A real-time fault detection method for photovoltaic inverters

Abstract

The application discloses a kind of photovoltaic inverter fault real-time detection methods, comprising the following steps: the internal operation signal of photovoltaic inverter is collected, fault feature extraction is carried out, and the mode matching of fault feature and pre-stored historical fault database is carried out, to generate fault diagnosis result;Real-time acquisition photovoltaic inverter grid-connected point's power grid parameter, according to the preset disturbance threshold value, the power grid parameter is analyzed, the type of grid disturbance event is identified, and according to grid disturbance event, power grid influence factor is calculated and generated;Synchronous reception fault diagnosis result and power grid influence factor, by inquiring the preset associated corresponding relationship, the theoretical influence value corresponding to current power grid disturbance event is obtained, and the initial confidence in the fault diagnosis result is corrected based on the theoretical influence value;By analyzing the theoretical influence value of grid disturbance to fault feature, and according to this, initial confidence is weightedly corrected, distinguish real fault from grid disturbance, significantly reduce false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method for real-time fault detection of photovoltaic inverters. Background Technology

[0002] As photovoltaic power generation accounts for an increasingly larger share of the energy mix, the operational reliability and fault diagnosis capabilities of photovoltaic inverters have become crucial. Existing fault diagnosis methods typically determine whether a fault has occurred by monitoring signals such as voltage, current, and temperature inside the inverter and comparing them with preset thresholds or historical fault modes.

[0003] However, in actual grid-connected operation, the power grid is not ideally stable. Grid disturbances such as voltage sags, frequency shifts, and harmonic distortion occur frequently. These external disturbances directly affect the operating status of photovoltaic inverters, causing their internal signals to exhibit transient characteristics very similar to real faults. For example, a severe voltage sag may lead to a sharp increase in switching stress on power devices, producing voltage overshoot and current oscillations similar to device aging.

[0004] Existing fault diagnosis systems often struggle to distinguish between false positive fault characteristics caused by external grid disturbances and those resulting from genuine inverter degradation, leading to a persistently high false alarm rate. Furthermore, frequent false alarms not only increase unnecessary maintenance costs but also erode maintenance personnel's trust in the diagnostic system, delaying the timely handling of genuine faults.

[0005] Therefore, how to decouple the impact of power grid disturbances and achieve accurate and reliable fault diagnosis is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This invention effectively suppresses false alarms caused by power grid events by introducing a power grid disturbance analysis channel and using a pre-built correlation model to correct the main diagnostic results.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a real-time fault detection method for photovoltaic inverters, comprising the following steps: Step S100, collecting internal operating signals of the photovoltaic inverter, extracting fault features, and performing pattern matching between the fault features and a pre-stored historical fault database to generate a fault diagnosis result containing fault type and initial confidence level; Step S200, collecting grid parameters at the grid connection point of the photovoltaic inverter in real time, analyzing the grid parameters according to a preset disturbance threshold, identifying the type of grid disturbance event, and calculating and generating a grid impact factor based on the duration and severity of the grid disturbance event; Step S300, synchronously receiving the fault diagnosis result and the grid impact factor, obtaining the theoretical impact value corresponding to the current grid disturbance event by querying a preset correlation relationship, and correcting the initial confidence level in the fault diagnosis result based on the theoretical impact value.

[0008] Preferably, step S100 specifically includes the following sub-steps: Step S110, simultaneously acquiring multiple internal operating signals through a sensor array integrated inside the photovoltaic inverter; wherein, the internal operating signals include: instantaneous values ​​of DC bus voltage and three-phase current on the AC output side acquired by Hall sensors, estimated junction temperature values ​​of power semiconductor devices obtained by infrared temperature measurement modules, and transient waveforms of gate-emitter voltage and collector-emitter voltage of power devices captured by high-speed differential probes during switching; Step S120, preprocessing the internal operating signals; wherein, the preprocessing includes noise reduction, Filtering and baseline correction; Step S130: Extract fault features using time-domain analysis, frequency-domain analysis, and wavelet transform methods; wherein, the fault features include: voltage overshoot peak value of the switching device, current tailing time, switching oscillation frequency and damping factor, energy loss value of a single switching process, and total harmonic distortion of the output current harmonics; Step S140: Calculate the similarity between the extracted fault features and the feature vectors stored in the historical fault database and labeled by confirmed fault cases using Euclidean distance, and generate the highest similarity value corresponding to the feature vector as the initial confidence level to form the fault diagnosis result.

