Fault warning method for cascaded H-bridge inverter

Abstract

The present invention discloses a fault warning method for a cascaded H-bridge frequency converter, comprising: obtaining the field current signal of each H-bridge unit, bidirectionally verifying the signal zero point using the reverse zero crossing reconstruction method, and generating a reconstructed timestamp matrix; performing phase space differential mapping based on the matrix, and extracting a characteristic matrix that can characterize the intrinsic dynamic characteristics of the system by calculating topological invariants such as trajectory length and curvature; after reducing the dimension of the characteristic matrix, using an immune memory prediction mechanism to perform affinity matching with a preset fault pattern library to achieve rapid fault identification and extrapolate to generate a predicted feature vector; combining the predicted features, residuals, and memory activation, introducing a quorum sensing collaborative decision model, and simulating the diffusion process of abnormal signals to achieve collaborative and robust decision-making on the fault type, occurrence time, and impact range, and generate early warning instructions. The present invention significantly improves the accuracy, reliability, and lead time of fault warnings.

Description

Technical Field

[0001] The present invention relates to an early warning method, in particular to a fault early warning method for a cascaded H-bridge frequency converter. Background Art

[0002] As a core component of modern power electronics systems, cascaded H-bridge inverters are playing an increasingly important role in renewable energy generation, motor drive, and power transmission. With the continuous improvement of power levels and the increasing complexity of application scenarios, the reliability of cascaded H-bridge inverters has become a key bottleneck restricting their large-scale application. Due to the large number of units and complex mutual coupling in the cascaded H-bridge structure, faults propagate quickly and have a wide impact range. The traditional post-fault maintenance model can no longer meet the stringent requirements of modern industry for continuous operation. Therefore, the development of high-precision and timely fault warning technology to achieve a transition from passive response to active prevention has important theoretical significance and practical value for ensuring the safe and stable operation of power systems, reducing operation and maintenance costs, and improving equipment utilization.

[0003] Current fault warning technologies for cascaded H-bridge inverters primarily focus on voltage harmonic analysis, current amplitude monitoring, and frequency domain feature extraction. Voltage harmonic analysis identifies faults by monitoring changes in the harmonic content of the output voltage, but this method is insensitive to minor faults and is susceptible to load fluctuations. Current amplitude monitoring utilizes the statistical characteristics of each unit current for fault detection, but its time resolution is limited, making it difficult to capture transient fault characteristics at the millisecond level. Frequency domain analysis uses FFT transforms to extract fault characteristic frequencies, but this is computationally complex and ineffective for non-stationary signals. In recent years, some research has focused on machine learning-based fault diagnosis methods, such as support vector machines and neural networks. While these methods have demonstrated advantages in feature extraction and pattern recognition, they generally suffer from large training sample requirements and limited generalization capabilities. Furthermore, model-based fault diagnosis methods, which rely on residual analysis by establishing a mathematical model of the system, face significant challenges in practical engineering applications due to their high model complexity and difficulty in parameter identification.

[0004] However, existing technical solutions still face numerous problems when addressing the specific technical challenges of cascaded H-bridge systems. For example, there is the instantaneous paradox of zero-crossing detection. Traditional zero-crossing detection assumes the existence of an ideal zero-point instant, but the actual current signal has tiny oscillations and noise interference near the zero point, resulting in multiple pseudo-zero-crossing triggers. While existing threshold filtering methods can suppress some interference, threshold setting relies on manual experience and cannot adapt to the dynamic changes in signal characteristics under different operating conditions. This problem of finding discrete time points in a continuous signal limits the accuracy of time measurement, which in turn affects the accuracy of subsequent fault feature extraction. Summary of the Invention

[0005] The purpose of the invention is to provide a cascade H-bridge inverter fault warning method, in order to solve the above-mentioned problems existing in the prior art.

[0006] Technical solution: A cascaded H-bridge inverter fault warning method, comprising:

[0007] Obtain the field current signal of each level of the H-bridge unit, apply the reverse zero-crossing reconstruction method to perform bidirectional verification processing on the zero point of the field current signal, and generate a reconstructed timestamp matrix;

[0008] Based on the reconstructed timestamp matrix, phase space mapping and dynamic feature extraction are performed to obtain the phase space feature matrix and perform dimensionality reduction processing to obtain the main feature matrix and residual matrix;

[0009] Predict immune memory based on the main feature matrix to generate prediction feature vectors and memory activation;

[0010] The eigenvector, residual matrix and memory activation degree are jointly predicted to make group sensing collaborative decisions and generate early warning instructions.

[0011] According to one aspect of the present application, a reverse zero-crossing reconstruction method is applied to perform bidirectional verification processing on the zero point of the field current signal, including:

[0012] For each candidate zero-crossing point, the positive value sequence before zero and the negative value sequence after zero are deconstructed, and the negative value sequence is subjected to time reversal and numerical inversion to shape the inversion sequence of the ideal positive value trajectory;

[0013] Apply interpolation algorithm to the inversion sequence to reconstruct the ideal positive trajectory that is precisely aligned with the positive sequence in time;

[0014] The comprehensive matching degree between the ideal positive trajectory and the actually observed positive sequence is quantified, and the candidate zero-crossing point is confirmed as the real zero-crossing point only when the comprehensive matching degree exceeds the preset verification threshold.

[0015] According to one aspect of the present application, an interpolation algorithm is applied to the inversion sequence to reconstruct an ideal positive trajectory, including:

[0016] Set first-order and second-order derivatives as boundary constraints for the inversion sequence;

[0017] Under boundary constraints, a cubic spline function is constructed for the inversion sequence to perform smooth extension in the time domain.

[0018] The cubic spline function is sampled at a high density to generate an ideal positive trajectory.

[0019] According to one aspect of the present application, quantifying the comprehensive matching degree between the ideal positive value trajectory and the actually observed positive value sequence includes:

[0020] From the three dimensions of trajectory morphological similarity, numerical deviation amplitude and random noise level, the normalized correlation coefficient, root mean square relative error and signal-to-noise ratio between the ideal positive trajectory and the positive sequence are calculated respectively, and weighted fusion is performed to generate a comprehensive matching degree.

[0021] According to one aspect of the present application, phase space mapping and dynamic feature extraction are performed based on the reconstructed timestamp matrix to obtain a phase space feature matrix, including:

[0022] Applying cross-unit differential operations to the reconstructed timestamp matrix to generate the original differential sequence that represents the multi-unit temporal relationship;

[0023] The original difference sequence is mapped to a three-dimensional phase space consisting of the current difference value, the difference value at the previous moment, and the difference change rate, to shape the phase space trajectory of dynamic evolution;

[0024] The topological invariants of the analytical phase space trajectory are fused into the phase space characteristic matrix after temperature rise drift self-calibration.

[0025] According to one aspect of the present application, the original differential sequence is mapped to a three-dimensional phase space to shape a dynamically evolving phase space trajectory, including:

[0026] Assign the current difference value and the previous moment difference value in the original difference sequence as the first axis coordinate and the second axis coordinate of the three-dimensional phase space respectively;

[0027] By performing numerical differentiation operations on the original difference sequence, the difference change rate that represents its change trend is obtained;

[0028] The differential change rate is filtered and smoothed, and the smoothed differential change rate is used as the third axis coordinate to construct a phase space trajectory.

[0029] According to one aspect of the present application, the topological invariant of the phase space trajectory is analyzed, including:

[0030] For the phase space trajectory, the trajectory length, average curvature and gyration radius are quantified respectively from the aspects of its overall path length, local curvature and spatial distribution compactness, and combined into topological invariants.

[0031] According to one aspect of the present application, the predicted eigenvector, residual matrix, and memory activation are combined to perform group sensing collaborative judgment and generate early warning instructions, including:

[0032] Each cascade H-bridge unit is abstracted as a biological individual, and the predicted feature vector and residual matrix are fused to calculate the initial abnormal concentration for each individual;

[0033] A diffusion propagation model is constructed based on the physical connection between individuals to simulate the spatiotemporal evolution of the initial abnormal concentration between adjacent individuals and form a diffusion concentration field.

