Fault diagnosis method and system for three-level npc grid-connected inverter based on disturbance subspace projection and excitation path excitable constraint

Abstract

The application discloses a three-level NPC grid-connected inverter fault diagnosis method and system based on disturbance subspace projection and conduction path excitation constraint, a diagnosis model under a synchronous rotating coordinate system is first constructed, and original residual errors between health condition estimated values and measured outputs are obtained through a state observer; a health disturbance subspace is constructed, and the original residual errors are projected to the orthogonal complement to obtain de-disturbance projection residual errors; the conduction path excitation of each fault is determined according to a current switch implementation state, a phase current direction, a midpoint balance modulation state and a midpoint voltage state, and a fault visibility mask is generated; fault matching score and posterior probability calculation is only performed in a current visible fault set; a rejection judgment is performed according to the maximum posterior probability, uncertainty and a posterior interval between the first two, and a fault category and a fault position or a rejection result are output. The application can effectively inhibit health residual error pollution caused by condition changes, and reduce misjudgment of a current physically non-excitable fault.

Description

Technical Field

[0001] This invention relates to the field of power electronic converter fault diagnosis technology, and in particular to a fault diagnosis method and system for three-level NPC grid-connected inverters based on disturbance subspace projection and conduction path excitation constraints. Background Technology

[0002] With the large-scale integration of photovoltaic power generation, energy storage systems, and high-performance power conversion equipment, three-level NPC grid-connected inverters have become one of the important topologies for new energy grid connection interfaces due to their advantages such as lower voltage stress on switching devices, better output voltage waveform quality, and suitability for medium- and high-voltage scenarios. However, the large number of power switching devices, strong conduction path redundancy, and significant midpoint potential coupling in three-level NPC inverters mean that under the combined effects of long-term high-frequency switching, thermal cycling stress, and grid disturbances, power switching devices, current sensors, and gate drive units may malfunction, potentially leading to inverter shutdown or even grid safety accidents. Therefore, achieving highly accurate and robust online fault diagnosis for three-level NPC grid-connected inverters under complex operating conditions has become a key technical problem that urgently needs to be solved in this field.

[0003] In existing technologies, fault diagnosis methods for three-level NPC grid-connected inverters can be broadly categorized into two main types: signal feature-based methods and system model-based methods. Signal feature-based methods typically utilize current harmonic analysis, Park vector trajectories, spectral distortion characteristics, or simple threshold criteria to identify faults; their principles are intuitive and their implementation is simple. System model-based methods, on the other hand, reconstruct the healthy reference output by constructing the dynamic equations of the inverter system and an observer, using the difference between the measured output and the reference output as the basis for fault diagnosis, exhibiting strong physical interpretability. Furthermore, some research has attempted to introduce intelligent reasoning methods such as Bayesian networks, DS evidence theory, or neural networks into the fault classification process, improving diagnostic accuracy through probabilistic reasoning or multi-source information fusion.

[0004] However, existing technologies still have the following shortcomings. Signal characteristic-based methods are easily affected by changes in operating conditions. When changes in operating conditions occur, such as photovoltaic power fluctuations, grid impedance changes, phase-locked loop transients, or midpoint balancing control actions, even if the system is in a healthy state, current and voltage measurement signals may exhibit waveform changes similar to fault signals, leading to a significant increase in the false alarm rate. Model-based methods typically assume that health disturbances such as changes in photovoltaic power setpoint, phase-locked loop angle and frequency deviations, grid impedance drift, and midpoint balancing modulation actions are uniformly treated as Gaussian noise terms, without explicitly modeling and separating the systematic impact of these disturbances on the observer residuals. Furthermore, three-level NPC inverters have redundant zero vectors and multiple internal conduction paths. Under different switching implementation states, phase current directions, and midpoint balancing strategies, the impact of the same type of fault on external measurable quantities varies, and may even be completely invisible at certain sampling moments. If the diagnostic tool does not utilize conduction path constraints and directly compares within the entire fault space, it may give a high confidence level judgment for faults that are currently physically unexciteable. Therefore, it is necessary to design a fault diagnosis method and system for three-level NPC grid-connected inverters based on perturbation subspace projection and conduction path excitation constraints to solve the above problems. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a fault diagnosis method and system for three-level NPC grid-connected inverters based on disturbance subspace projection and conduction path excitation constraints. It aims to overcome the problems in existing three-level NPC grid-connected inverter fault diagnosis methods, such as residual contamination caused by failure to explicitly separate healthy disturbance components, misjudgment of invisible faults due to failure to utilize conduction path constraints, and unreliable forced hard decision output due to lack of rejection mechanism. It features the characteristic of organically integrating orthogonal projection of healthy disturbance subspace with conduction path excitation constraints and multi-condition rejection mechanism to improve the robustness and reliability of online diagnosis.

[0006] To achieve the above technical solution, the technical solution adopted by the present invention is as follows: A fault diagnosis method for three-level NPC grid-connected inverters based on perturbation subspace projection and conduction path excitation constraints includes the following steps: Obtain the operational information set at the current sampling moment. The operational information set includes at least the grid-side current, split DC bus voltage, phase-locked loop variables, control quantities, and the current switch implementation status. A diagnostic model is constructed in a synchronous rotating coordinate system based on the operational information set. The diagnostic model includes at least the state variables of the LCL filter network, the state variables of the midpoint voltage, the operational condition disturbances, and the faults. Based on the diagnostic model and known control variables, the state estimate and estimated output value consistent with the health condition are obtained through the state observer, and the original residual is obtained based on the measured output and the estimated output. Based on the local sensitivity matrix of the original residual to the operating condition disturbance under the current operating point, a healthy disturbance subspace is constructed, and the original residual is projected onto the orthogonal complement of the healthy disturbance subspace to obtain the de-disturbed projected residual. Based on the current switch implementation state, phase current direction, midpoint balance modulation state, and midpoint voltage state, determine the excitation potential of each fault hypothesis's conduction path at the current moment and generate a fault visibility mask. For each fault hypothesis in the current set of visible faults, calculate its matching score with the descrambling projection residual, and perform posterior normalization only within the current set of visible faults to obtain the fault posterior probability. The rejection decision is made based on at least the maximum posterior probability, the uncertainty, and the interval between the top two posterior probabilities. Output the fault category and fault location when no rejection is triggered, and output the rejection result when rejection is triggered.