[0009] Preferably, step S200 specifically includes the following sub-steps: Step S210, the grid parameters of the photovoltaic inverter grid connection point are collected in real time and continuously through the monitoring module; wherein, the grid parameters include the instantaneous and effective values ​​of three-phase voltage and three-phase current, grid frequency, voltage imbalance, total harmonic distortion and harmonic content; Step S220, the collected grid parameters are compared with a preset disturbance threshold; Step S230, when any parameter exceeds its corresponding normal fluctuation range and continues to exceed a preset time window, a grid disturbance event is determined to have occurred; wherein, the grid disturbance event specifically includes: voltage sag, voltage swell, short interruption, voltage transient, frequency shift, voltage fluctuation and flicker, and harmonic distortion or interharmonic occurrence; Step S240, the grid impact factor is calculated and generated based on the amplitude deviation percentage, duration and rate of change corresponding to the grid disturbance event; wherein, the grid impact factor is normalized.

[0010] Preferably, step S300 specifically includes the following sub-steps: Step S310, inputting the type of the grid disturbance event, the grid impact factor, and the fault type included in the fault diagnosis result as joint query conditions into a pre-stored association correspondence for processing; Step S320, the association correspondence outputs one or more theoretical impact values ​​according to the input query conditions; wherein, the theoretical impact value is used to characterize the magnitude or probability of the expected characteristic change caused by the grid disturbance event in a healthy photovoltaic inverter, corresponding to the fault characteristics detected in the current fault diagnosis result; Step S330, using the theoretical impact value, the initial confidence level in the fault diagnosis result is weighted and corrected using a weighting strategy; the weighting strategy is configured to proportionally reduce the initial confidence level according to the magnitude of the theoretical impact value.

[0011] Preferably, the construction of the correlation relationship includes the following steps: In a laboratory environment, using a programmable AC power supply and a grid simulator, different types and intensities of grid disturbance events are simulated on a photovoltaic inverter in a healthy state; during and before and after each simulated disturbance event, transient response data of the internal operating signals of the photovoltaic inverter are synchronously recorded through a data acquisition system; wherein, the transient response data includes the switching waveforms and output electrical characteristics of the power devices within the photovoltaic inverter; the characteristics of each simulated grid disturbance event are used as input labels, and the changes in the transient response data corresponding to the internal operating signals of the photovoltaic inverter are used as output labels to generate a large number of training data pairs; multiple linear regression is used to learn the training data pairs to establish the correlation relationship.

[0012] Preferably, the weighted correction of the initial confidence level in step S330 follows the following multi-level rules: If the power grid impact factor is lower than a preset low threshold, the power grid disturbance is determined to be minor, the initial confidence level is not corrected, and the fault diagnosis result is directly output; if the power grid impact factor is between the preset low threshold and the preset medium threshold, the power grid disturbance is determined to cause some feature overlap, and the initial confidence level is reduced by a preset ratio; if the power grid impact factor is between the preset medium threshold and the preset high threshold, the power grid disturbance is determined to cause misjudgment, the initial confidence level is reduced to an extremely low safety value, and marked as "requires manual review"; if the power grid impact factor is higher than the preset high threshold, all current fault features are determined to be caused by power grid disturbance, the initial confidence level is directly set to zero, and the fault alarm is completely suppressed.

