Self-healing fault reconstruction control method for cascaded H-bridge frequency converter

Abstract

The invention discloses a self-healing fault reconstruction control method for a cascaded H-bridge frequency converter, and the method comprises the steps: analyzing original magnetic field intensity data collected by a magnetic field sensor array, and generating fault prediction information containing fault types and positions based on the spatial distribution characteristics of the original magnetic field intensity data; executing transient suppression operation based on the prediction information, and determining suppressed system state parameters; and performing topology reconstruction decision according to the state parameters, and generating and outputting a synchronous control instruction set. According to the method, the fault is predicted in advance through the magnetic field characteristics, and refined transient energy absorption, adaptive damping control and active phase pre-compensation reconstruction are combined, so that the problems of fault detection lag, large transient impact and serious reconstruction circulating current in the prior art are solved, and the reliability and continuity of system operation are improved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of power electronics and power transmission, and in particular to a self-recovery fault reconstruction control method for a cascaded H-bridge frequency converter. BACKGROUND

[0002] As a core device in the field of high-voltage and high-power frequency conversion, the cascaded H-bridge frequency converter plays an irreplaceable role in key fields such as new energy grid connection, industrial transmission and flexible DC power transmission, due to the high voltage adaptability, low harmonic output characteristics and low pressure bearing advantage of the modular topology structure. The system reliability directly determines the continuity of industrial production and the stability of power grid operation, so the development of self-recovery control technology that can realize continuous operation under fault conditions is a core technical requirement to improve the operation safety of the equipment.

[0003] The current fault handling technology of the cascaded H-bridge frequency converter mainly follows the traditional path of detection-isolation-reconstruction. The fault detection relies on current threshold monitoring, capacitor voltage deviation judgment or harmonic characteristic analysis, etc. The fault isolation cuts off the faulty unit from the main circuit through mechanical switches or semiconductor bypass devices. The topology reconstruction maintains the basic operation ability of the system by enabling redundant units or adjusting the PWM modulation strategy of the remaining healthy units, such as recalculating the carrier phase shift angle to rebalance the three-phase output voltage, and using a phase-locked loop (PLL) circuit to re-synchronize with the grid or load.

[0004] However, the passive response mechanism of the prior art is difficult to adapt to the microsecond fault characteristics of the power electronic system: the fault detection relies on the manifestation of the consequences of the fault, and there is a response blind area of tens to hundreds of microseconds, which easily leads to irreversible damage to the device; the transient energy management adopts a one-size-fits-all hard protection, which cannot dynamically match the damping demand according to the fault characteristics, and the voltage and current impact caused by energy release threatens the safety of healthy devices; the topology reconstruction lacks a phase pre-coordination mechanism, and the phase step deviation after fault isolation needs to rely on the passive tracking compensation of the phase-locked loop, during which the circulating current and impact current seriously affect the power quality. SUMMARY

[0005] The application provides a self-recovery fault reconstruction control method for a cascaded H-bridge frequency converter. The purpose is to solve the core technical defects in the prior art, such as the existence of a response blind area in fault detection which easily leads to irreversible damage to the device, the use of a one-size-fits-all hard protection in transient energy management which cannot dynamically match the damping demand, and the lack of a phase pre-coordination mechanism in topology reconstruction which leads to an impact on power quality, to realize active prediction, precise suppression and disturbance-free reconstruction of faults, and to improve the operation safety, continuity and power quality stability of the cascaded H-bridge frequency converter under fault conditions.

[0006] Technical solution: According to one aspect of the present application, it comprises: analyzing the original magnetic field intensity data collected by the magnetic field sensor array, generating fault prediction information containing fault type and location based on the spatial distribution characteristics of the original magnetic field intensity data; based on the fault prediction information, performing transient suppression operation, determining the system state parameters after suppression; according to the system state parameters after suppression, making topology reconstruction decision, generating and outputting synchronization control instruction set to the power unit of the frequency converter.

[0007] Beneficial effects: The present application collects original magnetic field intensity data by magnetic field sensor array and analyzes the spatial distribution characteristics to generate fault prediction information, breaking through the traditional passive detection mode relying on fault consequences, shortening the fault recognition window from tens to hundreds of microseconds to microseconds, effectively eliminating the response blind area, reducing the risk of irreversible damage of devices caused by over-stress; the transient suppression operation based on the fault prediction information realizes fine management of transient energy by dynamically adjusting the damping characteristics of power switching devices and sub-capacitance time-sharing collaborative energy absorption, avoiding the voltage and current impact caused by one-size-fits-all hard protection, and ensuring the safe operation of healthy devices; the topology reconstruction decision according to the system state parameters after suppression replaces the traditional passive tracking mode relying on phase-locked loop, eliminates the phase step deviation and circulating current after fault isolation, and improves the power quality stability during topology reconstruction process. Through the whole process of fault prediction, dynamic suppression and topology reconstruction, a complete closed-loop control system from early warning to self-healing is constructed, and the fault tolerance and operation reliability of cascaded H-bridge frequency converter are improved. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 The flowchart of the self-healing fault reconstruction control method of the cascaded H-bridge frequency converter provided by the embodiment of the present application.

[0009] Figure 2 The flowchart of analyzing original magnetic field intensity data to generate fault prediction information provided by the embodiment of the present application.

[0010] Figure 3 The flowchart provided by the embodiment of the present application before calculating the second-order spatial derivative.

[0011] Figure 4 The flowchart of identifying abnormal curvature distribution pattern in magnetic field curvature matrix to determine fault type and location provided by the embodiment of the present application.

[0012] Figure 5 The flowchart of controlling power switching devices provided by the embodiment of the present application. DETAILED DESCRIPTION

[0013] In order to better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts should fall within the scope of the present application.

[0014] It should be noted that the terms first, second, etc. in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms include and have and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to the clearly listed steps or units, but can include other steps or units that are not clearly listed or inherent to the process, method, product or device.

[0015] In order to solve the above problems, the applicant has conducted in-depth search and analysis, and found that: There is inevitable post-delay in fault detection, resulting in lag of protection action. The existing method relies on monitoring of fault consequences (such as overcurrent, undervoltage), and these electrical characteristics will only become significant after the fault has occurred and developed to a certain extent. In the blind area of tens or even hundreds of microseconds from the beginning of the fault to being detected, a huge short-circuit current may have caused irreversible damage to the power device. This detection cannot provide early warning and intervention at the initial stage of fault occurrence, or even before the occurrence, missing the best opportunity to actively suppress the development of the fault.

[0016] Furthermore, there is a lack of fine and adaptive management means for fault transient energy. After detecting the fault, the existing technology usually adopts a rough protection strategy, such as forming a hard short circuit through a crowbar circuit to absorb energy, or directly bypassing the fault unit. This one-size-fits-all approach cannot provide matching suppression measures according to the type, severity and development speed of the fault. The system lacks a mechanism that can provide precise damping as needed, like a variable resistor, to actively dissipate transient energy. Therefore, at the moment of fault isolation, the energy stored in the system stray inductance and capacitance will be released in the form of severe oscillation, causing serious voltage and current surges. This secondary impact not only threatens the safety of other healthy devices, but also seriously worsens the electromagnetic compatibility of the system.

[0017] On this basis, the topology reconfiguration process lacks active phase coordination, resulting in serious switching impact and circulating current problems. After the faulty unit is bypassed, the topology structure and electrical parameters of the system change abruptly, which will inevitably cause a step deviation between the output phase of the remaining healthy unit and the target phase of the system. The existing technology relies on a phase-locked loop to passively track and compensate for this phase difference after topology switching, but the inherent response bandwidth limitation of the phase-locked loop requires hundreds of microseconds or even milliseconds to relock. During this step-out time, the huge phase difference will produce serious inter-phase circulating current or impact current, which may trigger the upper-level protection to cause system outage and cause serious pollution to power quality. Since the existing technology cannot predict the phase state after switching and actively pre-compensate before the switching action occurs, to achieve seamless phase connection, the following scheme is provided: In combination Figures 1 to 5 The scheme of the present application is specifically described.

[0018] Embodiment one provides a self-healing fault reconfiguration control method for a cascaded H-bridge frequency converter, which overturns the passive protection mode of traditional detection-isolation-reconfiguration by building a prediction-suppression-reconfiguration trinity control system, and realizes pre-knowledge and disturbance-free self-healing of faults.

[0019] In specific embodiments, it includes: Step S100, analyzing the original magnetic field strength data collected by the magnetic field sensor array, and generating fault prediction information containing fault type and location based on the spatial distribution characteristics of the original magnetic field strength data; In this embodiment, the magnetic field sensor array, for example, can be a 4x4 matrix of Hall sensors arranged around each H-bridge power unit in the cascaded H-bridge frequency converter, for real-time monitoring of the magnetic field environment around the power unit. The original magnetic field strength data B _raw [i,j,t _k ] is the magnetic induction intensity value collected at the sensor array (i, j) position at discrete time t _k . Specifically, the analysis process includes high-frequency noise filtering (e.g., using a moving average filter) of B _raw data, and calculation and analysis of the spatial distribution characteristics of the filtered magnetic field matrix; comparing the extracted features with the preset fault model can identify the type of fault (such as IGBT short circuit, diode breakdown or bus flashover) and determine the physical location coordinates of the fault.