[0007] Preferably, the operating information set further includes photovoltaic power input, phase-locked loop angle, phase-locked loop frequency, and midpoint balance modulation related quantities; the split DC bus voltage is used to calculate the midpoint voltage deviation. ; in, The voltage across the upper capacitor. This is the voltage across the lower capacitor.

[0008] Preferably, the state vector of the diagnostic model includes at least the inverter-side d-axis current, inverter-side q-axis current, filter capacitor d-axis voltage, filter capacitor q-axis voltage, grid-side d-axis current, grid-side q-axis current, and midpoint voltage deviation.

[0009] Preferably, the operating condition disturbance includes at least one or more of the following: photovoltaic power given change, phase-locked loop angle change, phase-locked loop frequency change, grid impedance change, and midpoint balance modulation change.

[0010] Preferably, the faults include open-circuit faults of switching devices, current sensor faults, and gate abnormality faults, wherein open-circuit faults of switching devices and gate abnormality faults affect the status channel, and current sensor faults affect the measurement channel.

[0011] Preferably, the gate abnormality fault is a gate pulse failure fault, a gate jamming fault, or an equivalent open circuit fault caused by gate drive failure.

[0012] Preferably, the state observer includes a prediction phase and a correction phase; the prediction phase obtains a predicted state based on the diagnostic model, the known control variables, and the estimated values ​​of the operating condition disturbances; the correction phase corrects the predicted state based on the innovation between the measured output and the predicted output to obtain the state estimate and the estimated output value.

[0013] Furthermore, the estimated value of the operating condition disturbance... Obtained through one of the following methods: Will Set it to the zero vector, and then remove this part of the perturbation effect from the original residual by the subsequent health perturbation subspace projection step; or during the high-confidence health window, obtain the perturbation estimate at the current time by least squares estimation based on the historical residual sequence.

[0014] Preferably, projecting the original residual onto the orthogonal complement of the healthy perturbation subspace includes: ; ; ; in, This is the local sensitivity matrix of the original residual to operating condition disturbances. The residual weighting matrix, For regularization parameters, It is the identity matrix. For the original residual, For the estimated coordinates of the health disturbance, For the estimation components of the health disturbance residuals, To remove scrambling from the projected residuals.

[0015] Furthermore, the regularization parameters It is determined by the L-curve method, or taken as 0.01 to 0.1 times the ratio of the norm of the residual covariance matrix to the norm of the sensitivity matrix.

[0016] Furthermore, the local sensitivity matrix The estimated state is obtained analytically from the state already estimated by the state observer and the parameters of the current running point: the estimated state Substituting into the observer error dynamics equation, the steady-state response or one-step recursive response of the partial derivative terms of each health disturbance in the error dynamics equation is obtained respectively, thus forming... The column vectors of .

[0017] Preferably, the orthogonal complementary projection matrix of the health perturbation subspace is: ; The scrambling projection residual satisfies The healthy perturbation subspace and each fault subspace satisfy the condition of local separability. Or satisfy the minimum protagonist constraint ,in For health-related disturbance subspaces, For the first Fault-like subspace, The separable angle threshold.

[0018] Preferably, the fault visibility mask is determined jointly by the conduction path excitation function and the projected fault sensitivity, satisfying the following conditions: ; in, For the first Class of faults at the current moment Visibility mask, Indicates the first The result of determining the triggerability of this type of fault under the current conduction path. For the first The fault sensitivity vector after projecting the fault class onto the health perturbation subspace. This is the sensitivity threshold.

[0019] Furthermore, the sensitivity threshold Based on the health status of the first The statistical distribution of the fault sensitivity vector norm is determined, and a value above a certain quantile of the statistical distribution is selected to exclude non-zero sensitivity caused by noise from visibility determination.

[0020] Preferably, the conduction path excitation function is based on the current topology context variable. Confirmed, among which This represents the current state of the bridge arm. The symbol for the direction of three-phase current. For midpoint balanced modulation or zero vector allocation parameters, This represents the midpoint voltage deviation.

[0021] Furthermore, the conduction path can excite functions This is achieved through a lookup table logic: a truth table is pre-constructed based on combinations of bridge arm implementation status, phase current direction sign, zero vector allocation method, and midpoint balance modulation bias direction; and the truth table is then applied online based on the current topology context variables. Query No. Can this type of fault be triggered; or can it be calculated by the first... The local transmission sensitivity from a fault type to the output quantity is determined to be excitationable when the norm of this transmission sensitivity is greater than a preset threshold.

[0022] Preferably, the fault matching score satisfies ; in, For the first Class of faults at any time Match score, This is a fault visibility mask. To descramble the projected residuals, For the first The projected fault sensitivity vector of the fault class. The residual weighting matrix, This is for regularizing small quantities.

[0023] Preferably, the posterior normalization applies only to the currently visible set of faults. Internal execution, satisfying ; Let the health hypothesis be denoted as And together with each fault hypothesis in the current set of visible faults, they participate in posterior normalization. For the invisible fault hypothesis, we have: ;in For the first The posterior probability of a fault class For temperature parameters, This is for regularizing small quantities.

[0024] Furthermore, the temperature parameter The value is a positive real number, determined through cross-validation, or preset to a fixed value in the range of 1.0 to 5.0.

[0025] Preferably, the uncertainty is determined by the normalized entropy on the currently visible fault set, satisfying the following conditions: ; in, For a moment Uncertainty, The cardinality of the currently visible set of faults. To regularize small quantities, To regularize small quantities, Let be the posterior probability.

[0026] Preferably, rejection is triggered when any of the following conditions are met: the currently visible fault set is empty; the maximum posterior probability is lower than the first threshold; the uncertainty is higher than the second threshold; the posterior interval of the top two faults is lower than the third threshold.

[0027] Furthermore, the first threshold The value range is from 0.5 to 0.9, the second threshold. The value range is from 0.3 to 0.7, and the third threshold... The value range is from 0.1 to 0.3, and the specific value of the above threshold is determined through offline calibration based on the actual system's false alarm rate and false negative rate requirements.

[0028] Preferably, the method further includes an online update step: when no rejection is triggered, the maximum a posteriori corresponds to the health hypothesis, the health consistency score is higher than the health threshold, and the uncertainty is lower than the health update threshold, the local sensitivity matrix of the operating condition disturbance or its parameter dictionary is updated online.

[0029] Preferably, the high-confidence health window is determined by simultaneously satisfying the following conditions at multiple consecutive sampling times: health posterior dominance, health consistency score higher than a threshold, posterior uncertainty lower than a threshold, and no rejection trigger.