[0034] When the concentration in a local area of ​​the diffusion concentration field exceeds a threshold, a coordinated warning is triggered, thereby generating a warning instruction.

[0035] According to one aspect of the present application, a diffusion propagation model is constructed based on the physical connection relationship between individuals to simulate the spatiotemporal evolution of the initial abnormal concentration between adjacent individuals, including:

[0036] The physical connection relationship between individuals is abstracted into a weighted network graph, and a weighted graph Laplacian matrix is ​​constructed for the network graph to characterize the coupling strength and topological structure between individuals.

[0037] The diffusion partial differential equation is constructed based on the weighted graph Laplace matrix and numerically solved with the initial anomalous concentration as the initial condition to obtain the diffusion concentration field.

[0038] According to one aspect of the present application, immune memory prediction is performed based on the main feature matrix to generate a prediction feature vector and memory activation degree, including:

[0039] Perform affinity matching between the main feature matrix and the preset fault mode memory library, calculate the similarity between the current system state and each historical fault mode, and obtain the affinity matrix;

[0040] Based on the affinity matrix, when its value exceeds the activation threshold, the corresponding memory mode is triggered, and the main feature matrix is ​​extrapolated according to the activated memory mode to generate a predicted feature vector;

[0041] At the same time, the memory activation degree is calculated based on the affinity matrix and the inherent weight of the memory pattern.

[0042] According to one aspect of the present application, affinity matching is performed between the main feature matrix and the preset fault mode memory library to obtain an affinity matrix, including:

[0043] By integrating Euclidean distance, Mahalanobis distance and cosine similarity distance, the comprehensive distance between the main feature matrix and each memory mode in the fault mode memory library is calculated;

[0044] Estimate adaptive variance parameters for each memory pattern that take into account both intra-class compactness and inter-class separation;

[0045] Based on the comprehensive distance and adaptive variance parameters, the affinity matrix is ​​generated through exponential function transformation.

[0046] According to one aspect of the present application, feature extrapolation is performed on the main feature matrix according to the activated memory pattern to generate a predicted feature vector, including:

[0047] Using the memory activation degree as the weight, the multiple direction vectors pointing from the current system state to each activated memory mode are weighted and summed to determine the comprehensive weighted prediction direction;

[0048] Adaptively adjust the prediction step size based on the intensity of memory activation;

[0049] Starting from the current state of the main feature matrix, forward extrapolation is performed along the weighted prediction direction and with an adaptive prediction step size to generate a temporal prediction feature vector.

[0050] Beneficial effect: The reverse zero-crossing reconstruction theory fundamentally solves the instantaneous paradox of zero-crossing detection in the existing technology. This method no longer attempts to find an ideal discrete instantaneous point in a continuous signal full of noise and small oscillations, but creates a new paradigm for process reconstruction. It establishes a time reversal model, uses the negative value sequence with less interference in the second half of the cycle after the zero point, and reversely interpolates to reconstruct the theoretically perfect positive value change process in the first half of the cycle, and then calculates the matching degree between this ideal process and the actual process to achieve two-way verification. This mechanism improves the time measurement accuracy from the 10ps level of the traditional method to the 2ps level, allowing the system to capture the weak timing jitter caused by early faults between each H-bridge unit with extremely high fidelity at the source of data acquisition, providing a solid foundation for the accuracy of all subsequent analysis, thereby significantly improving the sensitivity and reliability of fault warning as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0052] Figure 2 This is a flow chart of the present invention for performing bidirectional verification processing on the zero point of the field current signal.

[0053] Figure 3 It is a flow chart of the present invention for reconstructing an ideal positive trajectory.

[0054] Figure 4 It is a flow chart of the present invention for obtaining the phase space characteristic matrix. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following Figures 1 to 4 The specific embodiments of the present invention are described in detail. It should be noted that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0056] In addition to the above problems, there is also the problem of phase structure loss in multi-unit differential processing: that is, the existing differential jitter analysis converts the time series into a difference matrix. Although it can eliminate common-mode interference, this scalar differential operation destroys the intrinsic dynamic structure of the time series, converting the originally ordered time information into disordered difference information, resulting in the loss of the phase space characteristics and topological invariants of the fault signal. In particular, for cascade system failure modes with complex spatiotemporal coupling characteristics, traditional linear differential methods cannot effectively extract their nonlinear dynamic characteristics, limiting the accuracy and advance time of fault warning.

[0057] Embodiment 1 describes a high-precision timestamp acquisition process based on reverse zero-crossing reconstruction. By using a bidirectional verification method, the zero-crossing point of the current signal is accurately determined, providing a high-precision time reference for fault feature extraction.

[0058] This embodiment mainly includes the following steps:

[0059] Step S101 : applying a reverse zero-crossing reconstruction method to perform a bidirectional verification process on the zero point of the field current signal to generate a reconstructed timestamp matrix.

[0060] In this embodiment, to address the issues of traditional zero-crossing detection methods, which are prone to false trigger points and low accuracy under noise interference, this method utilizes the symmetry of the half-cycles of sinusoidal alternating current before and after the zero-crossing point. The waveform of one half-cycle is used to reversely reconstruct the ideal waveform of the other half-cycle. The reconstructed waveform is then compared with the actual observed waveform. Only when there is a close match is it confirmed as a true zero-crossing point.

[0061] Specifically, this step is further refined as follows:

[0062] In the first step, for each candidate zero-crossing point, the positive value sequence before zero and the negative value sequence after zero are deconstructed, and the negative value sequence is subjected to time reversal and numerical inversion to shape the inversion sequence of the ideal positive value trajectory.

[0063] The candidate zero-crossing points are the signal polarity change points initially detected by traditional comparator methods. Time reversal and value inversion involve reversing the negative value sequence after zero on the time axis and taking the absolute values ​​of all their values, thereby obtaining a sequence that is morphologically similar to the positive value sequence before zero.

[0064] Get the synchronized current signal I from the current transformer of each H-bridge unit sync (t). For a i The candidate zero crossing point detected at the moment is separated into the positive sequence I of the first half cycle (for example, the first 5 sampling points) pos =I(t i -5:t i-1) and the negative value sequence I in the second half cycle (for example, the last 5 sampling points) neg =I(t i +1:t i +5). Then, generate the inversion sequence I rev =|I neg [end:-1:1]∣.

[0065] Because an ideal noise-free signal should have good symmetry before and after the zero crossing, an inversion operation is used to create an ideal first-half cycle sequence based on the second-half cycle information. This provides a benchmark for subsequent cross-validation and lays the foundation for the identification of noise and small perturbations.

[0066] In the second step, an interpolation algorithm is applied to the inversion sequence to reconstruct an ideal positive trajectory that is precisely aligned with the positive sequence in time.

[0067] Preferably, first-order and second-order derivatives are set as boundary constraints for the inversion sequence; under the boundary constraints, a cubic spline function is constructed for the inversion sequence to perform smooth extension in the time domain; and the cubic spline function is sampled with high density to generate an ideal positive trajectory.

[0068] In this embodiment, first, the inversion sequence I is calculated. rev The first and second order derivatives at the beginning and end are used as boundary conditions for the spline interpolation. Then, these boundary conditions are used to construct the piecewise cubic spline function S rev (t). Finally, in the time interval corresponding to the original positive sequence, rev (t) Perform high-density sampling (e.g., 10 times the original sampling rate) to obtain the reconstructed ideal positive trajectory I pos_ideal .

[0069] It should be noted that in addition to cubic spline interpolation, other shape-preserving interpolation algorithms, such as Lagrange interpolation and Akima spline interpolation, can also be used as long as they can smoothly reconstruct the waveform. The interpolation order and sampling density can be adjusted according to the requirements for computational accuracy and efficiency.