[0020] This step can capture early signs of failure within a time window of about 15-20 microseconds before the fault current or voltage appears significantly abnormal, which gains valuable time for subsequent suppression and reconfiguration.

[0021] Optionally, the specific analysis method of the spatial distribution characteristics will be further elaborated in Embodiment II and Embodiment III.

[0022] Step S200, based on the fault prediction information, performing a transient suppression operation, and determining a post-suppression system state parameter after the transient suppression operation is completed; Wherein, the transient suppression operation is used to actively intervene in the initial stage of the fault (for example, within 0-100μs), dissipate or transfer the transient energy generated by the fault, and limit the fluctuations of system voltage and current within a safe range. Further, the operation can include two aspects: one is adaptive damping control, that is, by accurately controlling the gate voltage of the power switch device (such as MOSFET) in the fault unit or its adjacent unit, it works in the linear region or sub-threshold region, and presents a 1Ω to 1kΩ continuous adjustable adaptive damper for suppressing the drastic change of current; the other is time-sharing energy absorption, that is, a large energy storage capacitor is divided into multiple independent sub-capacitor arrays (such as 10 sub-capacitors of 10μF), and according to the energy release curve calculated by the fault prediction information, these sub-capacitors are triggered in a predetermined sequence to realize fine and phased absorption of transient energy, avoiding the secondary impact caused by single absorption. The post-suppression system state parameter contains the key electrical parameters of all healthy H-bridge units in the system after the completion of the transient suppression, such as the residual DC bus voltage V _res , the residual current I _res , and the phase offset of the output AC voltage φ _res , which provides accurate initial conditions for subsequent topology reconstruction.

[0023] Step S300, according to the post-suppression system state parameter, making a topology reconstruction decision, and generating and outputting a set of synchronous control instructions to the power unit of the frequency converter.

[0024] Specifically, the decision-making process determines the fault unit that needs to be isolated according to the fault location determined in step S200, and based on the post-suppression system state parameter, a specific prediction algorithm (such as the algorithm based on RLS and coupled dynamics model in Embodiment VI) is used to predict the phase evolution trajectory φ _predBuilding upon this, optimization algorithms (such as the PSO algorithm or a rule-based engine) are used to search among all possible backup topologies to find the optimal reconfiguration topology configuration that minimizes system switching shock and achieves the most balanced power. Furthermore, the optimal topology configuration and the timing requirements for continuous phase switching are integrated into a synchronous control instruction set, CMD_set. This instruction set contains precise timing control commands for the power switches, disconnect switches, and PWM modulators of each H-bridge unit, and is synchronously distributed to each execution unit via a high-speed communication bus (such as the AXI bus), completing the self-healing process of the entire system.

[0025] Example 2 describes the specific process of fault prediction and location based on key feature vectors of magnetic fields, which is used in scenarios with limited computing resources or extremely demanding requirements for prediction speed.

[0026] This embodiment further includes: Step S210: Analyze the original magnetic field strength data to generate fault prediction information, including: filtering the original magnetic field strength data to obtain the filtered magnetic field matrix; calculating the second-order spatial partial derivatives of the filtered magnetic field matrix along at least two spatial dimensions to construct a magnetic field curvature matrix that characterizes the curvature of the magnetic field spatial distribution; identifying abnormal curvature distribution patterns in the magnetic field curvature matrix to determine the fault type and location, thereby forming fault prediction information.

[0027] In this embodiment, to reduce computational complexity, the process of calculating its second-order spatial partial derivatives to construct the magnetic field curvature matrix characterizing the spatial distribution curvature of the magnetic field is simplified to calculating the first-order spatial gradient and extracting key feature vectors from it. Specifically, the controller is based on the filtered magnetic field matrix B. _filtered [i, j, t] _k Calculate its first-order spatial gradient to obtain the gradient feature matrix G[i,j,t]. _k Based on this, instead of performing complex two-dimensional pattern matching, the following key feature vectors with minimal computational cost are extracted from the G matrix: Peak gradient G _max :G _max =max(G[i,j,t) _k ]); where G _max The maximum value in the gradient matrix is ​​expressed in Tesla per meter (T / m), and its magnitude directly reflects the intensity of the current anomaly; the gradient center [x] _c y _c []: The centroid coordinates of the gradient, which directly indicate the approximate physical location of the fault source, in meters (m); Gradient energy E: E _g =Σ(G[i,j,t) _k ]) 2 Among them, E _gis the sum of squares of all elements of the gradient matrix, whose magnitude reflects the overall scale of the magnetic field distortion caused by the fault; peak change rate dG _max / dt: dG _max / dt ≈ (G _max (t _k )-G _max (t _k-1 )) / (t _k -t _k-1 ); wherein, dG _max / dt is the change speed of the peak gradient, with the unit of Tesla / meter / second (T / m / s), which is used to evaluate the urgency of the fault development; Step S220, identifying the abnormal curvature distribution pattern in the magnetic field curvature matrix to determine the fault type and location, includes: performing normalization processing on the magnetic field curvature matrix to eliminate the influence of fault intensity and absolute position, and extracting the normalized shape feature representing the relative spatial distribution; correlating and matching the normalized shape feature with a pre-stored fault fingerprint library containing multiple known fault patterns, and determining the fault type according to the matching result; and locating the fault position based on the numerical maximum value point in the magnetic field curvature matrix.

[0028] Among them, the process of correlating and matching with the pre-stored fault fingerprint library containing multiple known fault patterns is realized as a decision tree or a small lookup table established in advance through simulation or experiment. The decision tree takes the key feature vector extracted in step S110 as input. For example, the simplified decision rule can be: if (G _max >Th _1 AND dG _max / dt>Th _2 ) then Fault _type =IGBT short circuit; if (G _max <Th _1 AND E _g >Th _3 ) then Fault _type =bus flashover; wherein, Th _1 , Th _2 , Th _3 are preset threshold values, for example, Th _1 may be set to 3 times the maximum gradient in normal operation, such as 30 T / m 2 . The process of locating the fault position based on the numerical maximum value point in the magnetic field curvature matrix is directly realized through the coordinates of the gradient center [x _c , y _c ] calculated in step S110, which is determined as the fault position.

[0029] In the embodiment, the key features are extracted and judged by using the decision tree, avoiding the complex two-dimensional cross-correlation operation such as FFT, so that a simple microcontroller can complete the fault prediction and positioning within several microseconds, improving the response speed and reducing the hardware cost.

[0030] Optionally, a more accurate identification method will be described in Embodiment Three, which adopts complete two-dimensional cross-correlation operation and can more strictly correspond to the technical features of the association matching with the pre-stored fault fingerprint library containing various known fault modes in this step.

[0031] The second-order difference calculation and feature extraction of the magnetic field spatial gradient specifically include: Step S211, dynamic determination of adaptive difference step; Read the filtered magnetic field matrix B _filtered [i,j,t], calculate the magnetic field intensity difference ΔB[i,j]=|B[i+1,j]-B[i,j]| between adjacent sensors, when ΔB exceeds 0.1T, dynamically reduce the difference step h from the default 10mm to 5mm, virtually generate intermediate measurement point data between the original 16 measurement points through linear interpolation, and obtain a high-density magnetic field matrix B _dense [i,j,t] (up to 32x32 data points), to ensure sufficient spatial resolution in the area where the magnetic field changes dramatically; Step S212, high-precision calculation of the second-order spatial derivative; Based on the high-density magnetic field matrix B _dense [i,j,t], the five-point central difference formula is used to calculate the second-order partial derivative: Ψ 2 B / Ψx 2 ≈[B(i-2,j)-2B(i-1,j)+2B(i+1,j)-B(i+2,j)] / (2h 2 ), which improves the accuracy level compared to the three-point formula. Calculate x and y directions respectively to obtain the second-order derivative matrix D 2 B _x [i,j,t] and D 2 B _y [i,j,t], both of which directly reflect the spatial curvature distribution of the magnetic field; Step S213, physical meaning analysis of the magnetic field curvature feature; The second-order derivative matrix D2B _x [i,j,t] and D2B _y [i,j,t] are physically interpreted: according to Maxwell's equations, Ψ 2 B / Ψx 2∝ΨJ / Ψt (J is current density), that is, the spatial curvature of the magnetic field directly corresponds to the time rate of change of the current; the curvature strength K = sqrt[(Ψ 2 B / Ψx 2 ) 2 +(Ψ 2 B / Ψy 2 ) 2 ] is calculated, when K > 10 T / m 2 , it is determined as abnormal curvature, and the curvature anomaly flag matrix K _flag [i, j, t] (1 = abnormal, 0 = normal) is output; Step S214, real-time detection of gradient mutation points; The curvature anomaly flag matrix K _flag [i, j, t] is subjected to connected domain analysis, and a continuous abnormal area is identified. The centroid coordinates of each connected domain are calculated as the gradient mutation center, and the moving trajectory of the mutation center in the continuous 3 sampling periods is counted, and when the moving speed exceeds 100 m / s, it is determined as a fault precursor; The gradient feature matrix G[i, j, t] is output, which contains the curvature value, mutation flag and moving speed information.