[0030] Preferably, a fault diagnosis system for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints is used to execute the aforementioned fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints. The system includes: The data acquisition module is used to obtain a set of operational information. The coordinate transformation and diagnostic model construction module has its input end connected to the output end of the data acquisition module, and is used to transform the running information into a synchronous rotating coordinate system and construct a diagnostic model. The state reconstruction module, whose input is connected to the output of the coordinate transformation and diagnostic model construction module, is used to obtain the state estimate and the original residual through the state observer. The perturbation projection module, whose input is connected to the output of the state reconstruction module, is used to construct a healthy perturbation subspace and project the original residual onto its orthogonal complement to obtain the de-perturbation projection residual. The conduction path can trigger the determination module to generate a fault visibility mask; The fault matching and posterior evaluation module has its first input terminal connected to the output terminal of the disturbance projection module to receive the descrambling projection residual, and its second input terminal connected to the output terminal of the conduction path excitation determination module to receive the fault visibility mask, which is used to calculate the fault matching score and posterior probability. The diagnostic output module, whose input is connected to the output of the fault matching and post-evaluation module, is used to perform rejection judgment and output diagnostic results. The online update module, whose input is connected to the output of the diagnostic output module, is used to update the perturbation sensitivity matrix within a high-confidence health window.

[0031] Furthermore, the data flow between the modules is as follows: the output of the data acquisition module is connected to the input of the coordinate transformation and diagnostic model construction module; the output of the coordinate transformation and diagnostic model construction module is connected to the inputs of the state reconstruction module and the disturbance projection module, respectively; the output of the state reconstruction module is connected to the input of the disturbance projection module to provide the original residual; the output of the disturbance projection module is connected to the input of the fault matching and posterior evaluation module to provide the de-scratched projection residual; the output of the conduction path excitation determination module is connected to the input of the fault matching and posterior evaluation module to provide the fault visibility mask; the output of the fault matching and posterior evaluation module is connected to the input of the diagnostic output module to provide the posterior probability; and the output of the diagnostic output module is connected to the input of the online update module to provide the health status confirmation signal.

[0032] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints.

[0033] Preferably, a computer-readable storage medium stores computer instructions for causing a computer to execute the described fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints.

[0034] The beneficial effects of this invention are as follows: 1. This invention constructs a healthy disturbance subspace based on the local sensitivity matrix of the original residual to the operating condition disturbance at the current operating point, and projects the original residual onto the orthogonal complement of this healthy disturbance subspace to obtain the de-disturbed projected residual. This achieves explicit stripping of the components in the original residual that can be explained by changes in healthy operating conditions, retaining only the unexplainable components caused by faults for subsequent diagnostic analysis. This solves the problem of existing model-based methods treating healthy disturbances uniformly as noise without differentiation, leading to residual contamination and fault misjudgment. It effectively improves the fault diagnosis robustness of inverters under operating condition changes such as photovoltaic power fluctuations, phase-locked loop transients, grid impedance changes, and midpoint balance control actions.

[0035] 2. This invention determines the excitability of the conduction path for each fault hypothesis at the current moment based on the current switch state, phase current direction, midpoint balance modulation state, and midpoint voltage state, and generates a fault visibility mask. Fault matching score calculation and posterior normalization are performed only within the current set of visible faults, and zero posterior probability is assigned to invisible fault hypotheses. This solves the problem that existing diagnostic methods do not utilize the conduction path constraints of three-level NPC inverters and directly compare in the entire fault space, resulting in high-confidence misjudgments of currently physically non-excitable faults. It effectively reduces the false alarm rate and improves the accuracy of fault location under multi-redundant path topologies.

[0036] 3. This invention solves the problem that existing diagnostic methods still force hard decisions when there is insufficient evidence, leading to unreliable erroneous label outputs, by performing rejection judgment based on the maximum posterior probability, uncertainty, and the interval between the top two posterior probabilities. Instead of forcibly outputting the fault category and fault location, it outputs the rejection result when the rejection condition is triggered. At the same time, by updating the local sensitivity matrix of the operating condition disturbance online within the high-confidence health window, the diagnostic model can adapt to the drift of slowly changing operating conditions, improving the engineering applicability and long-term deployment reliability of the online fault diagnosis system. Attached Figure Description

[0037] Figure 1 This is the overall flowchart in the embodiments of the present invention; Figure 2 This is a schematic diagram of the diagnostic model of the three-level NPC grid-connected inverter in the dq synchronous rotating coordinate system in an embodiment of the present invention; Figure 3 This is a schematic diagram of the health reference observer and the perturbation subspace projection module in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the determination of the excitation capability of the conduction path and the generation of the fault visibility mask in an embodiment of the present invention; Figure 5 This is a block diagram of fault matching, posterior normalization, uncertainty assessment, and rejection output in an embodiment of the present invention; Figure 6 This is a comparison waveform of the residuals before and after orthogonal projection of the health perturbation subspace in an embodiment of the present invention; Figure 7 This is a diagram showing the posterior probability of a fault and the rejection state under the excitation constraint of the conduction path in this embodiment of the invention. Figure 8 This is a structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0038] Example 1: A fault diagnosis method for three-level NPC grid-connected inverters based on perturbation subspace projection and conduction path excitation constraints includes the following steps: Obtain the operational information set at the current sampling moment. The operational information set includes at least the grid-side current, split DC bus voltage, phase-locked loop variables, control quantities, and the current switch implementation status. A diagnostic model is constructed in a synchronous rotating coordinate system based on the operational information set. The diagnostic model includes at least the state variables of the LCL filter network, the state variables of the midpoint voltage, the operational condition disturbances, and the faults. Based on the diagnostic model and known control variables, the state estimate and estimated output value consistent with the health condition are obtained through the state observer, and the original residual is obtained based on the measured output and the estimated output. Based on the local sensitivity matrix of the original residual to the operating condition disturbance under the current operating point, a healthy disturbance subspace is constructed, and the original residual is projected onto the orthogonal complement of the healthy disturbance subspace to obtain the de-disturbed projected residual. Based on the current switch implementation state, phase current direction, midpoint balance modulation state, and midpoint voltage state, determine the excitation potential of each fault hypothesis's conduction path at the current moment and generate a fault visibility mask. For each fault hypothesis in the current set of visible faults, calculate its matching score with the descrambling projection residual, and perform posterior normalization only within the current set of visible faults to obtain the fault posterior probability. The rejection decision is made based on at least the maximum posterior probability, the uncertainty, and the interval between the top two posterior probabilities. Output the fault category and fault location when no rejection is triggered, and output the rejection result when rejection is triggered.

