Multiple fault source identification method for cascaded H-bridge inverter

Abstract

This invention discloses a method for identifying multiple fault sources in a cascaded H-bridge inverter. The method involves acquiring real-time time series data from the inverter online; generating a real-time system residual function representing system deviations using a physical information health system operator dynamically weighted by the operating state vector; and decomposing the residual function using a three-layer cascade operator decomposition architecture that accounts for physical topology constraints to identify single faults and coupled effects. This method is adaptive to dynamic operating conditions, effectively suppresses spurious residuals, and accurately decouples concurrent faults, significantly improving diagnostic accuracy and robustness.

Description

Technical Field

[0001] The present invention relates to fault source identification, in particular to a method for identifying multiple fault sources of a cascaded H-bridge frequency converter. Background Art

[0002] Cascaded H-bridge (CHB) inverters, due to their significant advantages in modular design, multiple output levels, low harmonic content, and minimal voltage stress, have become core power electronics equipment for medium-, high-voltage, and high-power applications. Their application scenarios are broad, encompassing large-scale renewable energy grid integration (such as photovoltaic and wind power), flexible direct current transmission (HVDC), industrial heavy-duty drives (such as rolling mills and mine hoists), and national strategic industries such as electrified transportation. CHB systems, consisting of numerous power cells connected in series, have a complex structure and incorporate numerous power semiconductors (such as IGBTs), energy storage components (such as DC capacitors), and sensors. While this structural complexity offers performance advantages, it also results in a relatively high probability of failure and diverse failure modes. The failure of any single component can trigger a chain reaction, shutting down the entire inverter system, resulting in significant economic losses and potential safety hazards. Therefore, researching and developing a fast, accurate, and robust fault diagnosis technology for CHB inverters is of vital theoretical significance and practical engineering value for ensuring the stable operation of critical infrastructure, improving industrial production efficiency, and reducing operation and maintenance costs.

[0003] Research on fault diagnosis technologies for CHB inverters has made considerable progress, resulting in the development of various technical approaches. Among these, mathematical model-based methods are a traditional and established approach. These methods establish precise state-space equations, transfer functions, or parameter equivalent models of the inverter and then compare the differences (i.e., residuals) between the model outputs and actual system measurements online to detect faults. Another broad category involves signal processing-based methods. These methods do not rely on precise system models but instead directly process and analyze acquired electrical signals (such as output voltage and bridge arm current). Common techniques include using the Fast Fourier Transform (FFT) to analyze changes in harmonic components at specific frequencies, or using the Wavelet Transform to capture transient singular signal characteristics at the moment of fault occurrence. Additionally, methods such as the Hilbert-Huang Transform (HHT) analyze the instantaneous frequency and amplitude of signals. In recent years, with the advancement of artificial intelligence (AI), data-driven diagnostic methods have also gained widespread application. This type of method usually uses machine learning models such as support vector machines (SVM), traditional artificial neural networks (ANN) or decision trees to learn statistical features or frequency domain features extracted from the original signal and train a classifier that can classify the system status as healthy or in a specific failure mode.

[0004] However, when dealing with increasingly complex actual industrial application scenarios, existing technologies have gradually exposed their inherent limitations, mainly the problem of pseudo-residuals under dynamic working conditions and the nonlinear coupling and strong-weak masking of multiple fault characteristics. Summary of the Invention

[0005] The purpose of the invention is to provide a method for identifying multiple fault sources of a cascaded H-bridge inverter in order to solve the technical problems in the prior art that false alarms are easily generated under changing working conditions and concurrent coupling faults are difficult to diagnose accurately.

[0006] Technical solution: A method for identifying multiple fault sources of a cascaded H-bridge inverter, comprising:

[0007] Get real-time time series data of the inverter online;

[0008] Using pre-trained healthy system operators and real-time time series data, a real-time system residual function is generated to characterize the deviation of the actual operating behavior of the purifier from the healthy physical laws;

[0009] Based on a preset basic fault operator library containing multiple single fault modes or a three-layer cascade operator decomposition architecture that takes into account physical topology constraints, the real-time system residual function is decomposed to obtain the fault identification result.

[0010] According to one aspect of the present application, the steps of decomposing the residual function of the real-time system using the basic fault operator library specifically include:

[0011] Extracting at least part of the data from the real-time time series data as a real-time system input function, inputting the data into each basic operator in the basic fault operator library, and generating a candidate residual function;

[0012] By performing similarity matching between the candidate residual function and the current residual function, the best matching basic operator is selected from the basic fault operator library as the fault component identified in the current iteration;

[0013] The current residual function is updated for the next iteration until the energy of the current residual function is lower than a preset value, and the decomposition is completed.

[0014] According to one aspect of the present application, the process of identifying the fault component and updating the current residual function includes:

[0015] Calculate the projection of the current residual function on the candidate residual function, and determine the fault activation coefficient based on it to quantify the contribution of the fault component identified in the current iteration;

[0016] Determine the fault component through the fault activation coefficient and the best matching basic operator;

[0017] The fault component is subtracted from the current residual function to generate an updated current residual function for the next iteration.

[0018] According to one aspect of the present application, the health system operator is a physical information neural operator configured with a physical topology-aware architecture;

[0019] Each power unit of the inverter is a node of the graph, and the electrical connections between the power units are edges of the graph, constructing the physical graph of the system;

[0020] The features of the nodes in the system physical graph are processed through the graph neural network layer to simulate the transmission effect of physical quantities in the physical structure of the inverter.

[0021] According to one aspect of the present application, when operating, the physical information neural operator includes a working condition perception process, specifically:

[0022] Extract multiple characteristic parameters representing the current operating conditions from real-time time series data and encode them into an operating condition state vector;

[0023] By utilizing the weight modulation function of working condition perception, the inherent weight of the physical information neural operator is modulated in real time according to the working condition state vector to generate a dynamic health system operator that can adapt to the current operating conditions.

[0024] According to one aspect of the present application, a three-layer cascade operator decomposition architecture is used to decompose the residual function of the real-time system, including:

[0025] The operating condition pseudo-fault decomposition layer separates and identifies the pseudo-fault components caused by drastic changes in operating conditions from the initial residual function and outputs the first-layer decomposition residual;

[0026] The real fault decomposition layer separates and identifies the single real physical fault component from the first-layer decomposition residual and outputs the second-layer decomposition residual;

[0027] The fault coupling effect decomposition layer separates and identifies the nonlinear coupling effects caused by the concurrent occurrence of multiple real physical faults from the second-layer decomposition residuals.

[0028] According to one aspect of the present application, similarity matching in a three-layer cascade operator decomposition architecture adopts a physical topology constrained matching strategy;

[0029] When calculating the similarity between the current residual function and the candidate residual function, the physical topology constraint matching strategy introduces a topology matching function that quantifies the degree of matching between the physical positions and influence ranges of the two in the inverter circuit topology as a weight factor for the similarity calculation.