[0013] Preferably, the system further includes: when a final fault diagnosis is confirmed as a false alarm through manual inspection or equipment disassembly, the system automatically triggers a data recording mechanism to archive the complete data packet before and after the false alarm; wherein, the data packet includes the characteristics of the power grid disturbance event at that time, the original waveform of the internal operating signal of the photovoltaic inverter, and the generated fault characteristics and fault diagnosis results; after the false alarm case data packets accumulate to a preset number, the offline retraining process is started, and the data in the false alarm case data packets is used as new training samples to incrementally learn the correlation and correspondence relationship.

[0014] Preferably, the construction and updating of the historical fault database includes the following steps: by conducting accelerated aging experiments and fault injection tests on photovoltaic inverters in a laboratory environment, obtaining full life cycle evolution data of typical fault modes, and performing data migration and pattern learning to form the initial core content of the database; after standardizing, extracting features, and timestamping real fault cases and the types of grid disturbances at the time of their occurrence, the time-series data of fault evolution, and the fault diagnosis results, they are used as new learning samples and archived into the historical fault database.

[0015] Preferably, before step S100, a self-calibration and reference establishment step is included: by controlling the switching devices inside the inverter, a known, low-amplitude electrical excitation signal is injected into each sensing channel; the response data of each sensor channel to the electrical excitation signal is collected, and the collected response baseline is compared point by point with the reference response of the photovoltaic inverter; the error, offset, and phase delay of each sensor channel are calculated and generated through Kalman filtering; and the error, offset, and phase delay of each sensor channel are compensated.

[0016] Preferably, the power grid disturbance events are predefined based on their electrical characteristics and physical causes, and their specific classifications include: voltage amplitude change events, specifically rapid changes in the effective voltage value occurring within a power frequency cycle or several seconds, including voltage dips, voltage swells, and short-term or instantaneous interruptions; frequency change events, specifically steady-state or transient deviations of the power grid frequency from its nominal value, including frequency deviations and frequency fluctuations; and waveform distortion events, specifically the degree to which the voltage or current waveform deviates from an ideal sine wave, including harmonic distortion, interharmonic occurrence, and high-frequency conducted disturbances such as voltage gaps and transient overvoltages.

[0017] The beneficial effects of this invention: The real-time fault detection method for photovoltaic inverters provided by this invention lies in the "event decoupling" mechanism. By introducing an independent grid disturbance analysis channel, this invention can identify and quantify the impact of external grid events on the inverter in real time, and quantitatively analyze the degree of impact of grid disturbances on specific fault characteristics. Thus, this invention can achieve intelligent and refined correction of diagnostic results, retain the alarm capability of real faults, and enable the entire diagnostic system to adapt to the constantly changing grid environment and equipment aging characteristics, thus possessing a self-evolutionary function. Attached Figure Description

[0018] Figure 1 The flowchart illustrates the steps of a real-time fault detection method for a photovoltaic inverter, as provided in one embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Example 1: Reference Figure 1 The diagnostic process designed in this invention is a dual-channel parallel and final fusion decision process, which is executed through the following steps: Step S100: Execute the main diagnostic process, collect the internal operating signals of the photovoltaic inverter, extract fault features, and perform pattern matching between the fault features and the pre-stored historical fault database to generate a fault diagnosis result containing the fault type and initial confidence level; this process is like an "internal physician" that focuses on the inverter itself.

[0021] Step S110: Simultaneously collect various internal operating signals through a sensor array integrated inside the photovoltaic inverter; among which, the internal operating signals include: the instantaneous values ​​of DC bus voltage and three-phase current on the AC output side collected by Hall sensors, the estimated junction temperature of power semiconductor devices obtained by infrared temperature measurement modules, and the transient waveforms of gate-emitter voltage and collector-emitter voltage of power devices captured by high-speed differential probes during the switching process.

[0022] Step S120: Preprocess the internal operating signal; wherein, the preprocessing includes noise reduction, filtering and baseline correction.

[0023] Step S130: The system extracts features from the processed signal. For example, it extracts time-domain features such as voltage overshoot and current tailing, as well as frequency-domain features such as switching oscillation frequency, from the switching waveform of IGBT (Insulated Gate Bipolar Transistor).