[0039] Preferably, the operating information set further includes photovoltaic power input, phase-locked loop angle, phase-locked loop frequency, and midpoint balance modulation related quantities; the split DC bus voltage is used to calculate the midpoint voltage deviation. ; in, The voltage across the upper capacitor. This is the voltage across the lower capacitor.

[0040] Preferably, the state vector of the diagnostic model includes at least the inverter-side d-axis current, inverter-side q-axis current, filter capacitor d-axis voltage, filter capacitor q-axis voltage, grid-side d-axis current, grid-side q-axis current, and midpoint voltage deviation.

[0041] Preferably, the operating condition disturbance includes at least one or more of the following: photovoltaic power given change, phase-locked loop angle change, phase-locked loop frequency change, grid impedance change, and midpoint balance modulation change.

[0042] Preferably, the faults include open-circuit faults of switching devices, current sensor faults, and gate abnormality faults, wherein open-circuit faults of switching devices and gate abnormality faults affect the status channel, and current sensor faults affect the measurement channel.

[0043] Preferably, the gate abnormality fault is a gate pulse failure fault, a gate jamming fault, or an equivalent open circuit fault caused by gate drive failure.

[0044] Preferably, the state observer includes a prediction phase and a correction phase; the prediction phase obtains a predicted state based on the diagnostic model, the known control variables, and the estimated values ​​of the operating condition disturbances; the correction phase corrects the predicted state based on the innovation between the measured output and the predicted output to obtain the state estimate and the estimated output value.

[0045] Furthermore, the estimated value of the operating condition disturbance... Obtained through one of the following methods: Will Set it to the zero vector, and then remove this part of the perturbation effect from the original residual by the subsequent health perturbation subspace projection step; or during the high-confidence health window, obtain the perturbation estimate at the current time by least squares estimation based on the historical residual sequence.

[0046] Preferably, projecting the original residual onto the orthogonal complement of the healthy perturbation subspace includes: ; ; ; in, This is the local sensitivity matrix of the original residual to operating condition disturbances. The residual weighting matrix, For regularization parameters, It is the identity matrix. For the original residual, For the estimated coordinates of the health disturbance, For the estimation components of the health disturbance residuals, To remove scrambling from the projected residuals.

[0047] Furthermore, the regularization parameters It is determined by the L-curve method, or taken as 0.01 to 0.1 times the ratio of the norm of the residual covariance matrix to the norm of the sensitivity matrix.

[0048] Furthermore, the local sensitivity matrix The estimated state is obtained analytically from the state already estimated by the state observer and the parameters of the current running point: the estimated state Substituting into the observer error dynamics equation, the steady-state response or one-step recursive response of the partial derivative terms of each health disturbance in the error dynamics equation is obtained respectively, thus forming... The column vectors of .

[0049] Preferably, the orthogonal complementary projection matrix of the health perturbation subspace is: ; The scrambling projection residual satisfies The healthy perturbation subspace and each fault subspace satisfy the condition of local separability. Or satisfy the minimum protagonist constraint ,in For health-related disturbance subspaces, For the first Fault-like subspace, The separable angle threshold.

[0050] Preferably, the fault visibility mask is determined jointly by the conduction path excitation function and the projected fault sensitivity, satisfying the following conditions: ; in, For the first Class of faults at the current moment Visibility mask, Indicates the first The result of determining the triggerability of this type of fault under the current conduction path. For the first The fault sensitivity vector after projecting the fault class onto the health perturbation subspace. This is the sensitivity threshold.

[0051] Furthermore, the sensitivity threshold Based on the health status of the first The statistical distribution of the fault sensitivity vector norm is determined, and a value above a certain quantile of the statistical distribution is selected to exclude non-zero sensitivity caused by noise from visibility determination.

[0052] Preferably, the conduction path excitation function is based on the current topology context variable. Confirmed, among which This represents the current state of the bridge arm. The symbol for the direction of three-phase current. For midpoint balanced modulation or zero vector allocation parameters, This represents the midpoint voltage deviation.

[0053] Furthermore, the conduction path can excite functions This is achieved through a lookup table logic: a truth table is pre-constructed based on combinations of bridge arm implementation status, phase current direction sign, zero vector allocation method, and midpoint balance modulation bias direction; and the truth table is then applied online based on the current topology context variables. Query No. Can this type of fault be triggered; or can it be calculated by the first... The local transmission sensitivity from a fault type to the output quantity is determined to be excitationable when the norm of this transmission sensitivity is greater than a preset threshold.

[0054] Preferably, the fault matching score satisfies ; in, For the first Class of faults at any time Match score, This is a fault visibility mask. To descramble the projected residuals, For the first The projected fault sensitivity vector of the fault class. The residual weighting matrix, This is for regularizing small quantities.

[0055] Preferably, the posterior normalization applies only to the currently visible set of faults. Internal execution, satisfying ; Let the health hypothesis be denoted as And together with each fault hypothesis in the current set of visible faults, they participate in posterior normalization. For the invisible fault hypothesis, we have: ;in For the first The posterior probability of a fault class For temperature parameters, This is for regularizing small quantities.

[0056] Furthermore, the temperature parameter The value is a positive real number, determined through cross-validation, or preset to a fixed value in the range of 1.0 to 5.0.

[0057] Preferably, the uncertainty is determined by the normalized entropy on the currently visible fault set, satisfying the following conditions: ; in, For a moment Uncertainty, The cardinality of the currently visible set of faults. To regularize small quantities, To regularize small quantities, Let be the posterior probability.

[0058] Preferably, rejection is triggered when any of the following conditions are met: the currently visible fault set is empty; the maximum posterior probability is lower than the first threshold; the uncertainty is higher than the second threshold; the posterior interval of the top two faults is lower than the third threshold.

[0059] Furthermore, the first threshold The value range is from 0.5 to 0.9, the second threshold. The value range is from 0.3 to 0.7, and the third threshold... The value range is from 0.1 to 0.3, and the specific value of the above threshold is determined through offline calibration based on the actual system's false alarm rate and false negative rate requirements.