[0030] According to one aspect of the present application, before decomposing the residual function of the real-time system, a residual purification process is also included, specifically:

[0031] Separate the operating condition-related pseudo residuals caused by operating condition changes from the original residual function, and deduct the influence of the operating condition-related pseudo residuals to obtain a purer fault-related true residual for subsequent decomposition;

[0032] Among them, the effects of the pseudo residuals related to the working conditions are deducted from the original residual function, specifically including:

[0033] Matching the working condition-related pseudo residuals with a pre-trained library of pseudo fault signatures to identify and quantify the working condition switching effects and generate pseudo fault compensation vectors.

[0034] And deduct the pseudo fault compensation vector from the original residual function.

[0035] According to one aspect of the present application, in a three-layer cascade operator decomposition architecture, the decomposition process of the fault coupling effect decomposition layer is based on the single real physical fault component that has been separated and identified in the real fault decomposition layer to determine the candidate coupling effect operator and match and decompose the second-layer decomposition residual.

[0036] According to one aspect of the present application, the generation of the operating state vector includes:

[0037] Extracting at least one electrical quantity parameter representing the system operating load and control state from the real-time time series data;

[0038] The extracted electrical quantity parameters are encoded to generate an operating state vector;

[0039] The modulation of the intrinsic weight is a gradual modulation, including:

[0040] When a change in the operating state vector is detected, the weights are gradually adjusted over time under preset physical constraints that limit the rate of change of the weights until a dynamic health system operator matching the current operating condition is generated.

[0041] According to one aspect of the present application, the real-time time series data is multimodal time series data that fuses global electrical response and local physical effects, including:

[0042] Electrical quantity data, including voltage and current, to capture system-level electrical disturbances caused by faults;

[0043] Non-electrical quantity data, including temperature, high-frequency vibration, or leakage magnetic field signals at specific locations, are used to capture local effects caused by the same fault at the physical level and have high positioning value;

[0044] The health system operator is trained to learn and characterize the intrinsic physical relationship between electrical quantity data and non-electrical quantity data in a healthy state.

[0045] According to one aspect of the present application, the construction of a basic fault operator library includes:

[0046] Injecting a single fault mode into the simulation model in a parameterized manner to generate a parameterized single fault simulation data set;

[0047] For each single fault mode, a dedicated single fault residual operator is trained based on a parameterized single fault simulation dataset. The single fault residual operator is used to learn the mapping from system input to the residual function corresponding to the fault.

[0048] According to one aspect of the present application, the health system operator is a physical information neural operator;

[0049] The physical information neural operator is trained to learn the mapping rules from one or more system input functions to multiple system output functions, thereby simulating the physical dynamics of the inverter.

[0050] Beneficial effects: The present invention can adapt to dynamic working conditions, effectively suppress pseudo-residuals, and accurately decouple concurrent faults, significantly improving the accuracy and robustness of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the present invention.

[0052] Figure 2 This is a flow chart of the present invention using a basic fault operator library to decompose the residual function of a real-time system.

[0053] Figure 3 This is a flow chart of the present invention for identifying fault components and updating the current residual function.

[0054] Figure 4 It is a flow chart of the working condition perception process of the present invention. DETAILED DESCRIPTION

[0055] Combine Figures 1 to 4 And specific implementation cases, describing the technical details and principles of this application.

[0056] First, the applicant's research has found that existing diagnostic methods, whether model-based or data-driven, typically establish static health baselines, calibrating or training them for one or a few specific, steady-state operating conditions. In real industrial environments, the operating conditions of inverters (such as load level, speed command, grid voltage, etc.) frequently and dynamically change. When operating conditions change drastically, even if the system itself is completely healthy, its output electrical signal will produce a normal, but drastic, transient response. This response is a significant deviation from the static health baseline, resulting in a pseudo-residual peak in the residual signal whose amplitude may far exceed the true fault signature. Existing methods lack a mechanism that enables the health baseline to adaptively adjust in real time and dynamically to the physical laws of the operating conditions. This makes it very easy for these normal transient responses to be misclassified as faults under changing operating conditions, generating a large number of false alarms and severely limiting the reliability and credibility of diagnostic technology in practical engineering applications.

[0057] Secondly, when two or more faults occur concurrently in a CHB inverter, their combined impact on the system's overall electrical signal is not simply a linear summation of their individual fault signatures. Different faults can generate complex nonlinear coupling effects through the circuit's physical loops, generating new and unpredictable fault signatures and potentially distorting or suppressing the original single fault signature. Even more challenging, when one fault signature is prominent (such as a complete open circuit in a power device, a strong fault) while another is still in its early stages (such as a slight degradation in capacitor performance, a weak fault), the significant electrical disturbance generated by the strong fault can act like noise, completely masking the weak fault signature at the signal level. Traditional methods, relying on signal analysis to isolate multiple unknown sources from highly aliased, coupled, and noise-contaminated signals, are like blind men touching an elephant. They often only identify the most significant strong fault, leaving them unable to address weaker or concurrent faults that are more valuable as early warnings. This leads to missing the optimal maintenance window and fails to meet the urgent need for high-precision, multi-target concurrent diagnosis required for modern predictive maintenance.

[0058] To this end, the following implementation scheme is provided:

[0059] First, the present invention constructs two core mathematical models: the healthy system operator model (healthy system operator) and the basic fault signature operator model (basic fault signature operator). Both models are built and trained using the Physical Information Neural Operator (PINO) architecture (Physical Information Neural Operator).

[0060] The healthy system operator model serves as an ideal health reference. It receives the inverter's real-time system input functions (such as control signals) and outputs a predicted healthy system output function. In other words, it is a pre-trained deep neural network model with a physical information neural operator (PINO) architecture. This model is used to learn and characterize the deterministic mapping relationship between system input functions (such as control signals) and system output functions (such as multiple voltages and currents) for a cascaded H-bridge (CHB) inverter in a fully healthy state. In this invention, its role is to provide an ideal, dynamic health benchmark.

[0061] Basic fault signature operator models are organized into a library. Each model in the library uniquely corresponds to a single fault mode and receives the same system input function and outputs a candidate residual function caused only by that specific single fault.

[0062] In other words, it also uses a PINO architecture to train a deep neural network model specifically for a single, standardized fault mode. For example, the basic fault signature operator corresponding to an open-circuit IGBT in arm A of unit 1 learns the mapping from the system input function to the system residual function caused only by this specific fault. All these operators for different single faults constitute the basic fault operator library.

[0063] Next, the actual measured system output function of the inverter is compared (or subtracted) from the predicted healthy output function to obtain the real-time system residual function. Ideally, a healthy system's residual function should be close to zero. However, when a system fault occurs, the residual function will carry all the fault information.

[0064] The real-time system residual function (ε(t)) is a time series function defined in a function space. It is calculated by point-by-point subtraction of the measured inverter output function from the ideal healthy output function predicted by the healthy system operator model under the same inputs. This function is the target of all subsequent decomposition and identification steps.

[0065] Finally, the present invention provides a three-layer cascade decomposition architecture (three-layer cascade decomposition operator), which is a structured algorithmic process that processes the real-time system residual function generated in the previous step. Using multiple candidate residual functions generated by a basic fault signature operator model library as probes or substrates, the complex real-time system residual function is finely decomposed through an iterative matching and stripping process that takes into account physical topology constraints. The final product of the decomposition, the fault identification result, clearly indicates the type and severity of single faults, concurrent faults, and even their nonlinear coupling effects currently existing in the system. The complex residual decomposition problem is orderly divided into three progressive levels: operating condition pseudo-fault decomposition, single real fault decomposition, and fault coupling effect decomposition. In summary, the entire technical process can be summarized as: using the operator model to generate the residual function, then parsing the residual function through the decomposition architecture, and finally obtaining the fault identification result.