[0032] Embodiment three, a preferred scheme for accurate fault identification based on high-density magnetic field matrix and pattern matching is described. In this embodiment, adaptive spatial resolution and more precise pattern matching algorithm are introduced, which realizes higher accuracy of fault type and location identification, especially suitable for application scenarios with strict requirements on identification accuracy.

[0033] Step S310, before calculating the second-order spatial partial derivative, the method further comprises: evaluating the difference of magnetic field intensity of adjacent positions in the filtered magnetic field matrix, when the difference exceeds a preset spatial change threshold, dynamically reducing the difference step length for calculating the second-order partial derivative to obtain an adaptive difference step length; according to the adaptive difference step length, interpolation is performed between the original measurement points of the filtered magnetic field matrix to generate a high-density magnetic field matrix with higher spatial resolution containing virtual measurement point data; wherein the second-order spatial partial derivative is calculated based on the high-density magnetic field matrix.

[0034] In this embodiment, the filtered magnetic field matrix is B _filtered [i, j, t _k ]. Further, the controller calculates the difference of magnetic field intensity between adjacent sensors ΔB[i, j] = |B _filtered [i+1, j] - B _filtered[i,j]. When the difference AB exceeds a preset spatial variation threshold, such as 0.1T, the controller determines that the region has a dramatic magnetic field variation, which is likely to be a fault core region. Accordingly, the difference step size h used to calculate the spatial partial derivative is dynamically reduced from the default value (e.g., the physical spacing of the sensors 10mm) to a smaller value (e.g., 5mm). According to the reduced difference step size, the controller virtually generates data for intermediate measurement points between the original physical measurement points by linear interpolation or the like algorithm, to obtain a high-density magnetic field matrix B _dense [i,j,t _k ] with higher spatial resolution. For example, the original 4x4 sensor matrix data can be interpolated to an 8x8 or even 32x32 data matrix with higher density. In this embodiment, by using adaptive difference step size and interpolation technology, when a region with dramatic magnetic field variation is detected, the spatial sampling resolution of the region can be automatically improved, which provides a finer data basis for subsequent calculation of high-precision second-order spatial derivative and capture of weak curvature features in the early stage of fault, and improves the sensitivity and accuracy of fault recognition.

[0035] Step S320, calculate the second-order spatial partial derivative based on the high-density magnetic field matrix, and identify the abnormal curvature distribution pattern to determine the fault type and location.

[0036] Specifically, the controller calculates the second-order spatial partial derivative based on the high-density magnetic field matrix B _dense generated in step S310, using a five-point central difference formula with higher precision than the three-point formula, for example: Ψ 2 B / Ψx 2 ≈[B _dense (i-2,j)-2B _dense (i-1,j) + 2B _dense (i+1,j)-B _dense (i+2,j)] / (2h 2 ); where Ψ is the partial derivative. According to Maxwell's equations, the spatial curvature of the magnetic field is directly related to the time rate of change of the current (Ψ 2 B / Ψx 2 ∝ΨJ / Ψt), so the calculated second-order derivative matrix can directly reflect the development of the fault source current, achieving early mapping of the fault.

[0037] To identify the distribution pattern of the second-order derivative matrix (or the curvature matrix G synthesized therefrom), the matrix is normalized. The processing includes energy normalization G _norm [i,j]=G[i,j] / sqrtΣ(G[k,l] 2 ) and spatial centralization, which can eliminate the interference of the absolute intensity and position of the fault on pattern recognition, and only retain the shape features of the relative spatial distribution.

[0038] Furthermore, the normalized shape feature G _norm The system performs correlation matching with a pre-stored fault fingerprint database F_lib, which contains various known fault modes. This fingerprint database was constructed through extensive finite element simulations and experimental tests, and includes features such as Gaussian distribution characteristics corresponding to IGBT short circuits, dipole distribution characteristics corresponding to diode breakdowns, and linear distribution characteristics corresponding to bus flashovers. To achieve fast matching, this process preferably employs a two-dimensional cross-correlation algorithm based on Fast Fourier Transform (FFT), the calculation of which can be expressed as corr = ifft2(fft2(G _norm •conj(fft2(pattern))); where corr is the correlation result matrix; ifft2 is the two-dimensional inverse Fourier transform; fft2 is the two-dimensional Fourier transform; conj is the conjugate operation; and pattern is a fault pattern in the fault fingerprint database.

[0039] Preferably, based on the calculated matching scores of all fault modes, when the highest score exceeds a preset threshold (e.g., 0.85 determined based on ROC curve analysis of 1000 fault samples), the fault mode with the highest score is identified as the type of the current fault. Based on this, the fault location is precisely determined using the coordinates of the point with the maximum value in the original second derivative matrix or curvature matrix.

[0040] Example 4 describes a preferred implementation of adaptive damping control. This example transforms the power MOSFET from a pure switching device into a continuously adjustable adaptive damper, providing a finely controllable energy dissipation path for fault transient processes. The specific steps of this example are as follows: Step S410: Perform transient suppression operation, which includes: determining a target damping value based on the fault type and development speed contained in the fault prediction information, and controlling at least one power switching device in the inverter power unit to make the power switching device operate in its linear region or subthreshold region to provide variable damping corresponding to the target damping value.

[0041] Among them, power switching devices typically refer to MOSFETs in the H-bridge power unit; the subthreshold region refers to the gate-source voltage V of the MOSFET. gs Below its turn-on threshold voltage V th However, there is still a region where a small amount of leakage current flows, and in this region the on-resistance R of the MOSFET is... ds For V gs The changes are extremely sensitive, exhibiting characteristics similar to a variable resistor. Specifically, the controller determines the fault type F according to Embodiment Two or Three. _type and the rate of failure development v _fault Query the preset two-dimensional lookup table R _damp = LUT(F_type , v _fault ), determine the most suitable target damping value R _damp for the current operating condition. For example, for a rapidly developing IGBT short circuit, a small damping value (e.g. 10Ω) can be required to quickly discharge the current; while for a more moderate fault, a larger damping value (e.g. 500Ω) can be required for suppression. This step can effectively suppress the peak of overcurrent and overvoltage, avoiding damage to the device due to excessive electrical stress, and has higher control accuracy than the traditional crowbar circuit.

[0042] Step S420, controlling the power switching device includes: obtaining the real-time device temperature of the power switching device, and according to the target damping value and the real-time device temperature, calculating in a pre-stored lookup table representing the relationship among the gate voltage, the on-resistance and the device temperature to determine the initial gate voltage; monitoring the change rate of the real-time device temperature, and dynamically correcting the initial gate voltage according to the change rate to generate the final gate control voltage applied to the gate of the power switching device.

[0043] Specifically, to achieve accurate control of the on-resistance of the MOSFET, the controller obtains the real-time device temperature T _mos through a temperature sensor built-in or externally attached to the device; _damp According to the target damping value R _mos and the real-time temperature T _mos , in a three-dimensional lookup table (V gs -R ds -T relationship table) established in advance through device datasheet and actual measurement calibration, the initial gate voltage V gs_init required to achieve the target damping is calculated through a fast interpolation algorithm such as three-linear interpolation. This three-dimensional lookup table, for example, can cover the V gs range of 2.5V to 4.5V, the R ds range of 1Ω to 1kΩ (logarithmic distribution), and the temperature range of 25°C to 125°C.

[0044] Further, considering that the device junction temperature can change rapidly during the transient process, the controller also monitors the temperature change rate dT / dt in real time. When dT / dt exceeds a certain threshold (e.g. 5°C / ms), to compensate for the non-linear decrease in on-resistance due to temperature rise, the controller will dynamically negatively correct the initial gate voltage.

[0045] The correction formula can be: V gs_final =V gs_init -β×dT / dt; where V gs_final is the final control voltage applied to the gate of the MOSFET, in volts (V); β is a pre-set temperature compensation coefficient, for example, 0.002V / (°C / ms).

[0046] To ensure the accuracy of voltage control, the digital-to-analog converter (DAC) in the gate drive circuit is preferably a 12-bit or higher resolution automotive-grade chip to achieve a voltage resolution at the 1mV level. By introducing dual compensation for real-time temperature and temperature change rate, it can be ensured that the damping value provided by the MOSFET remains stable even under severe thermal shock, thereby improving the reliability and accuracy of transient suppression.

[0047] Optionally, to further improve the robustness of the damping control, a PI (proportional-integral) controller can be introduced on the basis of the above open-loop control; this PI controller measures the actual on-resistance (R) of the MOSFET in real time. _actual= V ds / I ds ) and compare it with the target value to form a closed-loop feedback, for V gs_final Fine-tuning is performed to eliminate damping value deviations caused by model errors or unmodeled dynamics.

[0048] Optionally, precise damping control of the MOSFET subthreshold region includes: Step S411: Fast interpolation algorithm for three-dimensional lookup table; Read device temperature T _mos And the target damping requirement, in the pre-stored V gs -R ds -T Locates the 8 nearest data points in the 3D lookup table and uses a trilinear interpolation algorithm to calculate the precise gate voltage: V gs =Σ(w _i ×V gs_i ), where the weight w _i The interpolation is determined based on the inverse relationship between temperature and resistance; the interpolation calculation is implemented in the FPGA using a parallel multiplier with a delay of less than 50ns, outputting the initial gate voltage V. gs_init .