[0070] Direct comparison using discrete inversion sequences can lead to time misalignment. Interpolation, smooth extension, and high-density sampling generate ideal trajectories in the continuous time domain, enabling precise comparison with the original, discrete, positive sequence at any point in time. This improves the accuracy of subsequent matching calculations.

[0071] The third step is to quantify the comprehensive matching degree between the ideal positive trajectory and the actually observed positive sequence, and only when the comprehensive matching degree exceeds the preset verification threshold, the candidate zero-crossing point is confirmed as the real zero-crossing point.

[0072] Preferably, the normalized correlation coefficient, root mean square relative error and signal-to-noise ratio between the ideal positive trajectory and the positive sequence are calculated from the three dimensions of trajectory morphological similarity, numerical deviation amplitude and random noise level, and weighted fusion is performed to generate a comprehensive matching degree.

[0073] By quantifying the consistency between the ideal positive trajectory and the actual observed positive sequence, the authenticity of the candidate zero crossing point can be verified. total =w1·R corr +w2·(1-RMSE rel )+w3·tanh(SNR / 10); among them, the normalized correlation coefficient R needs to be calculated separately corr , root mean square relative error RMSE rel and signal-to-noise ratio SNR.

[0074] In this embodiment, the weights of the three dimensions can be set to w1=0.4, w2=0.4, and w3=0.2. Verification threshold θ match It can be set to 0.95 according to the experiment. total >0.95, the current candidate point t i is a true zero crossing point.

[0075] Research has found that single metrics (such as the correlation coefficient alone) are susceptible to interference from certain types of noise. For example, a DC offset in phase can significantly impact the root mean square error (RMS) but has little effect on the correlation coefficient. Combining metrics from multiple dimensions allows for a more robust and comprehensive evaluation of the similarity between two sequences, effectively eliminating false zero crossings caused by glitches and oscillations, and improving timestamp accuracy from the traditional 10ps level to 2ps.

[0076] Through the above steps, the current signals of all H-bridge units are processed to finally obtain a high-precision, verified reconstructed timestamp matrix T recon [N×K], where N is the number of units and K is the number of time periods. This matrix will serve as the basis for all subsequent analyses.

[0077] Traditional zero-crossing detection technology is based on the principle of instantaneous comparison, which seeks the instantaneous moment when the voltage or current amplitude crosses zero in a continuous signal. However, this method faces the instantaneous paradox in signal processing theory: theoretically, a zero crossing should occur at an infinitesimal time point, but actual signals inevitably contain noise, glitches, and small oscillations near zero, resulting in multiple false crossings in the vicinity of zero. Existing techniques employ hysteresis thresholds or digital filtering to suppress interference, but these methods essentially forcibly determine a discrete time point in a continuous, noisy signal, limiting their accuracy due to the noise level and the subjectivity of threshold setting. This paper proposes a process reconstruction verification method: rather than attempting to find the ideal zero instant in the noise, this method leverages the time reversal invariance of an ideal sinusoidal signal (i.e., the mathematical property that I(t) = -I(-t)). Using waveform data from the half-cycle after zero, the ideal noise-free trajectory of the half-cycle before zero is mathematically and rigorously reconstructed. This reconstructed trajectory is then compared to the observed trajectory for multi-dimensional similarity verification. This self-referencing, bidirectional verification mechanism transforms the time point detection problem into a waveform matching problem, eliminating the impact of noise on time measurement accuracy, and achieving an improvement in accuracy from 10ps to 2ps. This lays the technical foundation for accurately capturing microsecond-level timing jitter between units in cascaded H-bridge systems.

[0078] In addition, the field of power system fault warning has long faced the technical challenge that early fault characteristics are weak and easily concealed. Traditional methods based on frequency domain analysis and statistical feature extraction can often only capture significant changes after the fault occurs, but lack sensitivity to weak signs in the incipient stage of the fault. The present invention introduces phase space topological analysis technology to achieve a transition from surface numerical analysis to intrinsic dynamic structure analysis. Specifically, traditional differential jitter analysis converts the time series Δt into a scalar difference matrix. Although it can eliminate common-mode interference, this dimensionality reduction operation irreversibly destroys the system dynamics information contained in the time series. The present invention constructs a three-dimensional phase space mapping [Δt(k), Δt(k-1), dΔt / dk] to upgrade the one-dimensional time series information into a high-dimensional geometric object, namely the phase space trajectory. The geometric form of the trajectory completely retains all the dynamic characteristics of the interaction between multiple units in the system. More importantly, the topological invariant extracted by the present invention (the trajectory length L traj , mean curvature k avg , Radius of gyration R gyration) is invariant to geometric transformations and robust to noise, and can identify subtle changes in the system's dynamic behavior from seemingly normal operating data. For example, the tiny phase lag caused by early aging of the IGBT is almost undetectable in traditional amplitude or harmonic analysis, but it will be clearly manifested in the phase space as a distortion of the trajectory shape and an increase in the radius of gyration. The analysis method based on the intrinsic dynamic characteristics of the system enables fault warning to break through the threshold dependence limitations of traditional methods, achieve advanced identification of gradual and hidden faults, and advance the warning time window from post-fault response to the fault's incipient stage, providing a new technical path for preventive maintenance of power systems.

[0079] Example 2: This example describes the process of extracting dynamic features based on phase-space mapping, based on the reconstructed timestamp matrix Trecon generated in Example 1. The core of this step is to map the one-dimensional timestamp difference sequence into a high-dimensional phase space and, by analyzing the topological structure of its trajectory, extract system dynamic features that are unattainable using traditional differential methods.

[0080] Step S201 : performing phase space mapping and dynamic feature extraction based on the reconstructed timestamp matrix to obtain a phase space feature matrix.

[0081] In this embodiment, this step mainly solves the problem that traditional differential jitter analysis destroys the intrinsic dynamic structure of the time series and causes information loss.

[0082] Specifically, this step is further refined as follows:

[0083] In the first step, a cross-unit differential operation is performed on the reconstructed timestamp matrix to generate an original differential sequence that represents the multi-unit temporal relationship.

[0084] For example, the process is as follows: Read the reconstructed timestamp matrix T recon [N×K]. According to the physical topology of the cascaded H-bridge (such as adjacent units, same-phase units), the difference across units is made: Δt ij =T recon (i)-T recon (j). For example, for a system with N units, N-1 adjacent unit pairs and N-2 cross-phase unit pairs can be selected, resulting in a total of P=2N-3 differential sequences. The final output is the original differential sequence matrix Δt raw [P×K].

[0085] In this embodiment, differential operations are used to eliminate common-mode noise and system-level timing drift, thereby highlighting relative timing differences between units. These differences often contain early signs of failure.

[0086] In the second step, the original differential sequence is mapped to a three-dimensional phase space consisting of the current differential value, the differential value at the previous moment, and the differential change rate, to shape the dynamically evolving phase space trajectory.

[0087] Preferably, the current differential value and the differential value at the previous moment in the original differential sequence are assigned as the first axis coordinate and the second axis coordinate of the three-dimensional phase space respectively; by performing numerical differentiation operation on the original differential sequence, the differential change rate representing its change trend is obtained; the differential change rate is filtered and smoothed, and the smoothed differential change rate is used as the third axis coordinate to construct a phase space trajectory.

[0088] In one embodiment, the X-axis represents the current differential value Δt(k). The Y-axis represents the differential value Δt(k-1) at the previous moment. The Z-axis represents the differential change rate dΔt / dk. The differential change rate can be calculated using numerical methods such as central difference and smoothed using a moving average filter to suppress noise.

[0089] Through this mapping, each one-dimensional difference sequence Δtraw becomes a trajectory X evolving in three-dimensional space. phase [P×K×3].