[0066] Example 1 describes the general process of the method for identifying multiple fault sources of a cascaded H-bridge inverter, which mainly includes the following steps:

[0067] Step S100: Acquire the real-time time series data of the inverter online.

[0068] In this embodiment, this step is used to obtain high-quality, multi-dimensional, time-synchronized data required for subsequent diagnostic analysis from the running CHB inverter.

[0069] Among them, real-time time series data is a collection of a series of sensor readings with precise time stamps that are continuously collected at a constant sampling rate.

[0070] For example, a distributed data acquisition system based on an FPGA and integrated with PTP (Precision Time Protocol) performs high-frequency synchronous sampling of raw multimodal sensor signals from each power unit and key nodes of the CHB inverter (such as unit output voltage, bridge arm current, DC link voltage, module temperature, etc.). The sampling rate is no less than 10 MHz, and the PTP protocol ensures that the timestamp error of multiple sensor signals across physical modules is within 100 nanoseconds.

[0071] The collected raw digital signal sequence is aligned, windowed, and normalized. Specifically, resampling and interpolation alignment are performed using a unified time base. Subsequently, the continuous data stream is divided into multiple data windows of length T (for example, T = 20ms, covering one power frequency cycle) and a 50% overlap rate. Finally, based on the maximum and minimum values ​​of each channel signal obtained from a large amount of historical health data, the data is normalized using min-max, scaling the values ​​to the range [0, 1] to obtain synchronized multimodal high-frequency time series data.

[0072] High-frequency synchronous sampling is essential for accurately capturing transient fault characteristics. Windowing converts continuous data streams into fixed-length input samples suitable for neural network model processing. Normalization eliminates the influence of different physical dimensions and accelerates the convergence of model training.

[0073] Step S200: using the pre-trained healthy system operator and real-time time series data, a real-time system residual function is generated to characterize the deviation between the actual operating behavior of the inverter and the healthy physical law.

[0074] In this embodiment, this step primarily constructs a baseline model and calculates the deviation of the real system from this baseline in real time. The healthy system operator is a deep model that accurately mimics the complete physical laws of a healthy CHB inverter. In this embodiment, it is implemented using the Physical Information Neural Operator (PINO-H). Learning involves the infinite-dimensional mapping from system input functions (such as control signals) to output functions (such as output voltage). The real-time system residual function is the difference between the actual system output function and the ideal healthy output function predicted by the healthy system operator. Its waveform and dynamic characteristics contain information about all unknown faults currently acting on the system.

[0075] For example, in each analysis window, according to the preset channel definition, the real-time time series data obtained in step S100 is dynamically divided into two parts: the real-time system input function urealtime(t) as the system driver and the real-time system actual output function y that represents the system state. realtime (t).

[0076] Enter the real-time system into function u realtime (t) is input into the pre-trained PINO-H operator G and a forward propagation calculation is performed to obtain the predicted health output function y predicted_healthy (t)=G(u realtime (t)).

[0077] y realtime (t) and y predicted_healthy For each corresponding dimension of (t), the difference calculation is performed point by point to obtain the real-time system residual function ε(t)=y realtime (t)-y predicted_healthy (t).

[0078] This transforms the fault diagnosis problem from directly analyzing complex raw signals to analyzing a residual function with a cleaner background and a strong correlation with the fault information. In a healthy state, this residual function should theoretically be close to zero; however, once a fault occurs, it will deviate significantly from zero, and its shape directly reflects the way the fault has disrupted the physical laws of the system.

[0079] Step S300: Based on a preset basic fault operator library containing multiple single fault modes, the real-time system residual function is decomposed to obtain a fault identification result.

[0080] In this embodiment, fault diagnosis is regarded as a decomposition problem in function space, which enables direct inversion from phenomenon to root cause.

[0081] The basic fault operator library is a structured collection of pre-built, offline operators. Each basic operator Fi in the library uniquely corresponds to a single, standardized fault mode (e.g., an open circuit in the IGBT on the A arm of unit 1).

[0082] The decomposition process refers to finding a set of the sparsest operators and their corresponding activation coefficients in the basic fault operator library, so that after these activated operators act on the current system input, their combination results can best reconstruct the observed real-time system residual function.

[0083] In one embodiment, an iterative decomposition algorithm is used:

[0084] Step S310: During the iteration process, each basic operator in the basic fault operator library is used to process a portion of the real-time time series data as a real-time system input function to generate a set of candidate residual functions.

[0085] Step S320: performing similarity matching between the candidate residual function and the current residual function, so as to select the best matching basic operator from the basic fault operator library as the fault component identified in the current iteration.

[0086] Step S330: Calculate the projection of the current residual function on the candidate residual function generated by the best matching basic operator, and determine the fault activation coefficient based on the projection to quantify the contribution of the fault component identified in the current iteration.

[0087] Step S340: Subtract the fault component determined by the fault activation coefficient and the best matching basic operator from the current residual function to generate an updated current residual function for the next iteration.

[0088] Step S350: determining the conditions for decomposition completion, including whether the energy of the current residual function is lower than a preset energy threshold.

[0089] For example, the main process is as follows:

[0090] During initialization, the initial residual R0(t)=ε(t) is set.

[0091] First iteration (k=1):

[0092] will u realtime (t) Input to all operators F in the libraryi , a candidate residual set {ε1*(t),..., ε N *(t)} is obtained.

[0093] Calculate the inner product (similarity) of each candidate residual and R0(t), and find the operator Fj with the largest absolute value of the inner product.

[0094] Calculate the activation coefficient α j = (R0(t), ε* j (t)) / ||ε* j (t)|| 2 2.

[0095] Update the residual R1(t) = R0(t) - α j ε* j (t).

[0096] Repeat the iteration until the energy (square of the L2 norm) of R k (t) is less than a preset threshold (e.g., 1% of the initial residual energy).

[0097] In this decomposition process, in each iteration, a most main and known fault component is stripped from the complex mixed residual until only noise remains in the residual. The finally obtained non-zero activation coefficient vector α is the fault diagnosis result. The positions of its non-zero elements indicate the fault type, and the magnitudes of its values are related to the fault severity.

[0098] Through the above steps, this embodiment can effectively identify the faults of the CHB inverter.

[0099] Embodiment 2: Further describe in detail the alternative solutions of step S200 and step S300.

[0100] By improving the dynamic adaptive ability of the health system operator and enhancing the physical constraints in the decomposition process, the accuracy and robustness of the diagnostic system under complex working conditions are improved.

[0101] Step S200': Generate purified multi-level residuals using a working condition adaptive dynamic health system operator.

[0102] In this embodiment, the health system operator is a physical information neural operator with a physical topology perception architecture, and this operator further includes a working condition perception process, which is as follows.