[0049] Step S412: Real-time correction mechanism for temperature compensation; Continuous monitoring of device temperature T _mos The rate of change of Vt indicates that the power consumption of the device is increasing when the temperature rises rapidly (dT / dt>5°C / ms), requiring a reduction in Vt. gs To reduce the conduction current; compensation formula: V gs_comp =V gs_init -β×dT / dt, where β=0.002V / (°C / ms) is the temperature compensation coefficient; the output temperature-compensated gate voltage V gs_temp This ensures that the damping value remains stable when the temperature changes.

[0050] Step S413: PI controller parameter self-tuning; Measuring actual drain-source current I ds , calculating actual damping R _actual =V ds / I ds , comparing with target damping R _target to get error e=R _target -R _actual ; PI controller output: AV gs =K p ×e+K i ×∫e•dt, wherein K p and K i are adaptively adjusted according to damping interval: when R _target <10Ω, K p =0.1, K i =0.01; when R _target >100Ω, K p =0.01, K i =0.001; Outputting PI-regulated gate voltage V gs_pi .

[0051] Step S414, online monitoring of subthreshold stability; Real-time calculation of transconductance gm=dI ds / dV gs , determining as deep subthreshold region when gm<0.01S, which has risk of out-of-control; determining as linear region when gm>1S, which has excessive power consumption; normal working interval is 0.01S<gm<0.1S; when exceeding safe interval, limiting V gs in safe range through limiting circuit; outputting gate control voltage V gs_target and real-time damping value R _damp (t) to realize precise controllable damping of 1Ω~1kΩ.

[0052] Embodiment five describes a sub-capacitance time-sharing trigger energy absorption strategy based on reverse integration, which regards fault energy as time-varying function rather than constant value, adopts sub-capacitance array for time-sharing and quantitative absorption, and avoids disadvantages of traditional single absorption mode.

[0053] Specific steps are as follows: Step S510, performing transient suppression operation, which further includes: constructing a customized energy timing model describing energy release process over time of the fault based on fault type and development speed in fault prediction information; generating a dynamic time-sharing trigger sequence for multiple sub-capacitances according to the customized energy timing model, and sequentially connecting multiple sub-capacitances in sequence to cooperatively absorb transient energy.

[0054] Wherein, the plurality of sub-capacitors refers to a preset capacitor array for absorbing transient energy, for example, the array can be composed of 10 independent 10 μF sub-capacitors, with a total capacity of 100 μF; this design replaces the traditional single large capacitor, allowing for precise control of the energy absorption process in stages. Further, the controller determines the fault type F _type and the fault development speed v _fault , obtains the typical energy release curve E _typical (t) of this type of fault by table lookup. Further, the time axis of this typical curve is linearly scaled using the fault development speed v _fault , to construct an energy timing model E _custom (t) tailored for this specific fault; this model describes the complete process of energy release over time in the next 100 μs from the occurrence of the fault (t = 0).

[0055] In this embodiment, by fine-tuning the energy absorption process, the release rhythm of energy can be accurately matched, avoiding voltage overshoot due to insufficient absorption, or causing new problems due to excessive absorption (which may cause secondary impact on the system or insufficient energy during subsequent operation).

[0056] Step S520, generating a dynamic time-sharing trigger sequence, includes: performing reverse integration on the customized energy timing model from the end time of energy release to obtain a remaining energy curve representing the remaining energy to be absorbed at each time; based on the remaining energy curve, calculating the energy to be absorbed in each preset time window, determining the trigger time and trigger threshold of each sub-capacitor to form a dynamic time-sharing trigger sequence.

[0057] Specifically, the generation of the dynamic time-sharing trigger sequence adopts a reverse integration strategy. Instead of sequentially calculating from the fault start time (t = 0), the controller starts from the estimated end time of energy release (e.g. t _final = 100 μs), and performs reverse integration on the power curve (P(τ), i.e. the differential of E _custom (t)) of the energy timing model to obtain a remaining total energy curve that needs to be absorbed in the future at any time t, with the calculation formula: E _remain (t) = ∫ t_final t (P(τ) dτ); where E _remain (t) is the remaining total energy that needs to be absorbed in the future at time t, with the unit of Joule (J); P(τ) is the instantaneous power function; τ is the integration time variable.

[0058] On this basis, the controller divides the entire 100 μs time window into 10 (or other number) time windows corresponding to the number of sub-capacitors; the kth time window needs to absorb energy ΔE _k The ΔE can be calculated by the residual energy curve: _k = E _remain (t _k-1 ) - E _remain (t _k ); where t _k and t _k-1 are the start and end times of the kth time window.

[0059] Correspondingly, the trigger voltage threshold V _trig_k of the kth sub-capacitor is determined by the energy formula: V _trig_k = sqrt (2 x ΔE _k / C _k ); where C _k is the capacitance of the kth sub-capacitor.

[0060] Using the reverse integration strategy, the entire energy absorption process is target-oriented, i.e. when the last sub-capacitor is triggered, all the remaining fault energy can be absorbed, achieving precise matching and exhaustive absorption of energy, and maximizing the utilization efficiency of the energy storage element.

[0061] Optionally, compensation for non-ideal characteristics of the actual circuit is introduced. For example, to consider the actual RC time constant of the capacitor charging, the trigger time can be set in advance, such as t _trigger = t _target - 3τ, where τ = R•C is the charging time constant of the sub-capacitor branch. In addition, to prevent excessive absorption due to model errors, a safety margin can be set, for example, when the cumulative absorbed energy reaches 90% of the estimated total energy, the subsequent sub-capacitors are forced to stop triggering, leaving a certain safety margin.

[0062] Embodiment six, for explaining the preferred implementation of topology reconfiguration. This embodiment changes from the traditional mode of passive waiting for phase-locked loop synchronization to the advanced control strategy of actively predicting phase and compensating in advance, to realize truly disturbance-free and impact-free reconfiguration. In this embodiment, the specific steps include: Step S610, making a topology reconfiguration decision, which includes: predicting the phase evolution trajectory of the healthy power unit in a future preset time window after isolating the faulty power unit, based on the fault location contained in the system state parameter after suppression and the electrical parameters of each healthy power unit, to generate a predicted phase trajectory matrix; based on the predicted phase trajectory matrix, determine the optimal reconfiguration topology configuration and the corresponding phase continuous switching sequence to form a synchronization control instruction set under the condition of meeting the system topology constraint.

[0063] wherein the topology reconfiguration decision refers to the process of deciding how to reorganize the inverter topology with the remaining healthy H-bridge units (including the possible spare units) after isolating the faulty H-bridge unit. Specifically, the decision process takes the post-suppression system state parameters (containing V _res , I _res , φ _res ) as initial conditions, generates a predicted phase trajectory matrix φ _pred describing the phase evolution of each healthy unit within the next 100 μs by a phase prediction algorithm (e.g. step S620 in the following).

[0064] Further, based on the predicted trajectory, an optimization algorithm (e.g. particle swarm optimization PSO or a simplified priority-based rule engine) is used to search among all possible topology configurations, with the optimization objectives typically including minimizing phase jumps, minimizing power imbalance, and minimizing the number of switching actions. The search process finally determines the optimal reconfigured topology T _optimal and the corresponding phase continuous switching sequence Seq _switch .

[0065] Step S620, predicting the phase evolution trajectory of the healthy power units, includes: establishing a coupled dynamic model for the healthy power units, which takes into account the phase traction effect between the healthy power units due to the isolation of the faulty power unit; and solving the coupled dynamic model to obtain the predicted phase trajectory matrix, with the post-suppression system state parameters as initial conditions.

[0066] The phase traction effect refers to the effect that, in a cascaded system, when a unit is cut off or its state changes, it will affect the voltage phase of other units through the common impedance of the power grid, like the effect of pulling each other. To accurately predict the phase trajectory, the controller establishes a coupled dynamic equation set for all healthy units. The equation set not only contains the basic oscillation characteristics of each unit itself, but also introduces a coupling term representing the phase traction effect.

[0067] The phase traction strength F _pull_i can be modeled as inversely proportional to the electrical distance d _i between the units, for example: F _pull_i = K _pull / d _i 2 ; where K _pull is a preset phase traction coefficient, for example 0.5 rad•m 2 / s; d _iis the electrical distance between the ith healthy unit and the removed faulty unit. After the model is established, the controller takes the post-suppression system state parameters as initial conditions, and solves the coupled differential equations by numerical integration method (e.g. fourth-order Runge-Kutta method) with a time step of 1 μs to obtain the phase evolution trajectories of each healthy unit in the next 100 μs, forming a predicted phase trajectory matrix φ _pred .

[0068] In step S630, the phase continuous switching sequence is determined, including: based on the predicted phase trajectory matrix, calculating the expected phase jump of each healthy power unit at the topology switching time; before the topology switching time arrives, a preset pre-compensation time period, actively adjusting the output phase of the healthy power unit to offset the expected phase jump.