[0090] This construction method allows the dynamic information implicit in a one-dimensional time series to be expanded into a higher-dimensional space. The X- and Y-axes represent the state of the system, while the Z-axis represents the trend of change in that state. The trajectory formed by these three axes can intuitively reflect changes in the system's dynamic behavior, such as from stability (the trajectory converges to a point or a small area) to anomaly (the trajectory diverges or the shape changes dramatically).

[0091] The third step is to analyze the topological invariants of the phase space trajectory and fuse them into the phase space characteristic matrix after temperature rise drift self-calibration.

[0092] Preferably, the calculation process may be as follows: for the phase space trajectory, the trajectory length, average curvature and gyration radius are quantified in order from the overall path length, local curvature and spatial distribution compactness, and combined into topological invariants.

[0093] Among them, the trajectory length L traj The length of each tiny line segment along the trajectory is integrated to reflect the cumulative amount of differential jitter. avg The average value of the local curvature of each point on the trajectory is calculated to reflect the severity of the trajectory bending. gyration The average distance from all points on the trajectory to its geometric centroid is calculated to reflect the distribution range and compactness of the trajectory in space.

[0094] Track length L traj =∑ k=1 K-1Δx k 2 +Δy k 2 +Δz k 2 Δx, Δy, and Δz are the coordinate differences between adjacent points on the three axes of the phase space.

[0095] Mean curvature κ avg =mean(κ loca l[k]), where the local curvature κ local It is calculated by cross product of tangent vectors of adjacent trajectory points.

[0096] Radius of gyration R gyration =sqrt(mean(∑(X phase [k,:]-center) 2 )). Xphase is the phase space coordinate matrix;

[0097] Combining these three invariants into a eigenvector [L traj , κ avg , R gyration ], and obtain the topological feature matrix Topo[P×3].

[0098] In addition to the three topological invariants mentioned above, other dynamic indicators, such as Lyapunov index and fractal dimension, can also be calculated to more comprehensively characterize the trajectory characteristics.

[0099] Because topological invariants represent a high-level summary of the trajectory geometry, they are less sensitive to noise than the trajectory itself and better reflect the inherent, essential dynamic properties of the system. For example, an initial insulation aging fault may not significantly change the mean of the differential value, but it may increase its volatility, which is directly reflected in a longer trajectory length and an increased radius of gyration.

[0100] Before fusion, the symmetry of the phase space trajectory is used to calculate the drift vector of the trajectory center of mass and perform translation correction on the coordinates to eliminate the systematic measurement drift caused by temperature changes. The final output is the calibrated phase space feature matrix Φ phase [N×M×3].

[0101] Through the above steps, the high-precision but single-information timestamp data is converted into a multi-dimensional feature matrix containing rich system dynamic information, providing high-quality input for subsequent intelligent prediction and judgment.

[0102] Example 3 describes the process of fault pattern matching based on the main feature matrix P obtained after feature dimensionality reduction. By drawing on the memory and recognition mechanism of the biological immune system, rapid matching of fault symptoms and early prediction of future status are achieved.

[0103] Step S301: Perform immune memory prediction based on the main feature matrix to generate a prediction feature vector and memory activation degree. In this embodiment, the main goal is to achieve rapid fault identification and early prediction, achieving a 300ms advance warning.

[0104] Specifically, this step is further refined as follows:

[0105] In the first step, affinity matching is performed between the main feature matrix and the preset fault mode memory library, and the similarity between the current system state and each historical fault mode is calculated to obtain the affinity matrix.

[0106] The Failure Mode Memory is a pre-built database that stores typical feature vectors for various known fault types, much like antibodies in the immune system. Affinity quantifies the degree to which the current system state (antigen) matches these antibodies.

[0107] Preferably, the comprehensive distance between the main feature matrix and each memory mode in the fault mode memory library is calculated by fusing the Euclidean distance, Mahalanobis distance and cosine similarity distance; the adaptive variance parameter that takes into account both intra-class compactness and inter-class separation is estimated for each memory mode; based on the comprehensive distance and the adaptive variance parameter, the affinity matrix is ​​generated by exponential function transformation.

[0108] First, for each fault mode k, the adaptive variance σ is calculated by analyzing the intra-class dispersion of its historical samples and the inter-class distance with other modes. k This makes the affinity calculation adaptive to the characteristic distribution of different failure modes.

[0109] Calculate the comprehensive distance D combined =w d1 ·d eucl +w d2 ·d maha +w d3 ·d cos ; Multi-scale metrics are more robust than single distance.

[0110] Generate affinity A(P,m k )=exp(-D 2 combined / 2σ k 2 ); The distance is converted into an affinity value A(P, m) in the range [0, 1] by Gaussian radial basis function k The higher the affinity, the more similar the current state is to the historical failure mode. Finally, the affinity matrix A[K×M] is obtained.

[0111] This step simulates the immune recognition process, in which the affinity function can accurately quantify the possibility of the current system state evolving into various known failure modes, providing a decision basis for subsequent activation and prediction.

[0112] In the second step, based on the affinity matrix, when its value exceeds the activation threshold, the corresponding memory pattern is triggered, and the main feature matrix is ​​extrapolated according to the activated memory pattern to generate a predicted feature vector.

[0113] Preferably, the memory activation degree is used as the weight, and a weighted sum is performed on multiple directional vectors pointing to each activated memory mode from the current system state to determine a comprehensive weighted prediction direction; the prediction step size is adaptively adjusted according to the strength of the memory activation degree; starting from the current state of the main feature matrix, forward extrapolation is performed along the weighted prediction direction and with an adaptive prediction step size to generate a temporal prediction feature vector.

[0114] When the affinity A(P,m k ) exceeds an adaptively adjusted activation threshold θ act When , the mode is activated. Activation degree α k Calculated based on affinity and the inherent weight of the pattern.

[0115] Calculate a weighted prediction direction vector pred _direction , which is the weighted average direction of the current state pointing to the center of all activated fault modes. _size It is proportional to the activation degree. The higher the activation degree, the clearer the trend of fault development and the larger the prediction step size. Then, starting from the current feature vector, linear extrapolation is performed along this direction to generate the predicted feature vector P for the next h steps. pred [d×h].

[0116] For example, the prediction time advance can be set to 300ms. If the feature sampling period is 10ms, the prediction step number h=30.

[0117] In this example, once a familiar antigen (fault symptom) is identified, the system can quickly recall its memory and predict its future direction and speed. This is more targeted and faster than purely time-series-based autoregressive models, and is particularly suitable for fault types for which prior knowledge is available.

[0118] In the third step, the memory activation degree is calculated based on the affinity matrix and the inherent weight of the memory pattern.

[0119] As mentioned above, the memory activation degree α[K×1] is a combination of affinity and the inherent weight W of the memory pattern. mem A vector calculated by factors such as αk represents the intensity of activation of the kth fault mode. This activation vector is not only used for weighted prediction, but also serves as one of the important inputs for subsequent collaborative decision making.

[0120] Through the above steps, the system can not only identify which fault the current state is similar to, but also predict the evolution trajectory of the features within a certain period of time in the future (for example, 300ms) and quantify the credibility of the prediction (memory activation), achieving a leap from diagnosis to prediction.

[0121] Example 4 describes the collaborative decision process based on quorum sensing based on the above-mentioned predicted feature vector Ppred, residual matrix R, and memory activation degree α. In this embodiment, multiple H-bridge units are treated as a whole, and the diffusion process of abnormal signals is simulated to achieve collaborative and robust fault decision.

[0122] Step S401 : Combine the predicted eigenvector, residual matrix and memory activation degree to perform group sensing collaborative judgment and generate early warning instructions.

[0123] In this embodiment, this step aims to address the problem that individual unit decisions are susceptible to noise interference and cannot accurately assess the scope of fault impact. It couples the decision-making processes of all units to form a collective decision-making system.