[0024] Step S140: The system calculates the similarity between the extracted current fault feature vector and the feature vectors labeled by confirmed fault cases stored in the historical fault database. For example, if the calculation shows that the current feature has the highest similarity (88%) with the "IGBT gate aging" case in the database, then the fault diagnosis result is generated: {Fault type: IGBT gate aging, initial confidence level: 88%}.

[0025] Step S200: Real-time acquisition of grid parameters at the photovoltaic inverter grid connection point, analysis of the grid parameters based on preset disturbance thresholds, identification of the type of grid disturbance event, and calculation and generation of grid impact factor based on the duration and severity of the grid disturbance event. This process is like an "environmental monitor" and is executed in parallel with S100, focusing on the external power grid.

[0026] Step S210: The monitoring module continuously and in real time collects grid parameters at the photovoltaic inverter's grid connection point. These grid parameters include instantaneous and effective values ​​of three-phase voltage and current, grid frequency, voltage imbalance, total harmonic distortion (THD), and the content of each harmonic. Step S220: The collected grid parameters are compared with a preset disturbance threshold. Step S230: When any parameter exceeds its corresponding normal fluctuation range and persists for more than a preset time window, a grid disturbance event is determined to have occurred. Specifically, the grid disturbance event includes: voltage sag, voltage swell, short-term interruption, voltage transient, frequency shift, voltage fluctuation and flicker, and harmonic distortion or interharmonic occurrence. Specifically, this invention uses a preset power quality monitoring module at the grid connection point to monitor parameters such as three-phase voltage, current, frequency, and THD in real time. For example, while this invention is performing step S100 for diagnosis, a voltage sag event occurs in the grid, with its amplitude dropping by 20% and a duration of 150ms. After comparing this parameter with a preset threshold, the system identifies a "voltage sag" event. Step S240: The system calculates a normalized grid impact factor based on parameters such as the magnitude (20%) and duration (150ms) of this event. Assuming the calculated factor is 0.75 (a value close to 1 indicates a severe impact), this result is recorded as {Event Type: Voltage Sag, Grid Impact Factor: 0.75}.

[0027] Step S300: Simultaneously receive fault diagnosis results and power grid impact factors, obtain the theoretical impact value corresponding to the current power grid disturbance event by querying the preset correlation correspondence, and correct the initial confidence level in the fault diagnosis results based on the theoretical impact value. This process is like a "chief reviewer" who is responsible for synthesizing the first two reports and making a final judgment.

[0028] Step S310: The type of the power grid disturbance event and the power grid impact factor, as well as the fault type contained in the fault diagnosis result, are used as joint query conditions and input into the pre-stored correlation relationship for processing; that is, the "chief review expert" receives two reports: {IGBT gate aging, 88%} and {voltage sag, 0.75}.

[0029] Step S320: The "lead expert" inputs this information as query conditions into the pre-stored "correspondence relationship". The "correspondence relationship" outputs a theoretical impact value through the query; for example, in this embodiment, the value is 0.8. This value indicates that for a healthy photovoltaic inverter, a single "voltage sag" event has a very high probability (80%) of causing a transient response highly similar to the currently detected "IGBT gate aging" characteristic.

[0030] Step S330: The "lead expert" uses the theoretical impact value of 0.8 to perform a weighted correction on the initial confidence level of 88%. According to the preset weighting strategy, due to the high theoretical impact value, the initial confidence level is significantly reduced. For example, the corrected final confidence level may drop to 15%. The system generates the final fault diagnosis conclusion: {Fault type: IGBT gate aging, final confidence level: 15%}. Due to the low confidence level, the system marks this event as "suspected grid disturbance," and does not trigger an alarm, thus successfully avoiding a false alarm. For example, the preset thresholds are set as follows: low threshold: 0.3 (corresponding to minor disturbances, such as voltage sag <10%), medium threshold: 0.6 (corresponding to medium disturbances, such as voltage sag 10%~30%), and high threshold: 0.9 (corresponding to severe disturbances, such as voltage sag >30% or frequency offset >1Hz).