[0069] Specifically, after the optimal reconstruction topology T _optimal and the switching time t _switch are determined, the controller calculates the expected phase jump of each healthy unit at the switching moment based on the predicted phase trajectory matrix φ _pred : Δφ _jump_i = φ _after_i - φ _before_i ; wherein φ _after is the steady-state target phase of the unit after switching to the new topology, and φ _before is the phase to be reached under the old topology. To offset the expected phase jump, the controller uses a phase pre-compensation technique. This technique actively and smoothly adjusts the output phase of each healthy unit by fine-tuning its PWM modulation wave within a preset pre-compensation time period (e.g. starting 50 μs in advance) before the switching time t _switch arrives. For example, a linear adjustment strategy can be used: φ _comp [i, t] = φ _orig [i, t] + Δφ _jump_i × (t - (t _switch - 50 μs)) / 50 μs; wherein φ _comp is the real-time phase after compensation, and φ _orig is the original predicted phase. Through this pre-compensation adjustment, it can be ensured that at the time t _switch , when the physical switch action occurs, the actual output phase of all healthy units has been absolutely aligned with the target phase under the new topology, realizing continuous transition of the phase and eliminating the circulating current and voltage surge caused by phase discontinuity.

[0070] Optionally, the step S620 includes: Step S621, mathematical modeling of the phase traction effect; based on the fault location coordinates [x _f , y_f ], calculate the electrical distance d between the fault unit and each healthy unit _i = sqrt[(x _i - x _f ) 2 + (y _i - y _f ) 2 ] ; the phase traction strength is inversely proportional to the distance: F _pull_i = k / d _i 2 , where k = 0.5 rad • m 2 / s is the traction coefficient; establish the phase dynamics equation: d 2 φ _i / dt 2 = -ω0 2 (φ _i - φ_ref) - ∑ F _pull [j] × (φ _i - φ[j]), output the phase traction matrix F _matrix [n × n].

[0071] Step S622, fast solution of the coupled differential equations; Write the phase dynamics equation set of n units in matrix form: Φ̈ = -Ω 2 Φ - F _matrix × Φ, where Φ = [φ1, φ2,..., φ n ] T ; solve using the fourth-order Runge-Kutta method, with a time step Δt = 1 μs. To speed up the calculation, pre- perform eigenvalue decomposition on F _matrix : F = QΛQ -1 , solve in the eigenvalue space and then transform back to the original space; output the eigenvalue decomposition result [Q, Λ] for subsequent use.

[0072] Step S623, real-time error compensation mechanism; Compare the predicted phase φ _pred [i, t] with the measured phase φ _meas_i every 10 μs, calculate the prediction error ε _i = φ _meas_i - φ _pred [i, t] ; estimate the system state deviation through a Kalman filter and update the state transition matrix; When the prediction error exceeds 2°, introduce an error correction term: φ _corr [i, t] = φ _pred [i, t] + α × ε _i × exp(-t / τ), where α = 0.8 is the correction strength and τ = 20 μs is the decay time constant.

[0073] Step S624, constraint condition setting of trajectory smoothness; Smoothness check is performed on the predicted phase trajectory, and the phase acceleration d 2 φ / dt 2 When the acceleration exceeds 1000 rad / s 2 , smoothing processing is performed: the trajectory is regenerated by using cubic spline interpolation, ensuring that the first derivative is continuous and the second derivative is bounded; and the predicted phase trajectory matrix φ _pred [1-n, 0-100 μs] is output, each element containing phase value and its first and second derivative information.

[0074] Example Seven, a specific numerical calculation case of prediction and transient suppression parameters for describing IGBT short circuit fault, to demonstrate how the method disclosed in the present application generates key control parameters for transient suppression and topology reconstruction from original sensor data through a series of fast and accurate calculations. Specifically: Step S71, scenario setting; It is assumed that in a cascaded H-bridge frequency converter system, a certain H-bridge power unit located near the origin of the coordinate system has a serious short circuit fault of IGBT (Insulated Gate Bipolar Transistor). The physical spacing of the 4x4 magnetic field sensor array in the system is 10 mm. At an early time t _k , the controller collects the following set of original magnetic field intensity data matrix B _raw (unit: Tesla T) and has performed filtering processing to obtain B _filtered (for simplicity, it is assumed that the filtered value is the same as the original value here): B _filtered [4x4]=[[0.01, 0.02, 0.03, 0.01], [0.03, 0.15, 0.45, 0.04], [0.04, 0.25, 0.35, 0.05], [0.02, 0.04, 0.06, 0.02]]; Step S72, first stage: fault prediction and positioning calculation; In this stage, the controller performs the efficient prediction method of Example Two.

[0075] Step S721, key feature vector calculation; The controller calculates the first-order spatial gradient matrix G based on B _filtered matrix and extracts the key feature vector therefrom. For illustration, only the calculation of part of the gradient points is shown here: G[2,3]≈|B[2,3]-B[2,2]| / h=|0.45-0.15| / 0.01=30T / m; Through calculation of the entire matrix, we obtain: peak gradient G _max: After the calculation, the maximum value in the gradient matrix is G _max = 55 T / m; Gradient center [x _c , y _c ]: By calculating the centroid of the gradient matrix, the coordinates of the gradient center are determined as (21 mm, 19 mm), which are initially determined as the fault location; Peak change rate dG _max / dt: Assuming that the peak gradient calculated at the last sampling time t _k-1 (for example, 1 μs ago) is 43 T / m, the current peak change rate is dG _max / dt = (55-43) / 1 μs = 12 T / m / μs; Step S722, the fault decision controller inputs the calculated feature vector into a preset decision tree for judgment. The decision rule is: if (G _max > 30 T / m 2 AND dG _max / dt > 10 T / m / μs) then Fault_type = IGBT serious short circuit.

[0076] After judgment, G _max = 55 T / m > 30 T / m and dG _max / dt = 12 T / m / μs > 10 T / m / μs, both conditions are met.

[0077] Therefore, the controller completes the decision within a few microseconds, and outputs the fault prediction information: fault type = IGBT serious short circuit, fault development speed = fast, and fault location = (21 mm, 19 mm).

[0078] Step S73, second stage: transient suppression parameter calculation; In this stage, the controller calculates the control parameters of adaptive damping and energy absorption in parallel based on the prediction information of the first stage.

[0079] Step S731, adaptive damping parameter calculation (corresponding to embodiment four); The controller determines the current most suitable target damping value R _damp = 15 Ω according to the determined IGBT serious short circuit and fast development speed, and determines the current most suitable target damping value R _type = 15 Ω according to the determined IGBT serious short circuit and fast development speed, and determines the current most suitable target damping value R _fault = 15 Ω according to the determined IGBT serious short circuit and fast development speed, and determines the current most suitable target damping value R _damp = 15 Ω according to the determined IGBT serious short circuit and fast development speed, and determines the current most suitable target damping value R _mos = 15 Ω according to the determined IGBT serious short circuit and fast development speed, and determines the current most suitable target damping value R _damp = 15 Ω according to the determined IGBT serious short circuit and fast development speed, and determines the current most suitable target damping value R_mos = 50°C, in V gs -R ds -T three-dimensional look-up table, the initial gate voltage V gs_init = 3.85V.

[0080] Further, the controller applies a temperature dynamic compensation formula for correction: V gs_final = V gs_init - β × dT / dt = 3.85V - 0.002V / (°C / ms) × 4°C / ms = 3.842V; this V gs_final voltage value will be immediately sent to the high-precision DAC and applied to the gate of the MOSFET.

[0081] Step S732, time-sharing energy absorption parameter calculation (corresponding to Example Five); Based on the determined IGBT serious short-circuit fault type, the controller retrieves its customized energy timing model E _custom (t) from the library, and estimates that the fault will release a total energy of about 5J within 100μs. Using the reverse integration strategy, the controller calculates that the energy ΔE _1 that needs to be absorbed within the first 10μs time window (i.e. t=0 to t=10μs) is 1.2J.

[0082] Correspondingly, the trigger voltage threshold V _1 of the first 10μF sub-capacitor C _trig_1 is calculated as follows: V _trig_1 = sqrt(2 × ΔE _1 / C _1 )=sqr(2 × 1.2J / (10 × 10 -6 F))≈489.9V; when the bus voltage of the fault unit rises due to the fault impact and reaches 489.9V, the first sub-capacitor will be triggered and put into use.

[0083] Step S74, calculation conclusion; In the IGBT short-circuit fault event simulated this time, the control method of the application generates the following set of accurate control parameters for executing subsequent transient suppression operations within a very short time (usually within 20μs) after the fault occurs through a series of fast and reproducible calculations: fault type: IGBT serious short-circuit; fault location: (21mm, 19mm); target gate control voltage: 3.842V; first sub-capacitor trigger threshold: 489.9V; Example Eight, describes a two-dimensional cross-correlation matching algorithm based on fast Fourier transform, as a specific implementation way of associating and matching the normalized shape features with the fault fingerprint library in the fault recognition process, including: Step S810, the normalized shape feature is matched with the fault fingerprint library, including: using a two-dimensional cross-correlation algorithm based on fast Fourier transform, including: performing two-dimensional fast Fourier transform on the normalized shape feature and the feature vector in the fault fingerprint library, respectively; in the frequency domain, multiplying the transformed normalized shape feature and the conjugate of the transformed feature vector to obtain a frequency domain representation of the cross-correlation result; performing two-dimensional inverse fast Fourier transform on the frequency domain representation of the cross-correlation result to obtain a correlation coefficient matrix, and determining a matching result based on the peak value of the correlation coefficient matrix.