[0124] Specifically, this step is further refined as follows:

[0125] The first step is to abstract each cascaded H-bridge unit into a biological individual. The predicted eigenvector and residual matrix are combined to calculate the initial anomaly concentration for each individual. The biological individual is a biomimetic abstraction of each H-bridge unit. The initial anomaly concentration is a quantitative indicator that represents the degree of anomaly for each individual at the current moment.

[0126] Specifically, for the i-th unit, its initial abnormal concentration c i It consists of two parts: one part comes from the known pattern abnormalities identified by the immune memory prediction module (predicted by the feature vector P pred The other part comes from the unknown pattern anomalies detected by the feature dimensionality reduction module (reflected by the norm of the residual matrix R). The calculation formula is: i =|||P pred [ i]∣∣ 2 +β|||R[i]||| 2 . Where β is a weight factor used to balance the contribution of known anomalies and unknown anomalies. The vector C obtained in this way local [N×1] is the initial abnormal concentration distribution of the entire group.

[0127] By fusing two kinds of abnormal information of different natures.pred Represents explainable anomalies that are similar to historical failures, while R represents new anomalies that are different from all known patterns. Combining the two allows for a more comprehensive assessment of the health status of each unit.

[0128] In the second step, a diffusion propagation model is constructed based on the physical connection relationship between individuals to simulate the spatiotemporal evolution of the initial abnormal concentration between adjacent individuals and form a diffusion concentration field.

[0129] Preferably, the physical connection relationship between individuals is abstracted into a weighted network graph, and a weighted graph Laplace matrix is ​​constructed for the network graph to characterize the coupling strength and topological structure between individuals; a diffusion partial differential equation is constructed based on the weighted graph Laplace matrix, and a numerical solution is performed with the initial abnormal concentration as the initial condition to obtain the diffusion concentration field.

[0130] In some embodiments, the process is as follows:

[0131] First, an adjacency matrix is ​​established based on the physical adjacency relationship of the H-bridge unit. Then, considering the distance of the electrical coupling, weights are assigned to the edges of the matrix to obtain a weighted adjacency matrix. Finally, the weighted graph Laplacian matrix L is calculated from the weighted adjacency matrix. weighted . Used to describe the difficulty and path of abnormal signals propagating between different units.

[0132] Construct a structure like dc / dt=D▽ 2 The partial differential equation of c+S(x, t)-γc. Among them, the Laplace operator ▽ 2 c is given by the graph Laplacian matrix L weighted Approximately, the source term S is the initial abnormal concentration C local , D is the diffusion coefficient, γ is the attenuation coefficient. The equation is solved using stable numerical methods such as Crank-Nicolson to obtain the spatiotemporal evolution matrix C of the abnormal concentration in the entire H-bridge system over a period of time in the future. diff [N×T pred ].

[0133] This step simulates the diffusion of signal molecules within a bacterial population. An abnormal signal generated by a cell diffuses to its neighbors, forming a concentration field. This mechanism provides spatial smoothing and filtering. Abnormal concentrations caused by transient noise disturbances in a single cell are averaged out due to diffusion, while persistent abnormal concentrations generated by actual fault sources accumulate and diffuse, forming distinct localized areas of high concentration, enhancing the robustness of the decision.

[0134] In some embodiments, the calculation formula is: i t+1 =c it +Δτ·[D ∑ j∈N(i) L ij c j t +S i t -γ·c i t ]; C i t is the anomalous concentration of the i-th H-bridge unit at time t. Δτ is the discrete time step in the numerical solution. D is the diffusion coefficient. L ij The element in the weighted graph Laplacian matrix that represents the connection relationship between units i and j. N(i) is the set of neighboring units of unit i. S i t is the abnormal concentration source term of unit i at time t (i.e., the initial abnormal concentration). γ is the attenuation coefficient.

[0135] In the third step, when the concentration in a local area of ​​the diffusion concentration field exceeds the threshold, a coordinated warning is triggered, thereby generating a warning instruction.

[0136] Specifically, for the diffusion concentration matrix C diff When the concentration value of one or several adjacent cells is continuously higher than a dynamically adjusted group threshold θ group When , the coordinated warning is triggered. This threshold can be adaptively calculated based on the mean and standard deviation of the baseline concentration during normal system operation, for example, θ group =μ baseline +k×σ baseline , where k is the sensitivity coefficient.

[0137] Once an alert is triggered, the system integrates the information and generates a structured alert instruction. This instruction includes the fault type (determined by the activated memory mode α), the predicted time of occurrence (current time + lead time), and a list of affected units (i.e., areas where concentrations exceed thresholds). This instruction is ultimately written to higher-level monitoring systems, such as SCADA, for operator decision-making.

[0138] Through the above-mentioned group sensing mechanism, the sublimation from individual decision-making to collective wisdom is achieved, which not only improves the accuracy and reliability of early warning, but also can accurately locate the scope of fault impact, providing more precise guidance for preventive maintenance.

[0139] Example 5: Describes the complete process of the cascaded H-bridge inverter fault warning method. This demonstrates an end-to-end application from data acquisition to warning output. It should be noted that the details of Examples 1 to 4 above have been described in detail and will not be repeated here.

[0140] In a typical application scenario, for example, a photovoltaic inverter system comprising N=30 H-bridge units, the execution process of the method of the present invention is as follows:

[0141] The system synchronously collects the current signals I(t) of all 30 units through GPSDO and FPGA-PLL. For each zero-crossing candidate point in each signal cycle, the reverse zero-crossing reconstruction method is applied for bidirectional verification. For example, for a candidate point, the ideal trajectory is reconstructed using cubic spline interpolation, and its comprehensive matching degree M with the actual trajectory is calculated. total Only when M total When the value is greater than 0.95, a 10ps TDC is used to record the precise time. Finally, the reconstructed timestamp matrix T is formed. recon [30×K].

[0142] Read T recon , calculate the difference sequence Δt between adjacent units raw . Each Δt raw Mapping to the three-dimensional phase space [Δt(k), Δt(k-1), dΔt / dk], the dynamic trajectory is obtained. The three topological invariants of the length, average curvature and gyration radius of each trajectory are calculated, and temperature rise calibration is performed to form the phase space characteristic matrix Φ phase .

[0143] For Φ phase Perform sparsification and sliding window SVD decomposition to obtain the main feature matrix P[dxM] and the residual matrix R[NxM]. P is input into the immune memory prediction module and affinity matched with the fault memory library. Assuming that the affinity between the current P and the IGBT open circuit fault memory mode is the highest and exceeds the activation threshold, the mode is activated. The system then extrapolates the features based on the activation mode to generate the predicted feature vector P for the next 300ms. pred , and outputs a higher memory activation degree α for IGBT open circuit fault k .

[0144] System Fusion P pred and R, calculate the initial abnormal concentration C of each unit local The graph Laplace matrix is ​​constructed based on the physical connection relationship between units, and the C is simulated by solving the diffusion equation. local The diffusion concentration field C is obtained during the propagation process in the entire system. diff If the concentrations of cells 5, 6, 7 and their adjacent areas continue to exceed the group threshold θ during the forecast period, group , then trigger the collaborative early warning.

[0145] The system generates a warning instruction W, which reads: Fault type: IGBT open circuit fault, predicted occurrence time: [current time + 300ms], affected units: units 5, 6, and 7. This instruction is sent to the monitoring center, instructing operations personnel to inspect or isolate the affected units in advance, thus avoiding an unplanned downtime.

[0146] Through the above complete implementation process, it is possible to issue an early warning in a highly reliable and accurate manner hundreds of milliseconds before a fault occurs, and indicate the nature and scope of the fault, demonstrating great engineering application value.

[0147] Example 6 describes the process of performing dimensionality reduction processing on a high-dimensional phase space feature matrix, extracting its main components and quantifying the residual information.