[0060] Preferably, the method further includes an online update step: when no rejection is triggered, the maximum a posteriori corresponds to the health hypothesis, the health consistency score is higher than the health threshold, and the uncertainty is lower than the health update threshold, the local sensitivity matrix of the operating condition disturbance or its parameter dictionary is updated online.

[0061] Preferably, the high-confidence health window is determined by simultaneously satisfying the following conditions at multiple consecutive sampling times: health posterior dominance, health consistency score higher than a threshold, posterior uncertainty lower than a threshold, and no rejection trigger.

[0062] Preferably, a fault diagnosis system for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints is used to execute the aforementioned fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints. The system includes: The data acquisition module is used to obtain a set of operational information. The coordinate transformation and diagnostic model construction module has its input end connected to the output end of the data acquisition module, and is used to transform the running information into a synchronous rotating coordinate system and construct a diagnostic model. The state reconstruction module, whose input is connected to the output of the coordinate transformation and diagnostic model construction module, is used to obtain the state estimate and the original residual through the state observer. The perturbation projection module, whose input is connected to the output of the state reconstruction module, is used to construct a healthy perturbation subspace and project the original residual onto its orthogonal complement to obtain the de-perturbation projection residual. The conduction path can trigger the determination module to generate a fault visibility mask; The fault matching and posterior evaluation module has its first input terminal connected to the output terminal of the disturbance projection module to receive the descrambling projection residual, and its second input terminal connected to the output terminal of the conduction path excitation determination module to receive the fault visibility mask, which is used to calculate the fault matching score and posterior probability. The diagnostic output module, whose input is connected to the output of the fault matching and post-evaluation module, is used to perform rejection judgment and output diagnostic results. The online update module, whose input is connected to the output of the diagnostic output module, is used to update the perturbation sensitivity matrix within a high-confidence health window.

[0063] Furthermore, the data flow between the modules is as follows: the output of the data acquisition module is connected to the input of the coordinate transformation and diagnostic model construction module; the output of the coordinate transformation and diagnostic model construction module is connected to the inputs of the state reconstruction module and the disturbance projection module, respectively; the output of the state reconstruction module is connected to the input of the disturbance projection module to provide the original residual; the output of the disturbance projection module is connected to the input of the fault matching and posterior evaluation module to provide the de-scratched projection residual; the output of the conduction path excitation determination module is connected to the input of the fault matching and posterior evaluation module to provide the fault visibility mask; the output of the fault matching and posterior evaluation module is connected to the input of the diagnostic output module to provide the posterior probability; and the output of the diagnostic output module is connected to the input of the online update module to provide the health status confirmation signal.

[0064] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints.

[0065] Preferably, a computer-readable storage medium stores computer instructions for causing a computer to execute the described fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints.

[0066] Example 2: like Figure 1 As shown, this embodiment provides a specific implementation process for a fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of the conduction path: S1. Establish the state variables and output variables in the synchronous rotating coordinate system.

[0067] Let the inverter-side current, filter capacitor voltage, and grid-side current be expressed as follows: , , ; Define the rotational coupling matrix as: ; The selected state vector and measurement output vector are as follows: ; ; in, This indicates the midpoint voltage deviation. The upper and lower capacitor voltages of the split DC bus are respectively... and Then we have: ; This definition is used to characterize the point potential balance state in a three-level NPC topology and serves as one of the state variables in the subsequent diagnostic model.

[0068] like Figure 2 As shown in Figure S2, establish a continuous-time model of the LCL filter network and the midpoint voltage deviation.

[0069] In the dq synchronous rotating coordinate system, the dynamic equations for the inverter-side inductor, capacitor branch, and grid-side inductor are as follows: ; ; ; In the formula, , These are the inverter-side inductance and its equivalent resistance, respectively. , These are the grid-side inductance and its equivalent resistance, respectively. For filtering capacitors, The equivalent voltage output by the inverter. This represents the grid voltage in the dq coordinate system.

[0070] For the midpoint voltage deviation state, assume that the values ​​of the upper and lower DC capacitors are both... The capacitance charges are respectively and Then we have: ; ; From the definition of midpoint voltage deviation, we can obtain: ; Define the unbalanced current at the midpoint. The positive direction is the direction of charging of the upper capacitor and discharging of the lower capacitor. Let: ; Differentiating the midpoint voltage deviation, we get: ; Because three-level NPC inverters have redundant conduction paths, different level implementations, phase current directions, and midpoint balance control will change the injection direction and weight of the midpoint current. Therefore, a conduction path coefficient is defined. , , ,That If we represent the current conduction context, then: ; Therefore, the dynamic midpoint voltage deviation can be written as: ; This expression reflects the coupling relationship between the midpoint deviation state and the conduction path context.

[0071] S3. Construct a unified disturbance-fault state space model.

[0072] The health condition disturbance vector is defined as: ; in, This represents a given change in photovoltaic power. and These represent the phase-locked loop angle and frequency deviation, respectively. This represents the change in equivalent grid-connected impedance. This indicates the change in midpoint balance modulation.

[0073] Define the fault vector as: ; in, For open-circuit fault components of switching devices, For the fault component of the current sensor, This represents the gate fault component. Therefore, the system can be uniformly written as: ; ; in, The current operating point parameter vector includes at least the phase-locked loop frequency, DC bus voltage, grid impedance, and midpoint balance control parameters. and These are process noise and measurement noise, respectively.

[0074] During the sampling period Next, by performing first-order discretization on the continuous-time model, we obtain: ; ; in: ; ; in , , , , , Let be the time-varying matrix obtained by discretizing the continuous model near the current running point, and let its subscript be . This indicates the dependency on changes in the running point. The aforementioned matrix changes slowly with the running point; therefore, this invention employs a combination of local linearization and online updates to achieve diagnosis.

[0075] S4. Construct a healthy reference observer and generate the original residuals.

[0076] Establish a discrete-time health reference observer: ; ; Therefore, the original residual is defined as follows: ; Let the state estimation error be Then the error dynamics approximately satisfy: ; in, The process noise, measurement noise, and higher-order linearization remainders are absorbed. Therefore, the original residual can be decomposed into: ; In the formula, The residual component caused by the disturbance of the healthy operating condition. The residual component caused by the fault. These are noise and modeling error components. If directly... Setting thresholds or classifying conditions can easily lead to misjudging changes in operating conditions as faults.

[0077] like Figure 3 As shown, S5, construct the healthy perturbation subspace and calculate the de-perturbation projection residual.

