timestamp, the corresponding fault state (severe aging S3), and key characteristic values ​​(abnormal deviation degree 18.5, local abnormal energy value 0.12, instantaneous amplitude peak value 0.45). For example, the entry is "Time: 4 hours 1.10-1.20 seconds, Fault state: severe aging, Abnormal deviation degree: 18.5, Local abnormal energy value: 0.12, Instantaneous amplitude peak value: 0.45". If multiple critical mutation points exist (such as the S3→S4 transition occurring again in the 6th hour monitoring period), all early warning entries are arranged in chronological order by timestamp to form an early warning sequence, for example, the sequence is [4th hour entry, 6th hour entry]. Finally, the warning sequence will be output in the form of a visual report or text, providing maintenance personnel with a clear timeline of fault development and severity information, so as to facilitate timely intervention strategies and prevent the fault from deteriorating further and causing the inverter to shut down.

[0075] In summary, this invention discloses a photovoltaic inverter fault detection method based on multi-source data, which solves the problem of inaccurate fault detection.

[0076] Reference Figure 2 The second embodiment of the present invention provides a photovoltaic inverter fault detection system based on multi-source data, comprising: The intrinsic modulus function acquisition module is used to acquire the original current signal and obtain the intrinsic modulus function sequence from the original current signal according to the preset extraction rules. The intrinsic modulus function screening and marking module is used to arrange the intrinsic modulus function sequence according to a preset arrangement rule. If the index in the intrinsic modulus function sequence is higher than the preset index threshold, it is marked as a risk intrinsic modulus function. The risk intrinsic modulus function decomposition module is used to decompose the risk intrinsic modulus function using a preset decomposition method to obtain the fine sub-band frequency component coefficients. The energy extraction fluctuation marking module is used to extract energy distribution features from the frequency component coefficients of the fine sub-band. If the local energy deviation value of the energy distribution feature exceeds a preset quantization threshold, the time period corresponding to the energy distribution feature is marked as a current fluctuation segment. The fluctuation segment demodulation module is used to perform envelope demodulation on the current fluctuation segment to obtain the instantaneous amplitude and instantaneous frequency; An abnormal energy calculation trend module is used to calculate local abnormal energy values ​​using the instantaneous amplitude and the instantaneous frequency, thereby forming an aging trend curve; The fault sequence early warning module is used to generate a fault evolution sequence based on the current fluctuation range and the aging trend curve, identify critical mutation points in the fault evolution sequence, obtain an early warning sequence based on the critical mutation points, and output it.

[0077] It should be noted that the photovoltaic inverter fault detection system based on multi-source data provided in this embodiment of the invention is used to execute all the process steps of the photovoltaic inverter fault detection method based on multi-source data in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0078] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a photovoltaic inverter fault detection program based on multi-source data. When the processor executes the computer program, it implements the steps in the various embodiments of the photovoltaic inverter fault detection method based on multi-source data described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the intrinsic modulus function acquisition module.

[0079] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0080] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0081] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0082] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0083] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0084] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0085] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A photovoltaic inverter fault detection method based on multi-source data, characterized in that, include: The original current signal is acquired, and the intrinsic mode function sequence is obtained from the original current signal according to the preset extraction rules. The intrinsic modulus function sequence is arranged according to a preset arrangement rule. If the index in the intrinsic modulus function sequence is higher than the preset index threshold, it is marked as a risk intrinsic modulus function. The risk intrinsic mode function is decomposed using a preset decomposition method to obtain the fine sub-band frequency component coefficients; Energy distribution features are extracted from the frequency component coefficients of the fine sub-band. If the local energy deviation of the energy distribution features exceeds a preset quantization threshold, the time period corresponding to the energy distribution features is marked as a current fluctuation segment. Envelope demodulation is performed on the current fluctuation segment to obtain the instantaneous amplitude and instantaneous frequency; The local abnormal energy value is calculated using the instantaneous amplitude and the instantaneous frequency to form an aging trend curve; A fault evolution sequence is generated based on the current fluctuation range and the aging trend curve. Critical abrupt change points in the fault evolution sequence are identified. Based on the critical abrupt change points, an early warning sequence is obtained and output.

2. The photovoltaic inverter fault detection method based on multi-source data according to claim 1, characterized in that, The process of acquiring the raw current signal and obtaining the intrinsic modulus function sequence from the raw current signal using preset extraction rules includes: Acquire raw current signals from multiple sensors and construct a multidimensional time series matrix; Initial screening parameters are generated based on the multidimensional time series matrix. Interpolation fitting and mean stripping are performed using the initial screening parameters to obtain the component to be determined. If the component to be determined meets the preset criterion threshold, it is confirmed as a candidate intrinsic mode function. Gaussian white noise is introduced to denoise the candidate intrinsic mode function to obtain the denoised component. The denoised components are filtered according to preset indicators to obtain the intrinsic mode function sequence.

3. The photovoltaic inverter fault detection method based on multi-source data according to claim 1, characterized in that, The step of arranging the intrinsic modulus function sequence according to a preset arrangement rule, and marking any index in the intrinsic modulus function sequence as a risky intrinsic modulus function if the index is higher than a preset index threshold, includes: Calculate the instantaneous frequency of each component of the intrinsic mode function sequence, and sort the intrinsic mode function sequence in descending order according to the instantaneous frequency to obtain the initial frequency mode components; If the instantaneous frequency of the initial frequency modal component is higher than a preset frequency threshold, then the initial frequency modal component is marked as a risky intrinsic mode function.