[0103] Step S210': Extract multiple characteristic parameters representing the current operating conditions from the real-time time series data and encode them into a working condition state vector.

[0104] Specifically, electrical parameters such as load current RMS value, power factor, and modulation index are extracted from real-time data. These parameters are input into a multi-layer perceptron (MLP) and encoded into a fixed-dimensional (e.g., 128-dimensional) operating state vector s(t).

[0105] Step S220 ′: using the weight modulation function of the working condition perception, the inherent weight of the physical information neural operator is modulated in real time according to the working condition state vector to generate a dynamic health system operator that can adapt to the current operating condition.

[0106] The modulation process is a gradual modulation. When it is detected that the change of the working state vector s(t) exceeds the threshold, a preset physical constraint (for example, |dW(s(t)) / dt|[cite s tart]≤β·E stored / E total (t)), the weights are gradually adjusted over time until a dynamic health system operator that matches the current operating conditions is generated. This dynamic, progressive modulation mechanism enables the health benchmark model to smoothly and stably adapt to changing operating conditions, fundamentally resolving the problem of spurious residuals caused by sudden changes in operating conditions and avoiding false alarms.

[0107] Step S300 ′: decomposing the residual function using a three-layer cascade operator decomposition architecture.

[0108] In this embodiment, the decomposition of the residual function is implemented using a three-layer cascade operator decomposition architecture, specifically including a working condition pseudo-fault decomposition layer, a real fault decomposition layer, and a fault coupling effect decomposition layer.

[0109] Step S310 ′: before decomposition, perform a residual purification process.

[0110] The operating condition-dependent pseudo residuals caused by operating condition changes are separated from the original residual function, and then the influence of the operating condition-dependent pseudo residuals is deducted from the original residual function. Specifically, the operating condition-dependent pseudo residuals are matched with a pre-trained operating condition pseudo-fault feature library to identify and quantify the operating condition switching effect, and then a pseudo-fault compensation vector is generated. The pseudo-fault compensation vector is deducted from the original residual function.

[0111] Step S320 ′: in the three-layer cascade operator decomposition architecture, a physical topology constraint matching strategy is adopted.

[0112] When calculating the similarity between the current residual function and candidate residual functions, a topology matching function is introduced to quantify the degree of matching between their physical locations and influence ranges in the inverter circuit topology. This function serves as a weighting factor in the similarity calculation. For example, if the residual function shows an abnormality in unit 5, it will be preferentially matched with the fault operator located in unit 5 in the fault database.

[0113] This strategy incorporates physical prior knowledge into the decomposition process, making the decomposition no longer a pure mathematical game, but a physically explainable reasoning process. It can effectively distinguish faults with different locations but similar signal morphologies, significantly improving positioning accuracy.

[0114] Step S330 ′: at the fault coupling effect decomposition layer, based on the single real physical fault component separated and identified in the real fault decomposition layer, determine candidate coupling effect operators, and match and decompose the second-layer decomposition residual.

[0115] Assume that after the second-level decomposition, the slight degradation of the IGBT in unit 1 (activation coefficient α1 = 0.3) and the aging of the capacitor in unit 5 (activation coefficient α5 = 0.2) are identified.

[0116] The third-level decomposition intelligently selects a specific compound fault operator from the fault library: the coupling of IGBT degradation (No. 1) and capacitor aging (No. 5). This operator is then matched against the residual error remaining after the second-level decomposition. If the match is successful and the activation coefficient α1,5 is significant, this indicates the presence of a significant nonlinear coupling effect in the system.

[0117] Through a layered and step-by-step decomposition strategy, complex mixed fault problems are broken down into smaller pieces and addressed one by one. First, operating condition interference is eliminated, then the main fault is identified, and finally its coupling effects are analyzed. The logic is clear and can handle multi-fault concurrent scenarios that are difficult to solve with traditional methods.

[0118] Through the above improvements, this embodiment can accurately identify up to three concurrent faults (including major and minor faults, and combinations of similar and heterogeneous faults) in a medium-sized CHB inverter (e.g., 11-level). Under conditions where the load suddenly changes from 20% to 90%, test data shows that the diagnostic accuracy can reach over 98%. This represents a significant technical improvement over Example 1 (accuracy of approximately 85%), which does not employ the improvements of the present invention, and the traditional method (accuracy below 70%).

[0119] Example 3: Detailed description of the implementation process of the working condition adaptive dynamic operator, which mainly includes the following steps:

[0120] Step S301: Construct and apply a dynamic health system operator with working condition perception process to generate a real-time system residual function. Describe a dynamic and adaptive health system operator.

[0121] In this embodiment, the health system operator not only possesses the physical topology perception architecture of Example 2, but also includes an operating condition perception process within the physical information neural operator. The operating condition state vector is a numerical vector that compactly and digitally represents the current macroscopic operating state of the inverter (e.g., load level, control mode, etc.). Progressive modulation primarily refers to the fact that operator weights do not jump instantaneously from one state to another, but instead evolve continuously over time along a smooth trajectory governed by physical laws, simulating the inertia of a real physical system.

[0122] For example, at least one electrical parameter representing the system's operating load and control state is extracted from real-time time series data and encoded to generate a state vector. Specifically, the effective value and power factor are extracted from each phase current, and the modulation index is extracted from the control signal. These three parameters are input into a small multilayer perceptron (MLP) and encoded into a 128-dimensional state vector s(t).

[0123] By utilizing the working condition-aware weight modulation function, the inherent weights of the physical information neural operator are modulated in real time according to the working condition state vector to generate a dynamic health system operator that can adapt to the current operating conditions.

[0124] Preferably, the modulation of the inherent weights is gradual. Specifically, when a change in the operating state vector is detected, the weights are gradually adjusted over time, subject to a preset physical constraint that limits the rate of change of the weights, until a dynamic healthy system operator matching the current operating conditions is generated. This physical constraint can be dynamically set based on the ratio of the system's total energy storage to its rated energy storage, ensuring that the rate of weight adjustment is consistent with the system's physical inertia.

[0125] Consider a typical operating condition sudden change scenario where the load of the CHB inverter suddenly increases from a light load of 20% to a heavy load of 90% within 0.5 seconds.

[0126] Since the static operator learns the physical laws under a fixed working condition, when the working condition changes suddenly, its predicted health output y predicted_healthy (t) will seriously deviate from the actual system output y realtime (t), thus generating a pseudo residual peak with a large amplitude lasting for hundreds of milliseconds in the residual function ε(t). The energy of this peak far exceeds the conventional fault alarm threshold, which can easily cause the system to issue a false fault alarm.

[0127] Using the dynamic operator of this embodiment, the operating state vector s(t) captures the load changes in real time. The weight modulation function starts a progressive modulation and smoothly adjusts the operator weights under physical constraints. Therefore, the predicted adaptive health output y predicted_adaptive (t) can closely follow the actual system output y under time-varying conditionsrealtime Experimental data show that during this sudden change in operating conditions, the amplitude of the residual function ε(t) generated by this embodiment always remains at a normal noise level, and the energy does not exceed 20% of the alarm threshold, successfully avoiding false alarms.

[0128] This step ultimately generates a real-time system residual function with a clean background that can effectively suppress interference from working condition changes.