[0148] This process inherits the phase space characteristic matrix Φ generated in Example 2. phase , specifically including the following steps:

[0149] Step S301: performing sparse dimension processing on the phase space feature matrix. Sparse processing refers to discarding redundant or less informative entries in the high-dimensional feature matrix while retaining key information, thereby reducing the matrix dimension and computational complexity.

[0150] For example, the input is the generated phase space feature matrix Φ phase [N×M×3]. Considering that in the cascaded H-bridge system, the strong coupling relationship between units mainly exists between units with adjacent physical locations or the same electrical phase, this step only retains the feature entries corresponding to these units with strong correlation according to the physical topology of the H-bridge. Subsequently, the three-dimensional phase space features [L traj , κ avg , R gyration ] are flattened and combined into a large two-dimensional matrix.

[0151] For a system consisting of N units, if the interactions between all units are considered, its feature dimension will reach O(N2) or even higher, resulting in a huge computational burden. Through sparse processing based on physical topology, the feature dimension can be significantly reduced from O(N2) or O(N3) level to O(N) level. While retaining most of the key fault information, the efficiency of subsequent processing is greatly improved, making it suitable for online monitoring scenarios with high real-time requirements. The output of this step is a sparse feature matrix Φ sparse [N′×M′].

[0152] Step S302: Use sliding window SVD decomposition to extract the main feature subspace.

[0153] SVD is used to decompose a matrix into the product of three matrices and identify the most important patterns or components in the data. Sliding windows move across the data using a fixed-length window, analyzing it piece by piece.

[0154] Read the sparse feature matrix Φ sparse [N′×M′]. Set a sliding window of fixed length. For example, the window length L can be set to 256 sampling points.

[0155] Perform a complete SVD decomposition on the data in the first window, that is, Φ sparse_window1 =U0Σ0V0 T Among them, the diagonal elements of the matrix Σ0 are singular values, which are arranged from large to small and represent the importance of different modes; U0 and V0 are left and right singular vector matrices, representing the specific forms of these modes.

[0156] When the window slides forward one time step, to avoid re-performing the computationally expensive full SVD decomposition, a Rank-1 correction algorithm is preferably used to quickly update the SVD decomposition result from the previous moment. This algorithm uses the new data entering the window and the data leaving the window to make small corrections to the original U, Σ, and V matrices, and is computationally much more efficient than a full decomposition.

[0157] In addition to SVD, other classic dimensionality reduction methods such as principal component analysis (PCA) can also be used. For real-time updates, in addition to Rank-1 correction, an incremental SVD algorithm based on subspace tracking can also be used.

[0158] SVD decouples the complex system dynamics into a set of orthogonal principal modes. Typically, a relatively small number of principal modes (corresponding to the largest singular values) are sufficient to capture over 99% of the system's dynamic behavior. This approach not only effectively compresses the data dimension but, more importantly, separates the information into the following matrix:

[0159] The primary characteristic matrix P[d×M], consisting of the first d largest singular values ​​and their corresponding singular vectors, is used to capture the most important and energy-intensive dynamic behavior characteristics of the system. This matrix is ​​the core basis for subsequent matching of known fault patterns.

[0160] The residual matrix R[N×M], consisting of the remaining small singular values ​​and their corresponding components, is used to represent weak or newly emerging signal components that are ignored by the main mode. This matrix is ​​valuable for detecting early-stage faults of unknown types or subtle signal fluctuations.

[0161] Through the above steps, the high-dimensional and complex phase space features are efficiently decomposed into a low-dimensional, information-concentrated main feature matrix P and a residual matrix R containing abnormal sprouts, providing ideal input for subsequent immune memory prediction and quorum sensing judgment.

[0162] Example 7 is used to describe in detail how the barrier pattern memory bank performs adaptive weight update and attenuation, so that the entire warning system has the ability to continuously learn and self-optimize. Based on Example 3, this solution further includes:

[0163] Step S404: Update and decay the weights of the fault signature memory database. Memory updating and decay simulates the proliferation and apoptosis of memory B cells in the biological immune system. This process strengthens proven memories while gradually forgetting long-term ineffective or erroneous memories.

[0164] This is executed after the system completes a warning judgment and obtains actual operational feedback (e.g., confirmation of the fault through the SCADA system or on-site verification by maintenance personnel). The input is the memory activation degree α[K×1] calculated in Example 3 and the verification result of the current fault.

[0165] For the memory mode k (i.e. α k >0), if the fault it predicts is finally verified to be true, then its weight W mem [k] is enhanced. The update rule adopts the decay update algorithm, and the formula is as follows: W mem_new [k]=γ·W mem [k]+(1-γ)·α k [cite:30]; among them, W mem_new [k] is the updated weight. γ is the memory decay coefficient, which is between 0 and 1, for example, it can be set to 0.9. This formula means that the new weight is the sum of the old weight and the new evidence (activation α k ), and γ controls the degree of retention of historical memory.

[0166] For memory patterns that have not been activated in this or multiple warnings, their weights will decay naturally. This can be achieved by multiplying their weights by a decay factor less than 1, or by directly applying the above formula (in this case α k = 0), the effect of which is that the weight W mem [k] decays exponentially with time.

[0167] The attenuation coefficient γ can be adjusted based on system stability and the rate of environmental change. In a relatively stable system, γ can be set to a larger value (such as 0.95) to rely more on historical experience. In systems with drastic operating conditions, γ can be appropriately reduced (such as 0.85) to more quickly adapt to new patterns.

[0168] This step introduces a feedback learning loop into the early warning system, allowing high-quality memory patterns that accurately predict faults to have a greater influence in future decision-making, thereby improving the sensitivity and accuracy of early warnings for similar faults. Weight decay gradually eliminates outdated (for example, due to a change in the fault mode after equipment maintenance) or erroneous memory patterns, preventing them from interfering with future decisions and keeping the memory bank clean and efficient.

[0169] Through a dynamic, adaptive update mechanism, the fault memory library is no longer a static set of rules, but a living knowledge base that grows and evolves with the equipment. This improves the robustness and adaptability of the entire fault warning method during long-term operation, ensuring its continued effectiveness under various complex operating conditions.

[0170] Example 8 describes the details of the reverse zero-crossing reconstruction algorithm and the time reversal invariance model.

[0171] The time reversal invariance model is an idealized physical model. Its core idea is based on the symmetry of the ideal sinusoidal signal. For an ideal current signal I without noise and distortion, ideal (t)=Asin(ωt), which has odd symmetry near the zero point (t=0), that is, it satisfies: I ideal (-t)=-I ideal This property is known as time reversal invariance. It states that reversing the time axis of a signal (t → -t) is equivalent to inverting its amplitude. Therefore, using the waveform information on one side of the zero point, this invariance relationship can be used to rigorously deduce the ideal waveform on the other side.

[0172] In actual engineering, the current signal I actual (t) always contains noise n(t) and distortion d(t), that is:

[0173] I actual (t)=I ideal (t)+n(t)+d(t). These interference terms destroy the perfect time reversal invariance. Therefore, by quantifying the degree of deviation of the real signal from this ideal model, we can reversely identify and eliminate the false zero crossings caused by noise and reconstruct the true zero crossing moment with high precision.

[0174] The mathematical derivation of the algorithm is as follows:

[0175] For the candidate zero-crossing point t0 initially determined by the comparator, a time window τ is taken before and after it (for example, τ is 5 sampling periods). The discrete sampling points in this window are defined as:

[0176] Positive sequence before zero: P actual={(t i , I i )|t i ∈[t0-τ, t0), I i >0};

[0177] Negative value sequence after zero: N actual ={(t j , I j )|t j ∈(t0, t0+τ], I j <0};

[0178] According to the time reversal invariance model, the negative value sequence N after zero is used actual To construct the ideal positive sequence that should appear before zero theoretically. actual Each point (t j , I j ) to transform:

[0179] Time reversal t j ′=t0-(t j -t0)=2t0-t j ;

[0180] Value Inversion I j ′=-I j This results in a new discrete point set, namely the inversion sequence: Prev={(t j ′,I j ′)}. In an ideal situation, P rev Should be with P actual Total overlap.