[0031] Example 2: Specifically, this example illustrates how the "association correspondence" is established.

[0032] In the laboratory, a brand-new, healthy photovoltaic inverter is connected to a grid simulator. First, by controlling the grid simulator, a slight voltage dip (e.g., a 10% drop) is simulated, while simultaneously acquiring transient response data for all key signals within the inverter at high speed. Then, a moderate voltage dip (e.g., a 30% drop) is simulated, and response data is acquired again. This process is repeated to simulate different types (frequency shifts, harmonic injections, etc.) and intensities of grid disturbances.

[0033] Each simulation takes the quantitative characteristics of the grid disturbance event (such as event type and voltage drop percentage) as input and the changes in the inverter's internal signals (such as the amount of volts added by switching overshoot) as output, forming an "input-output" data pair. By accumulating a large number of data pairs, multiple linear regression is used for training, ultimately establishing a correlation model that can accurately map the complex nonlinear relationship between "grid disturbance and inverter response," i.e., the "correlation correspondence." Specifically, when training the multiple linear regression model, the input variables include: grid disturbance event type (such as voltage sag, frequency shift, etc.), disturbance intensity (such as amplitude deviation percentage, duration), and fault type (such as IGBT gate aging); the output variable is the theoretical impact value, characterizing the degree of influence of the disturbance on the fault characteristics. After the model training is complete, a theoretical impact value between 0 and 1 can be output for confidence correction.

[0034] Example 3: When a final fault diagnosis is confirmed as a false alarm through manual inspection or equipment disassembly, the system automatically triggers a data recording mechanism to archive the complete data packet before and after the false alarm. This data packet includes records of the grid disturbance event characteristics at the time, the original waveforms of the photovoltaic inverter's internal operating signals, and the generated fault characteristics and fault diagnosis results. A "false alarm" is defined as follows: Suppose during a field operation, the system issues a low-confidence alarm. After on-site inspection, maintenance personnel confirm that the inverter is not faulty, indicating a false alarm. The maintenance personnel confirm this event as a "false alarm" through the human-machine interface. The system automatically triggers the recording mechanism to archive the complete data packet before and after the false alarm (including records of the grid disturbance at the time, the inverter's internal signal waveforms, and the preliminary diagnosis results generated by the system). After accumulating a certain number of false alarm cases, the system can initiate an offline retraining process in the background, using this real-world data as new training samples to incrementally learn or retrain the original correlation model. In this way, the model can continuously learn new and unforeseen grid-equipment interaction patterns, thereby continuously improving accuracy.

[0035] Example 4: By controlling the switching devices inside the inverter, a known, low-amplitude electrical excitation signal is injected into each sensor channel; the response data of each sensor channel to the electrical excitation signal is collected, and the collected response baseline is compared point by point with the reference response of the photovoltaic inverter. Through Kalman filtering, the error, offset, and phase delay of each sensor channel are calculated; wherein, the reference response is collected and stored under standard test environment before leaving the factory, and serves as the comparison benchmark for subsequent self-calibration; the error, offset, and phase delay of each sensor channel are compensated.

[0036] Example 5: Power grid disturbance events are predefined based on their electrical characteristics and physical causes. Specific classifications include: voltage amplitude change events, specifically rapid changes in the effective voltage value occurring within a power frequency cycle or several seconds, including voltage dips, voltage swells, and short-term or instantaneous interruptions; frequency change events, specifically steady-state or transient deviations of the power grid frequency from its nominal value, including frequency deviations and frequency fluctuations; and waveform distortion events, specifically the degree to which voltage or current waveforms deviate from an ideal sine wave, including harmonic distortion, interharmonic occurrence, and high-frequency conducted disturbances such as voltage gaps and transient overvoltages.

[0037] In summary, this invention decouples the impact of grid disturbances from inverter fault diagnosis through a dual-channel parallel diagnostic architecture, achieving accurate and reliable diagnosis in complex grid environments.