[0129] Step S302: Decomposing the real-time system residual function.

[0130] In this embodiment, to independently verify the effect of the dynamic operator, the residual function generated in step S301 is decomposed using the single-layer iterative decomposition method of embodiment 1. The specific decomposition process is the same as step S300 of embodiment 1 and its refined steps S310 to S350.

[0131] It can be seen from this embodiment that by introducing a dynamic health system operator that is adaptive to working conditions, the industry problem of variable working condition interference can be effectively solved, a large number of pseudo residuals can be eliminated from the root, and the robustness of the diagnostic system and its usability in actual industrial environments can be improved.

[0132] Example 4: This example is used to describe the decomposition strategy of the residual function of the real-time system, reflecting the use process of the multi-level topology constraint decomposition architecture. It mainly includes the following steps:

[0133] Step S401: Generate a real-time system residual function. In this embodiment, to independently verify the effect of the decomposition architecture, the static PINO health system operator of the first embodiment (ie, without the working condition sensing process) is used to generate the real-time system residual function ε(t).

[0134] The residual function ε(t) generated in this step may contain pseudo residuals caused by operating condition changes and complex features caused by multi-fault coupling. It will serve as the input of the subsequent advanced decomposition architecture.

[0135] Step S402: Decomposing the real-time system residual function using a three-layer cascade operator decomposition architecture.

[0136] This decomposition architecture abandons the traditional single-layer, linear superposition assumption and achieves refined analysis of complex residuals through systematic hierarchical decoupling.

[0137] The three-layer cascade operator decomposition architecture adopts a structured decomposition process, which includes the operating condition pseudo-fault decomposition layer, the real fault decomposition layer and the fault coupling effect decomposition layer.

[0138] The physical topology constraint matching strategy is an improved similarity matching method. Its main feature is that, while calculating the function similarity, a weight factor is introduced to quantify the matching degree between the residual and the physical location of the candidate fault in the inverter circuit topology.

[0139] Residual Purification and First-Level Decomposition: Before formal decomposition, a residual purification process can be performed. Specifically, the operating condition-related pseudo-residuals caused by operating condition changes are separated from the original residual function and their effects are subtracted from the original residual function. This constitutes the main function of the operating condition pseudo-fault decomposition layer.

[0140] Second-layer decomposition and topology constraints: At the true fault decomposition layer, the cleansed residuals are iteratively decomposed to identify single faults. During this process, a physical topology constraint matching strategy is employed. For example, if a feature in the residual function primarily affects the output voltage of cell 3 on phase A, the matching will assign a higher matching weight to fault operators located in cell 3 on phase A in the fault library (such as IGBT faults or capacitor faults in that cell).

[0141] Third-layer decomposition: In the fault coupling effect decomposition layer, candidate coupling effect operators are determined based on the single real physical fault component separated and identified in the real fault decomposition layer, and the remaining residual after the second-layer decomposition is finally decomposed.

[0142] Assume a complex scenario with multiple concurrent faults: the IGBT of phase A unit 2 experiences slight performance degradation (weak fault) and the DC bus capacitor of phase B unit 1 experiences significant aging (strong fault) at the same time.

[0143] If the single-layer decomposition method of Example 1 is used, the capacitor aging fault signal is strong and the IGBT degradation signal is weak. Therefore, the single-layer decomposition method will match and remove the strong capacitor aging fault component in the first iteration. In the remaining residual, the weak IGBT fault signature may be contaminated by noise or coupling effects, causing it to be missed in subsequent iterations or its severity to be seriously underestimated.

[0144] In this embodiment, the first layer first filters out the possible impact of operating condition fluctuations.

[0145] During the second-level decomposition, the strong capacitor aging fault was identified first. After updating the residual, although the IGBT fault signal was weak, the physical topology constraint matching strategy played a key role: the remaining disturbance in the residual was mainly concentrated in the topological location of cell 2 in phase A. Therefore, the IGBT degradation operator at this location received extremely high matching priority, allowing it to be accurately identified.

[0146] Further analysis at the third level may identify the small nonlinear coupling effect caused by the two faults acting together on the DC bus voltage.

[0147] Experimental data shows that this embodiment can successfully decouple the aforementioned strong and weak concurrent faults, achieving 100% accuracy in locating both faults and a severity estimation error of less than 15%. The comparative approach, however, completely missed the weak IGBT degradation fault.

[0148] By introducing an innovative multi-level topologically constrained decomposition architecture, we can achieve accurate decoupling and traceability of complex multi-fault residuals with coupling effects when the benchmark model is relatively simple, effectively solving the bottleneck problems of fault masking and nonlinear superposition faced by traditional decomposition methods.

[0149] Example 5. This example provides a complete numerical case of applying the method of the present invention to perform multi-fault diagnosis in a CHB inverter system with specific parameter settings, aiming to quantitatively illustrate the specific implementation details and technical effects of the present invention.

[0150] A 7-level cascaded H-bridge (CHB) three-phase inverter is used as the experimental object.

[0151] The specific CHB system parameters are as follows:

[0152] Topology: Each phase consists of three H-bridge power units in cascade.

[0153] DC side voltage: The DC side voltage of each power unit is supplied by an independent 300V DC source.

[0154] Power semiconductors: Each H-bridge uses an IGBT module, and the switching frequency is set to 2kHz.

[0155] Load: three-phase star-connected RL load, R=10Ω, L=20mH.

[0156] The multimodal data synchronization mechanism is as follows: the data acquisition scheme is the same as that in the first embodiment, and a system based on FPGA and PTP protocol is used.

[0157] Signals from different channels are resampled using a linear interpolation algorithm to align their timestamps.

[0158] When splitting the data window, the window length T is 20ms. This length is chosen so that it can completely cover a 50Hz power frequency cycle, facilitating periodic signal processing such as Fourier analysis. The data window overlap is 50%.

[0159] Build the PINO-H health system operator, including:

[0160] Build a physical topology-aware architecture: Model the three power units of each phase as a graph containing three nodes.

[0161] BranchNet: This model processes the input function u(t) (i.e., the SPWM control signal sequence for three units). Its network structure consists of a two-layer graph convolutional network (GCN). Each GCN layer has 128 hidden units. The ReLU function is used as the activation function.

[0162] TrunkNet: Used to process query coordinates (time t and output physical quantity type). Its network structure is a 4-layer fully connected network (MLP) with an output dimension of 128.

[0163] For example, consider the following input and output dimensions: In a 20ms data window, processed at a 10kHz frequency (downsampled from the original 10MHz), the time series length is 200 points. The input function u(t) has dimensions (200, 3), and the output function y(t) (the output voltages of three units) also has dimensions (200, 3).

[0164] The specific mathematical expression of the dual loss function is:

[0165] Total loss function L total =L data +λL phys .

[0166] Data-driven loss L data : The mean square error (MSE) loss is used to calculate the difference between the output voltage predicted by PINO-H and the actual output voltage in the healthy state generated by simulation.

[0167] Physical law residual loss L phys =(1 / N)∑ N i=1 ∫ t (R i (t)) 2 dt, where N is the number of units (N=3 in this example).