[0084] In the embodiment, in order to complete the accurate matching of the magnetic field distribution mode in a very short time of microseconds, the controller preferably uses an algorithm based on fast Fourier transform (FFT) to realize two-dimensional cross-correlation operation, and the process includes: The controller performs two-dimensional fast Fourier transform (2D-FFT) on the normalized shape feature matrix G _norm , and the feature vector pattern of each fault mode taken out from the pre-stored fault fingerprint library F _lib , respectively, and transforms them from the time domain (spatial domain) to the frequency domain to obtain their frequency domain representations G _fft and P _fft .

[0085] Further, in the frequency domain, the controller performs a point-by-point multiplication operation of complex matrices to obtain a frequency domain representation C _fft of the cross-correlation result. The calculation formula is: C _fft =G _fft ⊙conj(P _fft ); wherein C _fft is the frequency domain matrix of the cross-correlation result; ⊙ represents Hadamard product, that is, point-by-point multiplication of matrices; conj(P _fft ) is the complex conjugate of the fault mode frequency domain matrix P _fft . In the embodiment, using the convolution theorem, the time domain (spatial domain) convolution operation with extremely high calculation complexity is equivalently converted into the frequency domain multiplication operation with reduced calculation complexity, which is the key to realizing real-time matching.

[0086] On this basis, the controller performs two-dimensional inverse fast Fourier transform (2D-IFFT) on the frequency domain cross-correlation result C _fft to transform it back to the time domain (spatial domain) to obtain the final correlation coefficient matrix corr. Each element value in the matrix represents the matching degree of the fault mode pattern at the corresponding position of the G _norm matrix. The controller can determine the best matching degree and relative position of the current magnetic field distribution and the fault mode by finding the peak (maximum value) in the corr matrix. Through G _normThe above calculation is performed on all patterns in the fingerprint library one by one, and a set of matching degree scores is obtained, and the highest score is the final discrimination result. By using the FFT-based algorithm, the calculation complexity of two-dimensional cross-correlation can be reduced from O(N 4 ) to O(N 2 logN) (where N is the matrix dimension), which makes it possible to achieve real-time pattern matching in microseconds on high-performance FPGAs.

[0087] Embodiment Nine describes a preferred implementation of dynamic subdivision of the trigger window in the energy absorption strategy, which solves the problem of how to introduce an adaptive safety mechanism to cope with energy surges under extreme fault conditions in the process of generating a dynamic time-sharing trigger sequence. In a specific embodiment, the steps are as follows: Step S910, before determining the trigger time and the trigger threshold, the method further comprises: comparing the calculated energy to be absorbed in each time window with the safety absorption energy threshold of the sub-capacitor; when the energy to be absorbed in a certain time window exceeds the safety absorption energy threshold, dynamically subdividing the certain time window into at least two sub-windows to form an optimized trigger window sequence; wherein the determination of the trigger time and the trigger threshold is based on the optimized trigger window sequence.

[0088] In this embodiment, to prevent a single sub-capacitor from being subjected to an energy surge exceeding its safe operating range in a very short time, the controller, after calculating the energy ΔE _k to be absorbed in each fixed time window (e.g. 10 μs) in Embodiment Five, adds the steps of safety checking and dynamic adjustment.

[0089] Specifically, the controller compares the calculated ΔE _k with a preset safety absorption energy threshold E _safe of the sub-capacitor. E _safe is a parameter determined comprehensively according to the capacitance, rated voltage and heat dissipation capacity of the sub-capacitor, for example, it can be conservatively set as E _safe = 0.5 × C × (0.8 V _rated ) 2 ; where C is the capacitance of the sub-capacitor, and V _rated is the rated voltage of the sub-capacitor.

[0090] When the controller detects that the energy ΔE _k to be absorbed in a certain time window (e.g. the kth window) exceeds the safety threshold E _safeAt this moment, the controller determines that if the original plan is executed, it may cause excessive electrical or thermal stress to the kth sub-capacitor. Accordingly, the controller immediately starts the window dynamic subdivision mechanism to adaptively and dynamically subdivide this original 10 μs time window into two or more shorter sub-windows. For example, it can be subdivided into two 5 μs sub-windows, and the total energy ΔE _k is roughly evenly distributed into the two new sub-windows.

[0091] After the subdivision operation, the original trigger window sequence is updated to an optimized trigger window sequence containing shorter sub-windows. The subsequent trigger time and trigger threshold calculation of the controller will be strictly based on this new, optimized sequence; the huge energy originally planned to be absorbed by one sub-capacitor within 10 μs is now dispersed to be absorbed by two sub-capacitors (or the same capacitor at two different times) in two 5 μs windows.

[0092] In this embodiment, the introduction of this dynamic subdivision mechanism of trigger windows improves the adaptability and safety of the energy absorption strategy, which ensures that even under the most severe, extremely concentrated energy release fault impact, each energy storage sub-capacitor can always work within its safe operating area (SOA), avoiding device damage or secondary failure due to poor design of the energy absorption link, and enhancing the robustness of the entire self-healing system.

[0093] Optionally, the step S910 comprises: Step S911, reverse integral calculation of the energy timing curve; Read the customized energy timing model E _custom (t), and start reverse integration from the t = 100 μs moment: E _remain (t) = ∫[t to 100 μs] P(τ) dτ to obtain the residual energy curve E _remain (t) at each moment; since the energy is not sequentially absorbed, the energy required to be absorbed in each time period is calculated according to the residual energy, ensuring that the last capacitor absorbs all the energy.

[0094] Step S912, optimal matching of sub-capacitor capacity and time window; Based on the residual energy curve E _remain (t), divide 100 μs into 10 windows, each corresponding to a 0.1 μF sub-capacitor; calculate the energy required to be absorbed by the kth window: ΔE _k = E _remain ((k-1) × 10 μs) - E _remain (k × 10 μs); when ΔE _k > 0.5 × C × V 2max, the window is subdivided into two 5μs sub-windows, the trigger strategy is dynamically adjusted, and an optimized trigger window sequence T is generated _window [1-n] (n can be greater than 10).

[0095] Step S913, adaptive adjustment algorithm of trigger threshold; According to the optimized trigger window sequence T _window [1-n] and the energy requirement ΔE of each window _k , the trigger voltage threshold V is calculated: _trig_k =sqrt(2×ΔE _k / C _k ); considering the actual charging curve of the capacitor, a correction coefficient α=1.15 is introduced to compensate for the nonlinearity of RC charging, and a corrected trigger voltage threshold V _trig_adj [1-n] is obtained; a safety margin is set, and when V _trig_adj >0.9×V _rated , it is automatically assigned to the next capacitor; Step S914, safety boundary setting to prevent excessive absorption; The actual absorption energy E _absorbed_k of the triggered capacitor is monitored, and when the cumulative absorption energy reaches 90% of the total fault energy, the subsequent capacitor triggering is prohibited to prevent excessive absorption from causing insufficient system energy; a safety boundary flag Safe _flag is set, and when Safe _flag =1, the triggering is forcibly stopped; the sub-capacitor trigger timing table T _cap [1-10] and the trigger voltage threshold V _trig [1-10] are output, ensuring accurate controllability of energy absorption.

[0096] Embodiment Ten: Describes the preferred implementation of the damping closed-loop control based on PI feedback regulation, and explains how to upgrade the adaptive damping control from an open loop to a closed-loop control system with higher accuracy and stronger robustness. Specifically includes: The voltage generated by the initial gate voltage and the dynamic correction applied thereto is defined as the open-loop control voltage, and the control of the power switching device further includes: measuring the actual on-resistance of the power switching device to obtain an actual damping value; comparing the actual damping value with a target damping value to generate a damping error; generating a feedback regulation voltage based on the damping error using a proportional-integral controller; and wherein the final gate control voltage applied to the gate of the power switching device is generated by superimposing the feedback regulation voltage and the open-loop control voltage.

[0097] Optionally, the embodiment introduces a closed-loop feedback control loop. Specifically, the gate control voltage V gs_final, defined as the open-loop control voltage of the feedforward; this open-loop voltage can quickly set the working point of the MOSFET near the target damping value.

[0098] On this basis, the controller measures the drain current I ds and the drain-source voltage V ds across it in real time through the high-speed sampling circuit ds ; according to which the controller can calculate the actual on-resistance R _actual= of the device in real time, i.e. ds V ds / I _actual , which is defined as the actual damping value.

[0099] Correspondingly, the controller compares the actual damping value R _actual with the target damping value R _damp , obtains the real-time damping error e(t)=R _damp -R _actual (t), and inputs it into the digital proportional-integral (PI) controller. The PI controller generates the feedback adjustment voltage V _pi (t) based on the error. Its calculation formula is: V _pi (t)=K p ×e(t)+K i× ∫e(τ)dτ; where V _pi (t) is the feedback adjustment voltage output by the PI controller; K p is the proportional coefficient; K i is the integral coefficient; e(t) is the damping error at the current time; and ∫e(τ)dτ is the historical integral of the error. The values of K p and K i can be adaptively adjusted according to the interval in which the target damping value R _damp is located, so as to obtain the best dynamic response and stability in different damping intervals (for example, a lower damping interval requires faster response, and a higher damping interval requires higher stability).