[0038] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0039] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A photovoltaic inverter fault real-time detection method, characterized in that, The method comprises the following steps: Step S100, collecting internal operation signals of the photovoltaic inverter, extracting fault features, and performing pattern matching on the fault features and a pre-stored historical fault database to generate a fault diagnosis result containing a fault type and an initial confidence level; wherein the internal operation signals include DC bus voltage and three-phase current instantaneous values on the AC output side collected by a Hall sensor, junction temperature estimated values of power semiconductor devices obtained by an infrared temperature measurement module, and transient waveforms of gate-emitter voltage and collector-emitter voltage of the power devices in the switching process captured by a high-speed differential probe; Step S200, collecting grid parameters of a grid-connected point of the photovoltaic inverter in real time, analyzing the grid parameters according to a pre-set disturbance threshold, identifying the type of a grid disturbance event, and calculating and generating a grid influence factor according to the duration and severity of the grid disturbance event; Step S300, synchronously receiving the fault diagnosis result and the grid influence factor, obtaining a theoretical influence value corresponding to the current grid disturbance event by querying a pre-set associated correspondence; wherein the theoretical influence value is used to represent the amplitude or probability of an expected characteristic change corresponding to the detected fault feature in the current fault diagnosis result caused by the grid disturbance event in a healthy photovoltaic inverter; the initial confidence level in the fault diagnosis result is weighted and corrected by a weighting strategy using the theoretical influence value; the weighting strategy is configured to proportionally reduce the initial confidence level according to the size of the theoretical influence value.

2. The photovoltaic inverter fault real-time detection method according to claim 1, characterized in that, The step S100 specifically comprises the following sub-steps: Step S110, synchronously collecting multiple internal operation signals by a sensor array integrated in the photovoltaic inverter; Step S120, pre-processing the internal operation signals; wherein the pre-processing includes denoising, filtering and baseline correction; Step S130, extracting fault features by time domain analysis, frequency domain analysis and wavelet transform method; wherein the fault features include voltage overshoot peak value of the switching device, current tail time, switching oscillation frequency and damping factor, energy loss value of a single switching process, and total harmonic distortion of output current harmonics; Step S140, calculating the similarity of the extracted fault features and the feature vectors stored in the historical fault database and calibrated by confirmed fault cases by Euclidean distance, and generating the highest similarity value corresponding to the feature vectors as the initial confidence level to form the fault diagnosis result.

3. The photovoltaic inverter fault real-time detection method according to claim 1, characterized in that, The step S200 specifically comprises the following sub-steps: step S210, acquiring, by the monitoring module, the grid parameters of the grid-connected point of the photovoltaic inverter in real time and continuously; wherein the grid parameters comprise instantaneous values and effective values of three-phase voltage and three-phase current, grid frequency, voltage unbalance degree, and total harmonic distortion and harmonic content; step S220, comparing the acquired grid parameters with preset disturbance threshold values; step S230, when any parameter exceeds its corresponding normal fluctuation range and lasts for more than a preset time window, determining that a grid disturbance event occurs; wherein the grid disturbance event specifically comprises voltage sag, voltage swell, short-time interruption, voltage transient, frequency deviation, voltage fluctuation and flicker, and harmonic distortion or inter-harmonic appearance; step S240, calculating a grid influence factor according to the amplitude deviation percentage, duration, and change rate corresponding to the grid disturbance event; wherein the grid influence factor is normalized.

4. The photovoltaic inverter fault real-time detection method according to claim 1, characterized in that, The step S300 specifically comprises the following sub-steps: step S310, inputting the type of the grid disturbance event and the grid influence factor, and the fault type contained in the fault diagnosis result, as a joint query condition into a pre-stored association correspondence for processing; step S320, outputting one or more theoretical influence values by the association correspondence according to the input query condition; step S330, using the theoretical influence values to weight and correct the initial confidence in the fault diagnosis result by a weighting strategy; the weighting strategy is configured to proportionally reduce the initial confidence according to the size of the theoretical influence values.