[0078] At the current running point Nearby, the original residuals are locally linearized with respect to various health condition disturbances to obtain the local disturbance sensitivity matrix: ; Therefore, the original residual can be locally approximated as: ; in, Let the coordinates of the health perturbation be on the local perturbation basis. This is the fault sensitivity matrix. These are higher-order terms and noise terms. To estimate the components of the original residuals in the healthy perturbation subspace, the following regular least squares problem is constructed: ; Its closed-form solution is: ; The residual estimation components for health disturbances are: ; Therefore, the scrambled projection residual is defined as: ; Substituting the above equation and rearranging, we obtain the projection form: ; ; In the formula, This represents the orthogonal complement projection matrix of the health disturbance subspace. The physical meaning of this step is to remove only the portion of the original residual that can be explained by changes in health conditions, thereby preserving the unexplainable components caused by the fault to the greatest extent possible.

[0079] To ensure that fault information is not eliminated during projection, the following locally separable condition is preferably met: ; Alternatively, the minimum protagonist constraint can be used: ; in, This represents the health-perturbation subspace. Let represent the subspace of the j-th type of fault.

[0080] like Figure 4 As shown, S6, constructing a conduction path can trigger constraint and fault visibility masks.

[0081] Define the current topology context variable as: ; in, This represents the current switch state. The symbol for the direction of three-phase current. For midpoint balanced modulation or zero vector allocation parameters, This represents the midpoint voltage deviation.

[0082] Definition of the first The local sensitivity vector of the fault class to the original residual is: ; After projecting it with a health perturbation, we get: ; Further define the excitation function of the conduction path. When the first When a fault can propagate to an external measurement output under the current conduction path, ;otherwise Therefore, the fault visibility mask at the current moment is defined as: ; This allows us to construct the currently visible set of faults: ; if If the value is empty, it means that no fault assumptions meet the condition of "physically excitable and still recognizable after projection" at the current sampling time. Therefore, the rejection should be output directly, and a fault label should not be forcibly given.

[0083] like Figure 5 As shown, S7, fault matching score, posterior probability, and uncertainty assessment.

[0084] In a preferred embodiment, health status is also considered as a candidate hypothesis, denoted as... It participates in the posterior normalization evaluation together with each fault hypothesis in the currently visible fault set.

[0085] In a preferred embodiment, the normalized matching score between the descrambling projection residual and the fault projection signature is used as the diagnostic criterion. The matching score for the j-th type of fault is defined as: ; In the formula, To prevent small positive numbers with a denominator of zero, the larger the score, the more consistent the current descrambling residual is with the fault signature. Posterior normalization is performed only within the currently visible fault set, resulting in: ; For invisible faults, their posterior probability is defined as: ; The maximum posterior probability is defined as: ; Define the uncertainty of the normalized entropy on the currently visible set as: ; Define the posterior interval of the top two as: ; If we denote the health assumption as j=0, then the health consistency score can be expressed as: ; S8, Rejection Rules and Fault Output.

[0086] To avoid misjudgment due to insufficient evidence, this invention establishes the following rejection rules: ; That is, when there is no visible fault, the maximum posterior is insufficient, the posterior distribution is too scattered, or the distinction between the top two candidates is insufficient, the output is rejected.

[0087] when When the fault index is reached, output the fault index: ; Then, output the fault category based on the pre-established index mapping table. and location of the fault ;when When the condition is met, output an unknown state or a rejection result.

[0088] S9, Online Update Strategy.

[0089] To accommodate drift under slowly varying operating conditions, it is preferable to update the health perturbation sensitivity matrix or its parameter dictionary only within a high-confidence health window. The online update trigger condition is defined as follows: ; when When updating, the following recursive form is used: ; Where α∈(0,1) is the update step size. The local perturbation sensitivity matrix is ​​estimated for the current health window; when When updating, freeze the process to avoid corrupted samples contaminating the model.

[0090] Furthermore, the conduction path can excite function Construction: In a preferred embodiment, This is implemented using table lookup logic. A truth table is pre-constructed according to "bridge arm realization state * phase current direction * zero vector allocation method * midpoint balance bias direction". The truth table is then updated online based on the current... Query whether the j-th type of fault can be triggered.

[0091] In another preferred embodiment, It can be determined by local transmission sensitivity, that is, let: ; in, Indicates the first Local transmission sensitivity from fault type to output quantity The corresponding threshold.

[0092] Example 3: This embodiment uses a 100kW grid-connected three-level NPC inverter system as the object, and builds a fault diagnosis simulation experimental platform in the MATLAB / Simulink environment. The DC bus voltage is 700V, and the upper and lower DC side capacitors are each... Inverter-side filter inductor and its equivalent resistance AC filter capacitor Grid-side filter inductor and its equivalent resistance The grid phase voltage RMS value is 220V / 50Hz. IGBT modules are used as the power switching devices, with a switching frequency of 5kHz. A space vector modulation strategy is employed, and midpoint balance control is achieved using a zero-vector allocation method. The system sampling frequency... Each continuous sampling session lasts 0.2 seconds and includes 2000 sampling points.

[0093] The key parameters for this method are set as follows: residual weighting matrix. Take it as a unit diagonal matrix, regularization parameter Temperature parameters Regularization of small quantities Reject the first threshold Refusing to accept the second threshold Refusing to judge the third threshold Sensitivity thresholds for each fault Health threshold Health update threshold Online step size update Separable angle threshold .

[0094] The experiment covered four operating conditions: operating conditions For steady-state operation at 100kW rated power, operating conditions The photovoltaic power is changed from 100kW to 60kW at a ramp rate and then restored to 100kW. To generate a ±1.5Hz step disturbance in the phase-locked loop frequency based on the fundamental frequency of 50Hz, the operating condition is... For grid-connected equivalent inductance from Mutation to The fault types include open-circuit faults in switching devices, zero-drift faults in current sensors, and gate polarity pulse faults. There are a total of 12 fault modes for single-transistor faults and 6 modes for dual-transistor combination faults, plus 19 candidate hypotheses for healthy states. Each type of fault under each operating condition was tested 50 times in Monte Carlo simulation.