[0168] CHB system of differential equations: R i (t) is the residual term based on Kirchhoff’s current law (KCL). For the DC link capacitor C of the i-th unit, its voltage u dc_i The changing rule of C (du dc_i (t) / dt)=i in_i (t)-i out_i (t).

[0169] Therefore, the physical residual term is defined as: R i (t)=(du dc_i (t) / dt)-i in_i (t)+i out_i (t). In the ideal model, R i(t) should always be equal to 0. This calculation is performed through the backpropagation and automatic differentiation functions of the network.

[0170] The hyperparameter λ was determined by grid search on an independent validation set to balance data fitting accuracy and compliance with physical laws. Its typical value range is [0.01, 0.5]. In this example, experiments determined that a λ value of 0.1 was the best.

[0171] In the present invention, the residual operator ΔL f Operator L is used to characterize the effect of a specific fault f on the physical laws of a healthy system healthy The disturbance caused by the real system with faults, its physical law operator L real Can be modeled as L real =L healthy +ΔL f We observe the result after it acts on the input u(t), that is, the residual function ε(t)=(L real -L healthy )(u(t))=ΔL f (u(t)), to reversely infer ΔL f type, thereby identifying the fault f.

[0172] For example, consider building a fault operator for an open circuit in the upper IGBT Q1 of unit 2. After injecting this fault into the simulation, a dedicated PINO model is trained using the corresponding fault dataset. Training is considered converged when the model's loss function value on the validation set does not decrease for five consecutive training cycles (epochs), or decreases by less than 1e-5. The operator model at this point is saved.

[0173] The OMP algorithm process is described as follows: Assume that at a certain moment, two faults occur simultaneously in the inverter: F1: 'IGBTQ1 open circuit in the upper bridge arm of unit 2' (strong fault) and F2: 'ESR increase of the DC capacitor of unit 1 by 30%' (weak fault).

[0174] In discrete time series, the inner product of the function is calculated by the vector dot product: (A(t), B(t)) ≈∑ n=1 200 A[n]·B[n].

[0175] Set two termination conditions, and satisfy one of them:

[0176] The ratio of the current residual energy to the initial residual energy is less than a preset threshold η. In this embodiment, η is set to 2%.

[0177] To prevent infinite loops, a maximum number of iterations Kmax is set. In this embodiment, assuming that the number of concurrent failures does not exceed 3, K maxSet to 5.

[0178] Residual generation: The system generates a hybrid residual function ε(t) that contains two fault characteristics.

[0179] Matching is performed in the fault library and it is found that the candidate residual ε* generated by the F1 fault operator F1 (t) has the highest normalized similarity with the current residual R0(t)=ε(t), which is 0.92. The fault activation coefficient α is calculated F1 =0.95. Update residual R1(t)=R0(t)-0.95·ε* F1 (t).

[0180] Matching is performed again on the new residual R1(t), and it is found that the candidate residual ε generated by the F2 fault operator F2 (t) has the highest normalized similarity with R1(t), which is 0.75. The fault activation coefficient α is calculated as F2 =0.28. Updated residual R2(t)=R1(t)-0.28·ε* F2 (t).

[0181] The energy of R2(t) is calculated and the ratio of its energy to the energy of the initial residual ε(t) is found to be 1.6%, which is lower than the set threshold η = 2%. The iteration is terminated.

[0182] The system outputs the final diagnostic result list: [{fault: "IGBT open circuit", location: "unit 2 upper bridge arm Q1", activation coefficient: 0.95}, {fault: "capacitor ESR increased", location: "unit 1", activation coefficient: 0.28}].

[0183] This embodiment accurately identifies the types and locations of the two concurrent faults, and the activation coefficient (which can be considered a quantitative indicator of severity) is consistent with the intensity of the injected fault (an open circuit fault is a complete fault with a coefficient close to 1; a 30% increase in ESR is an early degradation fault with a smaller coefficient).

[0184] This embodiment can illustrate the diagnostic accuracy and reliability in multi-fault and strong-weak coupling scenarios, and solves the problem of insufficient disclosure in existing technical solutions.

[0185] According to one aspect of the present application, the operating state vector s(t) is a fixed-dimensional real number vector that can compactly and digitally represent the macroscopic operating conditions (such as load level, control state, etc.) of the inverter at time t.

[0186] Extract multiple physical parameters that characterize the current operating conditions from real-time time series data. In this embodiment, the effective value of the load current, I_rms, the power factor, cos(φ), and the modulation index, m, of the control signal are extracted. These three scalar parameters are concatenated into a three-dimensional input vector. A pre-trained multilayer perceptron (MLP) is used to encode this three-dimensional input vector to generate a 128-dimensional operating state vector, s(t). The specific network structure of this MLP is: an input layer (3 neurons), two hidden layers (containing 64 and 128 neurons, respectively, with ReLU activation function), and an output layer (128 neurons, with no activation function or a linear activation function).

[0187] Through nonlinear mapping (MLP encoding), discrete operating parameters with varying physical meanings are converted into a representation in a continuous vector space. This allows the health system operator model to smoothly and continuously adjust its weights based on this vector to adapt to dynamic changes in operating conditions, effectively suppressing spurious residuals caused by sudden changes in operating conditions.

[0188] Optionally, the neural network used for encoding is not limited to MLPs; it can also be other network structures capable of feature extraction and dimensionality reduction / increase, such as radial basis function networks (RBFNs) or the encoding portion of a small autoencoder. Furthermore, the input operating condition characteristic parameters are not limited to the three aforementioned ones; depending on the specific application scenario, they can be supplemented or replaced with other parameters that reflect the system's operating status, such as the average DC voltage, switching frequency, and active / reactive power commands.

[0189] In a specific system, when the load current is 80A RMS, the power factor is 0.9, and the modulation index is 0.85, the corresponding 3D input vector is [80, 0.9, 0.85] (normalized and fed into the MLP). After MLP encoding, a 128-dimensional specific vector s_t is generated.

[0190] Progressive modulation is used to describe the adjustment process in which the internal weight W of the health system operator model does not jump instantaneously when the working conditions change, but evolves continuously over time on a smooth trajectory constrained by physical laws.

[0191] In other words, when a change in the operating state vector s(t) is detected, the weight update is initiated. The rate of change of the weight W, dW(s(t)) / dt, is limited by the following physical constraint: |dW(s(t)) / dt|≤β·(E _rated / E _stored (t)). Where, E _stored (t) is the total instantaneous energy storage of the inverter at time t, which is mainly achieved by storing the energy of the DC side capacitors of all power units (Σ(0.5*C i *u _dc_i (t)2 )) to calculate the sum. E _rated is the total energy storage of the inverter under rated operating conditions and is a fixed design parameter. β is a dimensionless hyperparameter used to adjust the modulation rate.

[0192] The physical significance of the constraint is that the rate of change of the model's parameters (as a mathematical mirror of the system) should not exceed the rate of change of the physical system's energy state. This simulates the inertia of real power electronics systems, ensuring smooth, stable, and physically consistent adjustments to the health baseline. This prevents misinterpreting normal transient responses as faults.