[0181] In order to achieve accurate comparison, the discrete inversion sequence Prev needs to be converted into a continuous function. This embodiment uses the cubic spline interpolation method to construct this ideal trajectory. Using Prev as the interpolation node, a piecewise cubic polynomial function S is constructed. ideal (t). For any interval [t j ′, t j+1 ′], the function form is: S ideal (t)=a j (tt j ′)3+b j (tt j ′)2+c j (tt j ′)+d j ; Among them, the coefficient {a j , b j , c j , d j} is obtained by solving a system of linear equations that ensures continuity of function values, first-order derivatives, and second-order derivatives at the nodes. Continuous function S ideal (t) is the ideal positive trajectory without noise reconstructed by the time reversal model.

[0182] Therefore, the actual observation value P can be quantified mathematically exactly actual With the ideal trajectory S ideal (t) The deviation between. actual At each sampling time point t i , calculate the deviation value ε i =I i -S ideal (t i )All deviation values ​​constitute the deviation sequence {ε i If the candidate point t0 is a true, low-noise zero-crossing point, then {ε i} should appear as random white noise with zero mean. On the contrary, if t0 is a pseudo-zero crossing point, then {ε i}will show an obvious trend or a larger amplitude.

[0183] Therefore, the comprehensive matching degree M obtained in Example 1 is total include:

[0184] The root mean square relative error RMSErel comes directly from the energy of the deviation sequence, RMSE=sqrt(mean(ε i 2 )).

[0185] The signal-to-noise ratio (SNR) is calculated by comparing the variance of the ideal trajectory (signal energy) with the variance of the deviation sequence (noise energy). SNR = 20log 10 (std(S ideal (t i )) / std(ε i )).

[0186] Normalized correlation coefficient R corr :Calculate the sequence {I i} and {S ideal (t i )}Pearson correlation coefficient.

[0187] This model can achieve high-precision and high-reliability zero-crossing detection.

[0188] Example 9: Description of Adaptive Variance Parameter σ kfault estimation process.

[0189] In the affinity matching calculation of Example 3, the core formula is A=exp(-D 2 / 2σ2 kfault ). In this formula, the parameter σ kfault Plays the role of receptive field width.

[0190] If σ kfault If it is too large, the receptive field will be too wide, making the fault mode have a high affinity for dissimilar features, thus causing false positives.

[0191] If σ kfault If it is too small, the receptive field will be too narrow, making the fault mode insensitive to some of its own normal variants, thus causing false negatives.

[0192] Therefore, for each fault type k fault Estimate an optimal, adaptive σ kfault It is crucial.

[0193] The goal of the algorithm is to make the estimated variance parameter take into account both intra-class compactness and inter-class separation, as follows:

[0194] Failure Mode Memory: M lib ={m1, m2, ..., m Kfault}, where m kfault is the central feature vector (i.e. antibody) of the kfault class fault.

[0195] For each type of fault kfault, prepare a set S containing n historical samples kfault ={s1, s2, ..., s n}, where each s i Each is a feature vector belonging to this type of fault.

[0196] Calculate the intra-class compactness σ 2 intra , which is used to measure the degree of discreteness of the distribution of samples of a certain type of fault.

[0197] For each fault category kfault, calculate its sample set S kfault All samples in the class to the center m kfault The square of the average Euclidean distance. 2 intra (kfault)=(1 / n)∑ i=1 n ∣∣s i -m kfault ∣∣ 2 ;

[0198] The smaller the value, the more stable and concentrated the characteristics of this type of fault are.

[0199] Calculate the inter-class separation σ2 inter , used to measure the distance between a certain type of fault and other fault categories in the feature space. For each fault category kfault, calculate its center m kfault With all other fault categories center m j (where j = kfault) and take the minimum value. 2 inter (kfault)= min j≠kfault (∣∣m kfault -m j ∣∣ 2 The larger the value, the more recognizable the fault is, and the less likely it is to be confused with other faults.

[0200] Fusion calculation adaptive variance σ 2 kfault , combining the above two indicators to obtain a comprehensive variance parameter.

[0201] A weighted fusion approach is used to combine intra-class compactness and inter-class separation.

[0202] σ 2 kfault =α bal ·σ 2 intra (kfault)+(1-α bal )· σ 2 inter (kfault); α bal It is a balancing factor between [0, 1], which determines whether the final variance focuses more on reflecting the distribution of samples within the class or on ensuring the distance from other classes.

[0203] When the fault category itself has a large variation range (σ 2 intra large), you can increase α appropriately bal (such as 0.7), so that σ kfaul t can better cover these samples.

[0204] When the fault class is very close to other classes in the feature space (σ 2 inter small), it is necessary to reduce α appropriately bal (such as 0.3), so that σ kfault becomes smaller, forming a sharper receptive field to avoid incorrectly identifying samples of adjacent categories.

[0205] Through the above steps, instead of using a global uniform variance, a unique variance parameter σ is tailored for each failure mode. kfaultThis adaptive parameter estimation method can significantly improve the accuracy and robustness of fault affinity matching.

[0206] Embodiment 10 describes the specific process of temperature rise drift detection, which is used to eliminate the systematic measurement drift caused by the sensor or acquisition circuit due to temperature change before performing topological invariant analysis.

[0207] The intrinsic dynamics of the system (i.e., the shape of the trajectory) should not change with temperature under steady-state conditions. However, temperature changes primarily cause a translation of the entire trajectory in phase space. By calculating this translation (i.e., the drift vector) and subtracting it from the real-time trajectory, the true shape of the trajectory can be restored. The specific steps are as follows:

[0208] Perform this step when the drive system is in a healthy, stable operating state and the operating temperature is at a known reference base temperature (for example, 25°C, or the thermal equilibrium temperature after a long period of stable operation).

[0209] Collect one or more cycles of phase space trajectory data that can fully represent the reference state, denoted as X base [i, k time ,:], where i is the trajectory index, k time is a time point index.

[0210] For each trajectory i, calculate its geometric center of mass C in the three-dimensional phase space base [i,:]. The centroid is the mean of the coordinates of all points on the trajectory.

[0211] The calculation formula of the drift vector (reference center of mass part) is:

[0212] C base [i,:]=(1 / K)∑ K ktime=1 Xbase[i,k time ,:]; where K is the total number of points on the trajectory. This formula can be decomposed into three coordinate axes:

[0213] C basex [i] = (1 / K)∑ K ktime=1 X basex [i, k time ];

[0214] C basey [i] = (1 / K)∑ K ktime=1 X basey [i, k time ];

[0215] C basez[i] = (1 / K)∑ K ktime=1 X basez [i, k time ];

[0216] All the calculated reference centroid vectors C base [P × 3] (where P is the total number of trajectories) is stored in non-volatile memory and serves as a reference for subsequent corrections.

[0217] This step is performed periodically while the inverter system is online.

[0218] Get the real-time phase space trajectory data X within the current analysis window current [i, k time ,:].

[0219] Using the same method as the first step, calculate the real-time center of mass C of the current trajectory current [i,:].

[0220] Subtract the real-time center of mass from the stored reference center of mass to obtain the instantaneous drift vector V drift [i,:].

[0221] Drift Vector V drift [i,:]=C current [i,:]-C base [i,:],characterizes the overall translation amount and direction of the current trajectory relative to the reference state.

[0222] After the drift vector is calculated, the current trajectory data is corrected immediately.

[0223] Subtract the drift vector corresponding to the trajectory from each point on the current trajectory.

[0224] Mathematical formula for coordinate correction: X corrected [i, k time ,:]=X current [i, k time ,:]-V drift [i,:].