5. The photovoltaic inverter fault real-time detection method according to claim 4, characterized in that, The construction of the association correspondence comprises the following steps: in a laboratory environment, simulating different types and different intensities of grid disturbance events on a photovoltaic inverter in a healthy state by a programmable AC power supply and a grid simulator; synchronously recording transient response data of internal operating signals of the photovoltaic inverter during and before and after each simulated disturbance event by a data acquisition system; wherein the transient response data comprises switching waveforms of power devices in the photovoltaic inverter and output electrical characteristics; taking the characteristics of each simulated grid disturbance event as input labels and taking the change amount of the transient response data corresponding to the internal operating signals of the photovoltaic inverter as output labels to generate a large number of training data pairs; using multivariate linear regression to learn the training data pairs to establish the association correspondence.

6. The photovoltaic inverter fault real-time detection method according to claim 4, characterized in that, The weighting correction of the initial confidence in the step S330 follows the following multi-level rules: if the grid influence factor is lower than a preset low threshold value, it is determined that the grid disturbance has a slight impact, the initial confidence is not corrected, and the fault diagnosis result is directly outputted; if the grid influence factor is between the preset low threshold value and a preset medium threshold value, it is determined that the grid disturbance causes partial feature overlap, and the initial confidence is proportionally reduced; If the grid impact factor is between the preset medium threshold and the preset high threshold, it is determined that the grid disturbance will cause misjudgment, the initial confidence is reduced to a very low safety value, and is marked as "manual review"; if the grid impact factor is higher than the preset high threshold, it is determined that all current fault features are caused by grid disturbance, the initial confidence is directly set to zero, and the fault alarm is completely suppressed.

7. The photovoltaic inverter fault real-time detection method according to claim 1, characterized in that, Also comprising: When a final fault diagnosis conclusion is confirmed as a false alarm through manual inspection or equipment disassembly, the system automatically triggers a data recording mechanism to archive the complete data packet before and after the false alarm; wherein the data packet includes the grid disturbance event feature record at that time, the original waveform of the internal operation signal of the photovoltaic inverter, and the generated fault feature and fault diagnosis result; after the false alarm case data packet accumulates to a preset number, an offline retraining process is started, and the data in the false alarm case data packet is used as a new training sample to perform incremental learning on the association correspondence.

8. The photovoltaic inverter fault real-time detection method according to claim 1, characterized in that, The construction and updating of the historical fault database includes the following steps: through accelerated aging experiments and fault injection tests on photovoltaic inverters in a laboratory environment, full life cycle evolution data of typical fault modes are obtained, and data migration and mode learning are performed to form the initial core content of the database; after standardization processing, feature extraction and timestamp marking, real fault cases, the grid disturbance type at the time of occurrence, the time sequence data of fault evolution and the fault diagnosis result are archived into the historical fault database as new learning samples.

9. The photovoltaic inverter fault real-time detection method according to claim 1, characterized in that, Before step S100, a self-calibration and reference establishment step is further included: a known, low-amplitude electrical excitation signal is injected into each sensing channel by controlling the internal switching device of the inverter; the response data of each sensor channel to the electrical excitation signal is collected, and the response baseline collected this time is compared with the reference response of the photovoltaic inverter point by point, and the error, offset and phase delay of each sensor channel are calculated and generated through Kalman filtering; the error, offset and phase delay of each sensor channel are compensated.

10. The photovoltaic inverter fault real-time detection method according to claim 1, characterized in that, The grid disturbance event is predefined according to its electrical characteristics and physical causes, and its specific classification includes: voltage amplitude change event, specifically the rapid change of voltage effective value within a power frequency cycle or a few seconds, including voltage sag, voltage swell and short-time or instantaneous interruption; frequency change event, specifically the steady-state or transient-state deviation of grid frequency from its nominal value, including frequency deviation and frequency fluctuation; waveform distortion event, specifically the degree of deviation of voltage or current waveform from ideal sinusoidal wave, including harmonic distortion, interharmonic appearance, voltage notch and transient overvoltage and other high-frequency conducted disturbance.

Images