4. The photovoltaic inverter fault detection method based on multi-source data according to claim 1, characterized in that, The step of decomposing the risk intrinsic mode function using a preset decomposition method to obtain the fine sub-band frequency component coefficients includes: The frequency model component corresponding to the risk intrinsic mode function is subjected to multi-level wavelet packet decomposition to generate a wavelet packet decomposition tree structure. Extract the initial subband frequency component coefficients of the wavelet packet decomposition tree structure and construct the subband coefficient matrix; Calculate the energy entropy value of the sub-band coefficient matrix. If the energy entropy value exceeds a preset energy entropy threshold, mark the initial sub-band frequency component coefficients as abnormal sub-band frequency component coefficients. The frequency component coefficients of the anomalous subband are reconstructed by wavelet packet inverse transform to obtain the fine subband frequency component coefficients.

5. The photovoltaic inverter fault detection method based on multi-source data according to claim 1, characterized in that, The step of extracting energy distribution features from the fine sub-band frequency component coefficients, and marking the time period corresponding to the energy distribution features as a current fluctuation segment if the local energy deviation value of the energy distribution features exceeds a preset quantization threshold, includes: Calculate the energy values ​​of the frequency component coefficients of the fine subband to form energy distribution characteristics; The local energy deviation value is calculated by comparing the energy distribution characteristics with the preset benchmark energy distribution. If the local energy deviation exceeds a preset quantization threshold, the time period corresponding to the energy distribution feature is marked as a current fluctuation segment.

6. The photovoltaic inverter fault detection method based on multi-source data according to claim 1, characterized in that, The process of envelope demodulating the current fluctuation segment to obtain the instantaneous amplitude and instantaneous frequency includes: Obtain the sub-band signal sequence corresponding to the current fluctuation segment, perform Hilbert transform operation on the sub-band signal sequence, and construct an analytical signal sequence; Calculate the magnitude of the analytic signal sequence to obtain the instantaneous amplitude of the sub-band signal sequence; The instantaneous frequency of the sub-band signal sequence is obtained by taking the time derivative of the phase of the analytical signal sequence.

7. The photovoltaic inverter fault detection method based on multi-source data according to claim 1, characterized in that, The step of calculating local abnormal energy values ​​using the instantaneous amplitude and the instantaneous frequency to form an aging trend curve includes: A weighted product calculation is performed on the instantaneous amplitude and the instantaneous frequency to generate a local energy density sequence; The local energy density sequence is input into a preset benchmark health model, and an integral operation is performed to obtain the local abnormal energy value. The local abnormal energy value is mapped to a multidimensional aging feature space, and the Euclidean distance with the preset health state center vector is calculated to determine the degree of abnormal deviation. Spline interpolation fitting is performed on the degree of abnormal deviation to construct an aging trend curve.

8. The photovoltaic inverter fault detection method based on multi-source data according to claim 1, characterized in that, The process of generating a fault evolution sequence based on the current fluctuation range and the aging trend curve, identifying critical abrupt change points in the fault evolution sequence, and obtaining and outputting a warning sequence based on the critical abrupt change points includes: Map the current fluctuation segment to the aging trend curve and extract a local evolution trajectory segment; Gradient calculation is performed on the local evolution trajectory segment to obtain the evolution rate vector. If the evolution rate vector exceeds a preset steady-state threshold, a feature vector sequence is extracted. The feature vector sequence is input into a preset fault state transition matrix to generate a fault evolution sequence, and critical mutation points in the fault evolution sequence are identified. Extract the timestamps of the critical mutation points, construct early warning entries and sort them by time to obtain the early warning sequence and output it.

9. A photovoltaic inverter fault detection system based on multi-source data, characterized in that, include: The intrinsic modulus function acquisition module is used to acquire the original current signal and obtain the intrinsic modulus function sequence from the original current signal according to the preset extraction rules. The intrinsic modulus function screening and marking module is used to arrange the intrinsic modulus function sequence according to a preset arrangement rule. If the index in the intrinsic modulus function sequence is higher than the preset index threshold, it is marked as a risk intrinsic modulus function. The risk intrinsic modulus function decomposition module is used to decompose the risk intrinsic modulus function using a preset decomposition method to obtain the fine sub-band frequency component coefficients. The energy extraction fluctuation marking module is used to extract energy distribution features from the frequency component coefficients of the fine sub-band. If the local energy deviation value of the energy distribution feature exceeds a preset quantization threshold, the time period corresponding to the energy distribution feature is marked as a current fluctuation segment. The fluctuation segment demodulation module is used to perform envelope demodulation on the current fluctuation segment to obtain the instantaneous amplitude and instantaneous frequency; An abnormal energy calculation trend module is used to calculate local abnormal energy values ​​using the instantaneous amplitude and the instantaneous frequency, thereby forming an aging trend curve; The fault sequence early warning module is used to generate a fault evolution sequence based on the current fluctuation range and the aging trend curve, identify critical mutation points in the fault evolution sequence, obtain an early warning sequence based on the critical mutation points, and output it.

Images