[0100] On this basis, the final gate control voltage V gs_ultimate actually applied to the gate of the power switching device is generated by superimposing the feedback adjustment voltage V _pi (t) and the open-loop control voltage V gs_final in real time, i.e. V gs_ultimate (t)=V gs_final + V _pi (t).

[0101] The embodiment constructs a compound control structure of feedforward + feedback; the open-loop voltage of feedforward guarantees the rapidity of control, and the PI regulation of feedback is responsible for eliminating the steady-state error, guaranteeing the accuracy of control. By introducing the closed-loop control mechanism, the deviation of the damping value caused by various uncertain factors such as device aging, temperature drift, model inaccuracy, etc. can be effectively inhibited, so that the actual damping value can more accurately and robustly track the target value, and a higher level and more reliable control of the fault transient process is realized.

[0102] Embodiment eleven describes the specific process of phase accurate prediction based on RLS online identification and Kalman filter correction, and sets forth two core algorithms used in the phase prediction link to improve the prediction accuracy and adaptability: one is the recursive least square method for online updating of model parameters, and the other is the Kalman filter for real-time correction of predicted trajectories.

[0103] The specific steps of the embodiment include: Step S11.1, establishing a coupled dynamics model, including: reading a historical phase difference sequence from a circular buffer, and estimating the load power in real time; using the recursive least square method with a dynamically adjusted forgetting factor, based on the historical phase difference sequence and the real-time load power, online updating the model parameters of the coupled dynamics model, wherein the forgetting factor is dynamically adjusted according to the running stability of the system.

[0104] In the embodiment, in order to make the phase prediction model adaptive to the changing operating conditions of the frequency converter (especially the fluctuation of the load power), the controller uses the recursive least square method (RLS) to online identify and update the key parameters of the model. Specifically, the controller reads a historical phase difference sequence Δφ _history_k , from a circular buffer (for example, storing the historical data of the last 1000 power grid cycles), and estimates the load power P _load in real time by measuring the DC bus voltage and current.

[0105] Furthermore, the controller uses the RLS algorithm with a forgetting factor to online update the parameters θ (for example, θ = [a0, a1, a2] _history of the nonlinear model for describing the relationship between phase drift and load, based on Δφ _load and P T . _load , corresponding to the model dΔφ / dt = a0 + a1×P _load + a2×P 2 .

[0106] The update formula of the RLS algorithm is: θ(k) = θ(k-1) + P(k)×x(k)×[y(k)-x(k) Tx(k) = [1, P(k-1), x(k-1)] ; where θ(k) is the updated model parameter vector at time k; P(k) is the covariance matrix; x(k) is the input vector at time k, e.g. x(k) = [1, P(k-1), x(k-1)] ; y(k) is the actual observation at time k, i.e. dA / dt. _load (k) = P(k-1) - (P(k-1) x(k)T) / (1 + x(k)T P(k-1) x(k)). _load (k) 2 ] T

[0107] In particular, the forgetting factor λ in the RLS algorithm is dynamically adjusted according to the running stability of the system. For example, the controller judges the stability of the system by monitoring the fluctuation rate of the load power P _load . When the system is running in a steady state condition (small load fluctuation), λ is set to a larger value, e.g. 0.99, to retain more historical data information and enhance the stability of the model parameters. When the system experiences a transient process (such as load mutation), λ is dynamically adjusted to a smaller value, e.g. 0.95, to quickly forget outdated historical data, so that the model can quickly adapt to the new working condition and re-converge to accurate parameter estimation. Through this online identification mechanism, it is ensured that the phase prediction model can always accurately reflect the current dynamic characteristics of the system.

[0108] Step S11.2, after obtaining the predicted phase trajectory matrix, the method further comprises: comparing the predicted phase trajectory matrix with the actual phase measured in real time to obtain a prediction error sequence; using a Kalman filter to estimate the state deviation of the system based on the prediction error sequence to generate a phase error correction term; applying the phase error correction term to the predicted phase trajectory matrix to generate a higher-precision predicted phase trajectory matrix that is corrected in real time.

[0109] After obtaining the preliminary predicted phase trajectory matrix φ _pred by solving the coupled dynamic model, the controller introduces a real-time correction link based on Kalman filtering. Specifically, the controller compares the predicted phase trajectory φ _pred [i, t] with the actual phase φ _meas [i, t] of each healthy unit measured in real time by a high-precision phase-locked loop (PLL) or the like, to obtain a real-time prediction error sequence ε[i, t] = φ _meas [i, t] - φ _pred [i, t].

[0110] ​The prediction error sequence ε is inputted into the Kalman filter as an observation. The Kalman filter is an optimal state estimation algorithm, which can recursively estimate the internal states (e.g. the true phase, phase velocity and phase acceleration) of a dynamic system with noise. By using the noisy measurements ε to correct its estimation of the system states through its prediction and update steps, the Kalman filter can extract the true bias states of the system from the noisy signal and generate the optimal phase error correction term φ _corr .

[0111] Based on this, the controller applies the phase error correction term φ _corr outputted by the Kalman filter to the original predicted phase trajectory matrix φ _pred in real time and point by point, i.e. φ _final_pred = φ _pred + φ _corr , to generate the final predicted phase trajectory matrix with higher accuracy after real-time correction. By introducing the Kalman filter for closed-loop correction, the influence of measurement noise and model random disturbance on the prediction result can be effectively suppressed, so that the accuracy of phase prediction can reach the level of ±0.5°, providing an extremely reliable phase reference for subsequent disturbance-free switching.

[0112] Embodiment twelve describes a specific implementation of multi-objective topology optimization based on a particle swarm optimization algorithm. The implementation steps are as follows: Step S12.1, determining the optimal reconfiguration topology configuration includes: constructing the process of determining the optimal reconfiguration topology configuration as a multi-objective optimization problem with at least one of minimizing the phase jump, power imbalance degree and switch action number as the objective function; using a particle swarm optimization algorithm to search in the feasible topology space defined by the system topology constraints to solve the multi-objective optimization problem and obtain the optimal reconfiguration topology configuration.

[0113] Specifically, to find the best balance point among multiple conflicting performance indicators, the controller mathematically constructs the process of determining the optimal reconfiguration topology configuration as a multi-objective optimization problem. Specifically, the objective function of this optimization problem can include at least the following three items: Minimize phase jump: f1 = min(∑|Δφ _jump_i |); where Δφ _jump_i is the expected phase jump of the i-th healthy unit at the switching moment defined in embodiment six. This objective aims to ensure the electrical smoothness of the switching process; Minimize power imbalance degree: f2 = min(σ(P _load_i )); where σ(P _load_i ) is the load power P _loadthe standard deviation of the voltage of the system. This objective aims to guarantee the long-term stability of the reconfigured system; Minimize the number of switch actions: f3 = min(N _switch ); where N _switch is the total number of switch (including power switch and disconnect switch) actions required to implement the topology switching. This objective aims to reduce the switch loss and improve the device lifetime.

[0114] To solve this complex multi-objective optimization problem, the controller preferably adopts the Particle Swarm Optimization (PSO) algorithm. This algorithm is a swarm intelligence optimization algorithm simulating the foraging behavior of bird swarm. The specific implementation is as follows: In the feasible topology space defined by the system topology constraints (e.g., unit connection relationship, availability of backup units, etc.), the particle swarm is randomly initialized. Among them, the position vector of each particle represents a specific feasible topology reconfiguration scheme.

[0115] Further, the algorithm enters the iterative optimization process. In each iteration, each particle will update its own speed and position vector according to its current fitness value (calculated by the above multiple objective functions weighted or through Pareto sorting), its own historical optimal position (p Best ) experienced, and the global optimal position (g Best ) found by the whole swarm, searching in the solution space.

[0116] After a sufficient number of iterations, the whole particle swarm will gradually converge to the Pareto optimal frontier of this multi-objective optimization problem. This frontier is a set of non-inferior solutions, and any solution in the set cannot further optimize any of its objectives without sacrificing at least one other objective performance.

[0117] On this basis, the controller selects the solution that best meets the current system requirements from the solution set of this Pareto optimal frontier according to the pre-set higher-level decision rule (e.g., assigning different weights to different objectives, or prioritizing solutions with the smallest phase jump), and outputs it as the final optimal reconfiguration topology configuration T _optimal . By using PSO and other global optimization algorithms, compared with simple rule engines, it can more likely avoid falling into local optimum, find a reconfiguration scheme that performs well in multiple dimensions such as electrical smoothness, operational stability, and economy, and achieve higher-quality system self-healing.

[0118] Embodiment thirteen, provide a hardware implementation architecture based on asynchronous parallel processing, which provides high-performance and high-reliability physical carrier, expounds the real-time control signal comprehensive output link of asynchronous parallel architecture.