[0193] Optionally, the form of the constraint function can be adjusted. For example, a more complex function related to the power change rate can be used. The hyperparameter β can be a fixed value or a dynamic value that is adaptively adjusted according to the system temperature or other state parameters. The typical value range of β is [0.1, 2.0]. A smaller β value results in a smoother modulation process, which is suitable for systems with large inertia; a larger β value results in a faster response, which is suitable for scenarios that require rapid tracking of operating conditions. In a specific project, through offline simulation tests, it was determined that β=0.5 can achieve the best balance between suppressing pseudo-residuals and quickly adapting to operating conditions.

[0194] The physical topology constraint matching strategy refers to a method that introduces a weighting factor to quantify the degree of matching between the physical position and influence range of the two in the inverter circuit topology when calculating the similarity between the current residual function and the candidate residual function, thereby correcting the purely mathematical similarity calculation results. The specific process is as follows:

[0195] A fault-measurement impact matrix M is constructed offline.

[0196] The matrix is ​​N _m ×N _f Dimension, where N _m is the total number of system measurement points (such as each unit output voltage, bridge arm current, etc.), N _f is the total number of faults in the basic fault library. If the jth fault has a direct or strong correlation effect on the i-th measurement point, the matrix element M(i,j) = 1, otherwise it is 0. Step 2: When performing similarity matching online, first calculate the inner product (or correlation coefficient) of the current residual function R_k(t) and the candidate residual ε*_j(t) generated by the jth basic fault feature operator in the traditional way, denoted as Sim _math. Step 3: Calculate the topological matching degree Sim_topo. Specifically, the indexes of the measurement points with significant energy (for example, the amplitude exceeds the noise threshold) in R_k(t) and ε*_j(t) are used to form two sets S_R and S_ε. The topological matching degree can be calculated as Sim_topo=|S_R∩S_ε| / |S_R∪S_ε|(Jaccard similarity). Step 4: The final similarity score is Sim _final =Sim _math ×Sim_topo. Select Sim _f The fault with the highest inal value is taken as the recognition result of the current iteration.

[0197] This strategy incorporates prior knowledge of the circuit's physical structure into the decomposition algorithm. It addresses the long-range mismatching problem that can occur with traditional methods. For example, a fault on phase A might produce a residual waveform that is mathematically similar to the waveform of a different type of fault on phase B, even though their physical origins are completely different. By using topological constraints, only when the physical influence ranges match are the two considered truly correlated, significantly improving fault location accuracy.

[0198] The influence matrix M can be weighted rather than binary, with the weight indicating the degree of influence. The topological matching degree Sim_topo can also be calculated in other ways, such as using the distance between nodes in graph theory (electrical distance on a circuit diagram). The final similarity fusion method can be multiplication or weighted summation w1*Sim _math +w2*Sim_topo.

[0199] Assume that the residual error is mainly reflected in the output voltage of the second unit of phase A, and there are two faults with high mathematical similarity in the candidate fault library: the IGBT fault of the second unit of phase A (Sim _math =0.9) and the first unit capacitor fault of phase B (Sim _math =0.88). The topological match Sim_topo of the former is 0.95, while the Sim_topo of the latter is 0.1 due to the position mismatch. The final scores are 0.9*0.95=0.855 and 0.88*0.1=0.088, respectively. Therefore, the system will accurately select the former.

[0200] Fault coupling effect refers to the nonlinear additional system disturbance generated by the interaction of two or more single physical faults through the physical loop of the circuit when two or more single physical faults occur simultaneously, which cannot be characterized by the simple linear superposition of the characteristics of each single fault.

[0201] The mechanism is initiated after the real fault decomposition layer is completed and processes the second-layer decomposition residual R_2(t). Step 1: Based on the list of single faults identified by the real fault decomposition layer (e.g., faults F_i and F_j), determine the candidate coupling effect types.

[0202] From a pre-built coupled fault signature library, find the _i ,F _j ) Combine the corresponding pre-trained coupling effect feature operator F _ij_couple If found, the candidate residual generated by it is matched with R_2(t).

[0203] If no corresponding pre-trained operator is found in the library (for unforeseen coupling combinations), an online approximate synthesis method is used. Specifically, by _i and F _j The corresponding basic fault feature operators are combined in some form of nonlinear combination to generate an approximate coupled candidate residual.

[0204] For example, ε* _couple (t)≈c _1 *ε* _i (t)+c _2 *ε* _j (t)+c _3 *(ε* _i (t).*ε* _j (t)), where * represents point-by-point multiplication and c is the coefficient to be determined.

[0205] The coupling effect component with the highest matching degree is removed from R _2 Subtract (t).

[0206] This embodiment focuses on the complex and potentially infinite coupling fault problem by performing combined analysis based on a few identified single faults, thus narrowing the search space. By combining offline pre-training with online approximate synthesis, it balances diagnostic accuracy for common coupling modes with coverage for rare coupling modes, enabling analysis of nonlinear aliasing fault characteristics that traditional methods cannot handle.

[0207] The construction of the coupled fault signature library is not limited to PINO and can also be other generative models. The online approximate synthesis method can be more complex, such as using a neural network to learn the mapping from two single fault residuals to their coupled residuals.

[0208] When the energy of the second-level decomposition residual is still higher than 5% of the initial residual energy, coupling effect decomposition is initiated. In the scenario of Example 5, after identifying the IGBT open circuit and the capacitor ESR increase, the system specifically searches for the coupling operator between the two and matches and decomposes the remaining 1.6% energy residual to identify any small nonlinear coupling components.

[0209] By extracting electrical parameters representing the system's operating load and control state from real-time time series data and encoding them into a state vector, this vector is then used to modulate the inherent weights of the healthy system operator in real time. This method enables the diagnostic system's healthy physical law benchmark to dynamically and adaptively adapt to changes in the drive's actual operating conditions. The principle behind this adaptive benchmark is the deep coupling of abstract operating conditions (input parameters) with specific operator models (algorithm features), transforming the operator from a static mapping function into a dynamic function that reflects time-varying physical laws. This fundamentally addresses the technical challenge of traditional static models, which generate significant spurious residuals due to their inability to distinguish between normal system transient responses and true fault characteristics when faced with operating conditions such as sudden load changes and grid voltage fluctuations. Consequently, this method effectively suppresses false alarms caused by varying operating conditions, significantly improving the robustness and reliability of the diagnostic system in non-ideal, highly dynamic industrial environments, and ensuring the accuracy of diagnostic results.

[0210] By gradually, rather than instantaneously, adjusting operator weights upon detecting a change in the operating state vector, within pre-set physical constraints that limit the rate of weight change, the dynamic health benchmark adjustment process is smooth, stable, and consistent with the inertia of the physical system. This approach is implemented by linking the abstract algorithmic weight update process with specific physical domain knowledge, such as the electromagnetic energy storage and response inertia determined by the large number of capacitors and inductors in the CHB system. Specifically, the weight adjustment rate is limited by the system's current energy storage state. This avoids quadratic residuals caused by mathematical overshoot or oscillation during rapid weight modulation, ensuring that the diagnostic system maintains high-precision health behavior predictions throughout the transient period of drastic operating condition switching. This not only further enhances the ability to suppress variable operating condition disturbances, but also improves the numerical stability and physical interpretability of the entire diagnostic model, making the adaptive process itself more reliable.