[0095] For comparison, two existing methods are selected: Method 1 is a diagnostic method based on an adaptive sliding mode observer and current residual location variables, referenced in Xu Shuiqing et al., *Journal of Electrical Engineering Technology*, 2023, Vol. 38. Its observer uses the same power circuit parameters and adaptive reaching law parameters. , Sliding surface coefficient is taken as The threshold for fault location variables is set at 5% of the fundamental current amplitude, as per the original text. Comparison Method Two is a diagnostic method based on residual evaluation plus a norm evaluation function. The reference is Chen Chaobo et al., *Journal of Electrical Engineering Technology*, 2024, Vol. 39. Its evaluation function uses the 2-norm, and the threshold is set adaptively according to 5% of the residual standard deviation. The data are shown in Tables 1 and 2 below.

[0096] Table 1: Diagnostic accuracy and rejection rate of various faults under different operating conditions using this method;

[0097] Table 2: Comparison of the overall performance of this method with two comparative methods under the same experimental conditions;

[0098] The comparative data above demonstrates that this method exhibits significant advantages in accuracy, robustness, and reliability for fault diagnosis of three-level NPC grid-connected inverters. Under steady-state conditions, the correct diagnosis rates of the three methods are not significantly different, with this method achieving 95.8%, while Comparison Method 1 and Comparison Method 2 achieve 95.2% and 94.6%, respectively, indicating that all three methods possess basic fault detection capabilities under ideal conditions. However, the differences widen considerably when operating conditions change: under varying photovoltaic power conditions, the correct diagnosis rate of this method is 92.7%, while Comparison Method 1 drops to 84.3%, and Comparison Method 2 drops to 86.5%. The differences are even more pronounced under phase-locked loop frequency disturbance conditions, with this method achieving 94.0%, Comparison Method 1 at only 82.1%, and Comparison Method 2 at 87.3%. The fundamental reason for this difference is that this method effectively removes the systematic contamination of the residuals by the changes in operating conditions through orthogonal projection of the healthy disturbance subspace, so that the de-disturbed projection residuals mainly retain fault information. In contrast, the two comparative methods use the residuals as a whole for diagnosis, and cannot distinguish between healthy disturbance components and fault components. Therefore, their performance is significantly reduced under disturbed operating conditions.

[0099] Regarding the handling of invisible faults, both comparative methods 1 and 2 lack a conduction path excitation constraint mechanism, resulting in misclassification rates of 8.4% and 6.7%, respectively. Our proposed method forces the posterior probability of invisible faults to zero using a fault visibility mask, achieving a zero misclassification rate and completely eliminating such errors. At sampling moments with insufficient evidence, both comparative methods 1 and 2 unconditionally output hard decisions, with error rates of 35.2% and 28.6%, respectively. Our method achieves a correct rejection rate of 91.5%, effectively avoiding the risk of incorrect labels generated by forced hard decisions. In terms of online update mechanisms, our method updates the sensitivity matrix through a high-confidence health window, achieving a long-term drift adaptation rate of 96.3%, superior to the two comparative methods. The average diagnosis time of our method is 5.7ms, slightly longer than the two comparative methods, but considering the added perturbation projection and rejection decision steps, this diagnostic delay remains within an acceptable range.

[0100] Figure 6 This paper demonstrates the core advantages of our proposed method over traditional residual-based diagnostic methods. When photovoltaic power ramp changes and phase-locked loop (PLL) frequency step disturbances are superimposed, the original residual contains systematic deviation components caused by these health condition changes. The amplitude of these deviations is on the same order of magnitude as the residual components caused by faults. If the original residual is directly input into the diagnostic tool, the changes in operating conditions are easily misjudged as faults. Our method constructs a health disturbance subspace at the current operating point and projects the original residual onto its orthogonal complement to obtain a de-disturbed projected residual. The time-domain waveform shows that the low-frequency drift components caused by photovoltaic power changes and PLL frequency disturbances are effectively suppressed in the de-disturbed projected residual, with only a significant amplitude jump occurring at the moment of fault occurrence. This achieves the separation of health disturbances and fault components, improving the accuracy and robustness of fault detection.

[0101] Figure 7 This paper demonstrates the impact of conduction path excitation constraints on fault diagnosis results in this method. Under specific switch states and phase current directions, certain types of faults cannot propagate to external measurement outputs through the conduction path at the current moment. Traditional methods do not utilize this constraint and directly compare across the entire fault space, which may lead to non-zero posterior probabilities or even high-confidence misjudgments for these invisible faults. This method generates a fault visibility mask, forcing the posterior probability of invisible faults to zero and normalizing it only within the currently visible fault set. As shown in the figure, after applying the conduction path constraint, the posterior probability distribution is concentrated on the truly visible fault categories, while the rejection status flag is correctly triggered at the sampling moment with insufficient evidence, avoiding forced hard decisions.

[0102] Example 4: like Figure 8As shown, this embodiment of the invention also provides a computer device, which includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory for displaying graphical information of a GUI on external input / output devices, such as display devices coupled to the interfaces. In some alternative embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations, for example, as a server array, a group of blade servers, or a multiprocessor system.

[0103] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0104] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0105] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0106] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0107] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0108] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0109] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A fault diagnosis method for three-level NPC grid-connected inverters based on perturbation subspace projection and conduction path excitation constraints, characterized in that, Includes the following steps: Obtain the operational information set at the current sampling moment. The operational information set includes at least the grid-side current, split DC bus voltage, phase-locked loop variables, control quantities, and the current switch implementation status. A diagnostic model is constructed in a synchronous rotating coordinate system based on the operational information set. The diagnostic model includes at least the state variables of the LCL filter network, the state variables of the midpoint voltage, the operational condition disturbances, and the faults. Based on the diagnostic model and known control variables, the state estimate and estimated output value consistent with the health condition are obtained through the state observer, and the original residual is obtained based on the measured output and the estimated output. Based on the local sensitivity matrix of the original residual to the operating condition disturbance under the current operating point, a healthy disturbance subspace is constructed, and the original residual is projected onto the orthogonal complement of the healthy disturbance subspace to obtain the de-disturbed projected residual. Based on the current switch implementation state, phase current direction, midpoint balance modulation state, and midpoint voltage state, determine the excitation potential of each fault hypothesis's conduction path at the current moment and generate a fault visibility mask. For each fault hypothesis in the current set of visible faults, calculate its matching score with the descrambling projection residual, and perform posterior normalization only within the current set of visible faults to obtain the fault posterior probability. The rejection decision is made based on at least the maximum posterior probability, the uncertainty, and the interval between the top two posterior probabilities. Output the fault category and fault location when no rejection is triggered, and output the rejection result when rejection is triggered.

2. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitation constraints of conduction paths as described in claim 1, characterized in that, The operational information set also includes photovoltaic power input, phase-locked loop angle, phase-locked loop frequency, and midpoint balance modulation related quantities; the split DC bus voltage is used to calculate the midpoint voltage deviation. ； in, The voltage across the upper capacitor. This is the voltage across the lower capacitor.

3. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, The state vector of the diagnostic model includes at least the inverter-side d-axis current, inverter-side q-axis current, filter capacitor d-axis voltage, filter capacitor q-axis voltage, grid-side d-axis current, grid-side q-axis current, and midpoint voltage deviation.

4. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, The operating condition disturbances include at least one or more of the following: changes in photovoltaic power input, changes in phase-locked loop angle, changes in phase-locked loop frequency, changes in grid impedance, and changes in midpoint balancing modulation.

5. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitation constraints of conduction paths according to claim 1, characterized in that, The faults include open circuit faults of switching devices, current sensor faults, and gate abnormality faults. Among them, open circuit faults of switching devices and gate abnormality faults affect the status channel, while current sensor faults affect the measurement channel.

6. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitation constraints of conduction path as described in claim 5, is characterized in that, The gate abnormality fault is an equivalent open circuit fault caused by gate pulse failure, gate jamming, or gate drive failure.

7. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, The state observer includes a prediction phase and a correction phase; the prediction phase obtains the predicted state based on the diagnostic model, the known control variables, and the estimated values ​​of the operating condition disturbances; the correction phase corrects the predicted state based on the innovation between the measured output and the predicted output to obtain the state estimate and the estimated output value.

8. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, Projecting the original residuals onto the orthogonal complement of the healthy perturbation subspace includes: ； ； ； in, This is the local sensitivity matrix of the original residual to operating condition disturbances. The residual weighting matrix, For regularization parameters, It is the identity matrix. For the original residual, For the estimated coordinates of the health disturbance, For the estimation components of the health disturbance residuals, To remove scrambling from the projected residuals.

9. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 8, characterized in that, The orthogonal complement projection matrix of the health disturbance subspace is: ； The scrambling projection residual satisfies The healthy perturbation subspace and each fault subspace satisfy the condition of local separability. Or satisfy the minimum protagonist constraint ,in For health-related disturbance subspaces, For the first Fault-like subspace, The separable angle threshold.

10. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, The fault visibility mask is determined by the conduction path excitation function and the projected fault sensitivity, satisfying the following conditions: ； in, For the first Class of faults at the current moment Visibility mask, Indicates the first The result of determining the triggerability of this type of fault under the current conduction path. For the first The fault sensitivity vector after projecting the fault class onto the health perturbation subspace. This is the sensitivity threshold.

11. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 10, characterized in that, The conduction path excitation function depends on the current topology context variable. Confirmed, among which This represents the current state of the bridge arm. The symbol for the direction of three-phase current. For midpoint balanced modulation or zero vector allocation parameters, This represents the midpoint voltage deviation.

12. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitation constraints of conduction paths according to claim 1, characterized in that, The fault matching score satisfies ； in, For the first Class of faults at any time Match score, This is a fault visibility mask. To descramble the projected residuals, For the first The projected fault sensitivity vector of the fault class. The residual weighting matrix, This is for regularizing small quantities.

13. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, The posterior normalization applies only to the currently visible set of faults. Internal execution, satisfying ； Let the health hypothesis be denoted as And together with each fault hypothesis in the current set of visible faults, they participate in posterior normalization. For the invisible fault hypothesis, we have: ;in For the first The posterior probability of a fault class For temperature parameters, This is for regularizing small quantities.

14. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, The uncertainty is determined by the normalized entropy on the currently visible fault set, satisfying the following condition. ； in, For a moment Uncertainty, The cardinality of the currently visible set of faults. To regularize small quantities, To regularize small quantities, Let be the posterior probability.

15. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitation constraints of conduction paths according to claim 1, characterized in that, A rejection is triggered when any of the following conditions are met: the currently visible fault set is empty; the maximum posterior probability is lower than the first threshold; the uncertainty is higher than the second threshold; or the posterior interval between the top two faults is lower than the third threshold.

16. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 1, characterized in that, It also includes an online update step: when no rejection is triggered, the maximum a posteriori corresponds to the health hypothesis, the health consistency score is higher than the health threshold and the uncertainty is lower than the health update threshold, the local sensitivity matrix of the operating condition disturbance or its parameter dictionary is updated online.

17. The fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable constraints of conduction path as described in claim 16, characterized in that, The high-confidence health window is determined by simultaneously satisfying the following conditions at multiple consecutive sampling times: health posterior dominance, health consistency score higher than the threshold, posterior uncertainty lower than the threshold, and no rejection trigger.

18. A fault diagnosis system for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable conduction path constraints, used to execute the fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and excitable conduction path constraints as described in any one of claims 1-17, characterized in that the system... include: The data acquisition module is used to obtain a set of operational information. The coordinate transformation and diagnostic model construction module has its input end connected to the output end of the data acquisition module, and is used to transform the running information into a synchronous rotating coordinate system and construct a diagnostic model. The state reconstruction module, whose input is connected to the output of the coordinate transformation and diagnostic model construction module, is used to obtain the state estimate and the original residual through the state observer. The perturbation projection module, whose input is connected to the output of the state reconstruction module, is used to construct a healthy perturbation subspace and project the original residual onto its orthogonal complement to obtain the de-perturbation projection residual. The conduction path can trigger the determination module to generate a fault visibility mask; The fault matching and posterior evaluation module has its first input terminal connected to the output terminal of the disturbance projection module to receive the descrambling projection residual, and its second input terminal connected to the output terminal of the conduction path excitation determination module to receive the fault visibility mask, which is used to calculate the fault matching score and posterior probability. The diagnostic output module, whose input is connected to the output of the fault matching and post-evaluation module, is used to perform rejection judgment and output diagnostic results. The online update module, whose input is connected to the output of the diagnostic output module, is used to update the perturbation sensitivity matrix within a high-confidence health window.

19. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the fault diagnosis method for a three-level NPC grid-connected inverter based on the excitation constraints of the perturbation subspace projection and the conduction path, as described in any one of claims 1 to 17.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the fault diagnosis method for a three-level NPC grid-connected inverter based on perturbation subspace projection and conduction path excitation constraints as described in any one of claims 1 to 17.

Images