[0119] In the disclosed self-healing fault reconfiguration control method, multiple tasks with different time scales and computational complexities are involved. For example, the preliminary fault prediction and the highest priority protection actions need to be responded within nanoseconds (ns); the generation and update of pulse width modulation (PWM) signals need to be completed within 100 nanoseconds (100 ns); and complex pattern matching, phase prediction, and topology optimization algorithms may require microseconds (μs) level of computation time. Traditional single-processor or single-FPGA architectures face serious interrupt response delays and resource competition problems when processing such multi-time scale complex tasks, making it difficult to guarantee the real-time requirements of all tasks.

[0120] To solve this technical problem, the present embodiment discloses an asynchronous parallel processing architecture based on three FPGAs. In this architecture, complex control tasks are explicitly functionally decomposed according to their time scales and assigned to three cooperating high-performance FPGAs. Specifically: Step S13.1, task decomposition and FPGA unit configuration, includes: First, the protection unit (FPGA1): this unit preferably uses an FPGA chip with low delay characteristics, such as the Artix-7 series from Xilinx; it performs all the highest priority fast protection actions that need to be completed within 10 ns; for example, this FPGA directly processes raw data from the magnetic field sensor array and performs the most basic gradient or curvature threshold judgment; once any signal exceeding the limit safety threshold is detected, FPGA1 bypasses all upper-layer algorithms and directly outputs trip or clamping instructions to the drive stage of the power device, achieving the fastest hardware-level protection; Second, the modulation unit (FPGA2): this unit preferably uses an FPGA chip with rich I / O resources and strong PWM generation capability, such as the Spartan-7 series from Xilinx; it performs all high-precision PWM waveform generation and real-time update that need to be completed within 100 ns; this FPGA receives modulation parameters (such as duty cycle, phase, dead time, etc.) from the coordination unit (FPGA3) and independently and in parallel generates high-frequency PWM control signals for each H-bridge power unit; Thirdly, the coordination and calculation unit (FPGA3): this unit preferably adopts a system on chip (SoC) FPGA integrated with a hard core processor (such as an ARM core), for example, the Zynq-7000 series of Xilinx Corporation; it performs all the upper coordination control and complex algorithm operations that need to be completed within 1 μs; it includes the FFT pattern matching of embodiment three, the reverse integral energy calculation of embodiment five, the PI closed-loop control of embodiment ten, the RLS online identification and Kalman filtering of embodiment eleven, and the PSO topology optimization of embodiment twelve, etc.; the built-in ARM processor can be responsible for processing communication with the host computer, data recording, and management of the entire system state machine.

[0121] Step S13.2, synchronization and data interaction mechanism of the architecture; To ensure that the three independent FPGAs can work cooperatively as a whole, the architecture adopts a set of precise synchronization and data interaction mechanisms, specifically: Synchronization mechanism: all three FPGAs are provided with homogenous clocks by a common, high-stability temperature-compensated crystal oscillator (TCXO), eliminating clock drift; more importantly, all the instructions and data packets exchanged between the FPGAs are attached with 64-bit high-resolution time stamps. The time stamps are generated by a synchronized nanosecond-level counter; upon receiving an instruction, the receiving FPGA does not execute it immediately, but waits for its local synchronized counter to reach the time stamp of the instruction before executing it accurately; in this way, nanosecond-level synchronization of control actions between the three FPGAs can be achieved; on this basis, the system automatically measures the transmission delay of each channel during initialization through a loopback test and pre-compensates the time stamps; Data interaction mechanism: the three FPGAs are interconnected through a high-speed parallel bus (for example, an AXI bus implemented in the internal logic resources of the FPGA); to avoid read-write conflicts, the communication between the FPGAs preferably adopts a dual-port RAM (BRAM) as a data buffer and uses a ping-pong buffering mechanism. That is, two RAMs are set up, when the sending FPGA writes data to area A, the receiving FPGA can safely read the data of the previous cycle from area B, and in the next cycle, they are alternately used, realizing non-conflict, pipelined data transmission.

[0122] By decomposing control tasks into different hardware units according to time scales to realize true parallel processing, the hardware architecture of the embodiment can ensure that the protection task with the most stringent real-time requirement is never blocked by complex algorithms with longer time consumption. This architecture design guarantees the deterministic delay and fast response of each link in the entire prediction-suppression-reconstruction self-healing control process, and is an important foundation for successfully implementing the disclosed method with high performance and high reliability.

[0123] The preferred embodiments of the present application are described in detail above, but the present application is not limited to the specific details of the above-described embodiments, and various equivalent transformations of the technical solutions of the present application can be made within the technical concept of the present application, and these equivalent transformations all belong to the protection scope of the present application.

Claims

1. A self-healing fault reconfiguration control method for cascaded H-bridge frequency converters, characterized in that, include: The raw magnetic field strength data collected by the magnetic field sensor array is analyzed, and based on the spatial distribution characteristics of the raw magnetic field strength data, fault prediction information containing fault type and location is generated. Based on fault prediction information, perform transient suppression operations and determine the post-suppression system state parameters after the transient suppression operations are completed. Based on the suppressed system state parameters, a topology reconfiguration decision is made, and a synchronous control command set is generated and output to the power unit of the frequency converter.

2. The method according to claim 1, characterized in that, The original magnetic field strength data is analyzed to generate fault prediction information, including: The original magnetic field strength data is filtered to obtain the filtered magnetic field matrix. For the filtered magnetic field matrix, calculate its second-order spatial partial derivatives along at least two spatial dimensions to construct a magnetic field curvature matrix that characterizes the curvature of the spatial distribution of the magnetic field. Identify abnormal curvature distribution patterns in the magnetic field curvature matrix to determine the fault type and location, and generate fault prediction information.

3. The method according to claim 2, characterized in that, Before calculating the second-order spatial partial derivatives, the method also includes: The difference in magnetic field strength between adjacent positions in the filtered magnetic field matrix is ​​evaluated. When the difference exceeds a preset spatial variation threshold, the differential step size used to calculate the second partial derivative is dynamically reduced to obtain an adaptive differential step size. Based on the adaptive differential step size, interpolation is performed between the original measurement points of the filtered magnetic field matrix to generate a high-density magnetic field matrix with higher spatial resolution that includes virtual measurement point data. The second-order spatial partial derivative is calculated based on the high-density magnetic field matrix.

4. The method according to claim 2, characterized in that, Identifying abnormal curvature distribution patterns in the magnetic field curvature matrix to determine fault type and location includes: The magnetic field curvature matrix is ​​normalized to eliminate the influence of fault intensity and absolute position, and the normalized shape features representing the relative spatial distribution are extracted. The normalized shape features are correlated and matched with a pre-stored fault fingerprint database containing multiple known fault modes, and the fault type is determined based on the matching results. The location of the fault is determined by the maximum value point in the magnetic field curvature matrix.

5. The method according to claim 1, characterized in that, Perform transient suppression operations, including: Based on the fault type and development speed contained in the fault prediction information, the target damping value is determined, and at least one power switching device in the inverter power unit is controlled to make the power switching device work in its linear region or subthreshold region, providing variable damping corresponding to the target damping value.

6. The method according to claim 5, characterized in that, Controlling power switching devices includes: The real-time device temperature of the power switching device is obtained, and the initial gate voltage is determined by calculation based on the target damping value and the real-time device temperature in a pre-stored lookup table that characterizes the relationship between gate voltage, on-resistance and device temperature. The rate of change of the real-time device temperature is monitored, and the initial gate voltage is dynamically corrected based on the rate of change to generate the final gate control voltage applied to the gate of the power switching device.

7. The method according to claim 1, characterized in that, Performing transient suppression operations also includes: Based on the fault type and development speed in the fault prediction information, a customized energy time series model is constructed to describe the energy release process of the fault over time. Based on a customized energy timing model, a dynamic time-division triggering sequence is generated for multiple sub-capacitors, and multiple sub-capacitors are turned on sequentially according to the sequence to collaboratively absorb transient energy.

8. The method according to claim 7, characterized in that, Generate a dynamic time-sharing trigger sequence, including: For a customized energy time series model, the reverse integration over time is performed starting from the end of its energy release to obtain the curve representing the remaining energy to be absorbed at each time. Based on the remaining energy curve, the energy to be absorbed in each preset time window is calculated, the triggering time and triggering threshold of each of the multiple sub-capacitors are determined, and a dynamic time-division triggering sequence is formed.

9. The method according to claim 1, characterized in that, Making topology reconfiguration decisions includes: Based on the fault location and electrical parameters of each healthy power unit contained in the system state parameters after suppression, the phase evolution trajectory of the healthy power unit within a preset time window after isolating the faulty power unit is predicted, and a predicted phase trajectory matrix is ​​generated. Based on the predicted phase trajectory matrix, under the condition of satisfying the system topology constraints, the optimal reconfiguration topology configuration and the corresponding phase continuous switching sequence are determined, forming a synchronization control instruction set.

10. The method according to claim 9, characterized in that, Predicting the phase evolution trajectory of healthy power cells includes: Taking into account the phase traction effect caused by the isolation of faulty power units, a coupled dynamics model is established for healthy power units, where the phase traction effect characterizes the phase interaction between healthy power units. Using the suppressed system state parameters as initial conditions, the coupled dynamics model is solved to obtain the predicted phase trajectory matrix.

Images