[0211] By employing a three-tiered decomposition architecture that structures the residual decomposition process into three cascaded layers: the operating-condition pseudo-fault decomposition layer, the real-physical fault decomposition layer, and the fault coupling effect decomposition layer, a divide-and-conquer approach and refined analysis of complex mixed residuals is achieved. This architecture decomposes a complex, multi-objective optimization problem into three sub-problems with clear physical meaning and sequential logic. First, the operating-condition pseudo-fault decomposition layer (with raw residuals as input and purified residuals as output) filters out interference unrelated to the faults. Then, the real-fault decomposition layer (with purified residuals as input and a list of single faults as output) identifies the primary fault sources. Finally, the fault coupling effect decomposition layer (with quadratic residuals as input and coupled effects as output) identifies the nonlinear superposition effects caused by multiple concurrent faults. This systematic decoupling strategy overcomes the technical bottlenecks of traditional single-layer decomposition methods, which struggle to distinguish residual components from different sources and handle nonlinear superposition effects. It significantly improves diagnostic clarity and accuracy in concurrent multi-fault scenarios and provides deep insight into the interactions between faults.

[0212] By introducing a topology matching function, which quantifies the degree of match between the residual's physical location and impact range in the inverter circuit topology, as a weighting factor in the similarity calculation, the decomposition process adds a crucial physical space constraint. This method combines the abstract similarity calculation (algorithmic characteristics) with the specific circuit topology (technical characteristics). In the specific technical field of CHB inverter diagnosis, an IGBT fault occurring in the second unit of phase A will generate residual disturbances that are primarily reflected in the voltage and current channels associated with that unit. The topology matching function leverages this prior knowledge, enabling the decomposition algorithm to prioritize fault sources with more reasonable physical locations when searching the fault database. This resolves the diagnostic ambiguity caused by the potential similarity in signal morphology between faults of different locations and types. In particular, in the scenario of concurrent strong and weak faults, the method significantly improves the detection sensitivity of weak faults that are masked by strong fault signals, enhancing fault location accuracy and early warning capabilities.

[0213] The PINO model limits the update rate of operator weights through a mathematical formula constrained by the system's real-time energy storage, enabling the abstract neural network model to simulate the electromagnetic inertia of a real power electronic system, thereby eliminating the possibility of transient responses of operating conditions being misjudged as faults.

[0214] OMP introduces a circuit topology influence matrix to guide the similarity matching process, transforming the purely mathematical signal matching problem into a physically interpretable causal reasoning process. This solves the ambiguity and misjudgment inherent in conventional decomposition algorithms when faced with faults that have similar signals but different physical locations. It significantly improves the ability to detect subtle faults in the presence of strong background noise.

[0215] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A method for identifying multiple fault sources of a cascaded H-bridge inverter, characterized in that: include: Get real-time time series data of the inverter online; Using pre-trained healthy system operators and real-time time series data, a real-time system residual function is generated to characterize the deviation of the actual operating behavior of the purifier from the healthy physical laws; Based on a preset basic fault operator library containing multiple single fault modes or a three-layer cascade operator decomposition architecture that takes into account physical topology constraints, the real-time system residual function is decomposed to obtain fault identification results; Among them, the basic fault operator library is used to decompose the residual function of the real-time system, including: Extracting at least part of the data from the real-time time series data as a real-time system input function, inputting the data into each basic operator in the basic fault operator library, and generating a candidate residual function; By performing similarity matching between the candidate residual function and the current residual function, the best matching basic operator is selected from the basic fault operator library as the fault component identified in the current iteration; Update the current residual function for the next iteration until the energy of the current residual function is lower than the preset value, and the decomposition is completed; Alternatively, a three-layer cascaded operator decomposition architecture is used to decompose the real-time system residual function, including: The operating condition pseudo-fault decomposition layer separates and identifies the pseudo-fault components caused by drastic changes in operating conditions from the initial residual function and outputs the first-layer decomposition residual; The real fault decomposition layer separates and identifies the single real physical fault component from the first-layer decomposition residual and outputs the second-layer decomposition residual; The fault coupling effect decomposition layer separates and identifies the nonlinear coupling effects caused by the concurrent occurrence of multiple real physical faults from the second-layer decomposition residuals; Similarity matching in a three-layer cascade operator decomposition architecture uses a physical topology constraint matching strategy; When calculating the similarity between the current residual function and the candidate residual function, the physical topology constraint matching strategy introduces a topology matching function that quantifies the degree of matching between the physical positions and influence ranges of the two in the inverter circuit topology as a weight factor for the similarity calculation.

2. The method according to claim 1, characterized in that The process of identifying the fault component and updating the current residual function includes: Calculate the projection of the current residual function on the candidate residual function, and determine the fault activation coefficient based on it to quantify the contribution of the fault component identified in the current iteration; Determine the fault component through the fault activation coefficient and the best matching basic operator; The fault component is subtracted from the current residual function to generate an updated current residual function for the next iteration.

3. The method according to claim 1, characterized in that The health system operator is a physical information neural operator configured with a physical topology-aware architecture; Each power unit of the inverter is a node of the graph, and the electrical connections between the power units are edges of the graph, constructing the physical graph of the system; The features of the nodes in the system physical graph are processed through the graph neural network layer to simulate the transmission effect of physical quantities in the physical structure of the inverter.

4. The method according to claim 3, characterized in that When working, the physical information neural operator includes the working condition perception process, specifically: Extract multiple characteristic parameters representing the current operating conditions from real-time time series data and encode them into an operating condition state vector; By utilizing the weight modulation function of working condition perception, the inherent weight of the physical information neural operator is modulated in real time according to the working condition state vector to generate a dynamic health system operator that can adapt to the current operating conditions.

5. The method according to claim 1, characterized in that Before decomposing the residual function of the real-time system, the residual purification process is also included, specifically: Separate the operating condition-related pseudo residuals caused by operating condition changes from the original residual function, and deduct the influence of the operating condition-related pseudo residuals to obtain a purer fault-related true residual for subsequent decomposition; Among them, the effects of the pseudo residuals related to the working conditions are deducted from the original residual function, specifically including: Matching the working condition-related pseudo residuals with a pre-trained library of pseudo fault signatures to identify and quantify the working condition switching effects and generate pseudo fault compensation vectors. And deduct the pseudo fault compensation vector from the original residual function.

6. The method according to claim 1, characterized in that In the three-layer cascade operator decomposition architecture, the decomposition process of the fault coupling effect decomposition layer is based on the single real physical fault component that has been separated and identified in the real fault decomposition layer to determine the candidate coupling effect operators and match and decompose the second-layer decomposition residuals.

7. The method according to claim 4, characterized in that The generation of the operating state vector includes: Extracting at least one electrical quantity parameter representing the system operating load and control state from the real-time time series data; The extracted electrical quantity parameters are encoded to generate an operating state vector; The modulation of the intrinsic weight is a gradual modulation, including: When a change in the operating state vector is detected, the weights are gradually adjusted over time under preset physical constraints that limit the rate of change of the weights until a dynamic health system operator matching the current operating condition is generated.

Images