[0225] The corrected trajectory data X corrected It is output to the subsequent topological invariant calculation module.

[0226] It can dynamically and in real time compensate for the measurement errors introduced by slowly changing environmental factors such as temperature, ensuring the accuracy of subsequent fault feature extraction and sensitivity to real faults.

[0227] According to one aspect of the present application, achieving nanosecond or picosecond multi-channel synchronization is a mature technology in fields such as test and measurement, communications, and power. This is typically achieved using a commercial GPSDO (GPS Disciplined Oven Oscillator) module, which receives atomic clock signals from GPS satellites and outputs a high-precision 1PPS (pulse per second) signal and a 10MHz frequency reference. Each H-bridge acquisition unit receives these synchronization signals via optical fiber or dedicated coaxial cable and uses them as a trigger or phase-locked loop reference to achieve timebase synchronization for each channel. This is standard engineering practice within the skill of those skilled in the art.

[0228] According to one aspect of the present application, managing and distributing input reference clocks within an FPGA is a standard process for digital logic design. Mainstream FPGA vendors (such as Xilinx, Intel / Altera) offer powerful, silicon-proven PLL (phase-locked loop) or MMCM (mixed-mode clock manager) hard-core IP. Designers can easily configure parameters such as frequency multiplication, division, and phase shift of the IP core using a graphical configuration interface or instantiated code to generate the various clock signals required by the system. The FPGA development environment automatically performs subsequent clock tree synthesis and routing, which is also well-known in the art.

[0229] According to one aspect of the present application, a large amount of published literature and established methods exist for implementing a high-precision time-to-digital converter (TDC) and its calibration in an FPGA. The Vernier delay line mentioned in the article is a classic TDC architecture. Its calibration typically employs code density testing, which involves inputting a signal with a known random distribution and counting the number of hits per time quantization step (bin) to measure and correct the TDC's differential nonlinearity (DNL) and integral nonlinearity (INL). Accuracy verification can be accomplished through comparative testing with a higher-precision commercial time interval analyzer.

[0230] According to one aspect of the present application, in a monitoring or diagnostic system based on data analysis, setting and tuning threshold parameters is a standard step. The method mainly includes:

[0231] Statistical analysis method collects a large amount of data under normal working conditions, calculates the mean (μ) and standard deviation (σ) of relevant features, and sets the threshold to μ+kσ (such as the 3σ principle).

[0232] Receiver operating characteristic (ROC) curve analysis, by continuously adjusting the threshold on the test set, plots the relationship curve (ROC curve) between the true positive rate (sensitivity) and the false positive rate (1-specificity) under different thresholds, and selects the optimal working point on the curve as the final threshold based on the tolerance for false positives and false negatives.

[0233] Empirical setting and online fine-tuning: Initially, a conservative initial value is set based on experience. During actual operation, the threshold is then fine-tuned online and continuously based on early warning results and feedback from operations and maintenance personnel. Those skilled in the art are fully capable of employing one or more of the above methods to find appropriate settings for each threshold in the present invention.

[0234] In short, the reverse zero-crossing reconstruction algorithm achieves unprecedented accuracy and reliability at the source of fault warning, the time measurement link. By establishing a time reversal invariance model: that is, using the current waveform data of the half cycle after zero point (usually less affected by arc interference) Ineg Through time reversal and numerical inversion, a noise-free, theoretically ideal trajectory of the first half of the cycle is mathematically reconstructed. Ipos_ideal Then, the reconstructed ideal trajectory is compared with the actual observed trajectory of the first half of the cycle I pos Perform multi-dimensional matching calculations. This self-referencing, bidirectionally verified algorithm transforms the challenge of finding discrete time points into evaluating the similarity of two curves, fundamentally resolving the technical pain point of traditional comparator methods, which are susceptible to noise and glitches and produce pseudo-zero crossings. In the specific scenario of cascaded H-bridge inverters, characterized by high-frequency switching and strong electromagnetic interference, this high-precision timestamp acquisition capability (up to picoseconds) enables the system to capture previously imperceptible inter-unit timing anomalies at the microsecond or even nanosecond level, caused by early IGBT aging and weak drive signal jitter. This provides a high-quality data foundation for subsequent, precise fault feature extraction, an advantage unmatched by traditional amplitude or harmonic analysis methods.

[0235] By mapping the timestamp difference sequence Δt into the three-dimensional phase space [Δt(k), Δt(k-1), dΔt / dk] and analyzing the topological invariants of its trajectory, the present invention addresses the core problem that traditional differential analysis methods can destroy the intrinsic dynamic structure of time series and lead to information loss. This technical advantage is achieved by creatively transforming one-dimensional, discrete time series difference information into a geometric object in a high-dimensional space—a phase space trajectory. The trajectory's shape, length, curvature, and spatial distribution intuitively and completely preserve the full dynamic information of the interactions between system units. For example, in a cascaded H-bridge application scenario, the slow aging of the capacitor of a power unit may not immediately cause a significant change in the output current amplitude, but it will cause its dynamic response to produce slight phase lags and irregular fluctuations. This change in dynamic characteristics may be obscured by traditional differential mean or variance statistics, but in phase space, it is clearly manifested as a distortion of the trajectory shape, an increase in the trajectory length (Ltraj), and an increase in the radius of gyration (Rgyr). Therefore, by calculating these topological invariants that are insensitive to noise, the present invention can extract deep features that characterize the degradation of the system health status from seemingly normal operating data, achieve accurate identification of early and slowly varying faults, and greatly improve the sensitivity and depth of early warning.

[0236] By incorporating an immune memory prediction mechanism inspired by biological immunology, this invention achieves a leap from passive diagnosis to active prediction, providing valuable advance warning of faults. This technical achievement is achieved through the deep integration of business rules (building an expert knowledge base using historical fault data) with specialized algorithms (affinity matching and feature extrapolation). Specifically, the system stores historically verified fault signatures (such as IGBT open circuits and busbar capacitor failures) as antibodies in a fault pattern memory library. During real-time monitoring, the system calculates the affinity between the current system's main eigenvector P and each antibody m_k in the library to achieve millisecond-level rapid pattern matching of fault signs. Once the affinity exceeds the activation threshold, the system immediately identifies the potential fault type and, based on the fault evolution direction information contained in the activated memory pattern, extrapolates the current signature forward, predicting the state hundreds of milliseconds (e.g., 300ms) into the future. In scenarios with extremely stringent requirements for continuous operation, such as renewable energy power generation and high-end equipment drives, these hundreds of milliseconds of lead time are crucial. They are sufficient to allow the upper-level control system to perform protective actions, such as starting backup units, safely isolating faulty units, or performing orderly shutdowns. This can transform a destructive failure that could cause a production line interruption or system downtime into a planned preventive maintenance. The advantage of its application is that it greatly reduces the huge economic losses caused by unplanned downtime.

[0237] Through a unique quorum sensing collaborative decision-making model, this invention addresses the technical challenges of single-unit decisions being unreliable and prone to false alarms in large multi-unit systems due to local noise or individual differences. This is achieved by abstracting each H-bridge unit as a biological individual and using a diffusion equation based on a graph Laplace matrix to simulate the propagation of abnormal concentrations across a physically connected population of units. The input parameter, abnormal concentration c, is determined by the predicted feature P_pred (representing known risk) and the residual R_res (representing unknown risk), ensuring comprehensive information input. The core of the algorithm is that the abnormal concentration generated by a real, persistent fault source will form a concentration field that continuously accumulates and diffuses around it. Meanwhile, transient abnormalities in a single unit caused by random noise are rapidly averaged and diluted by diffusion. Therefore, the system triggers a final warning only when the concentration in a local area exceeds an adaptive quorum threshold θ_group, derived from normal background noise statistics. In systems with a large number of tightly coupled units, such as cascaded H-bridges, this mechanism offers the advantage of extremely high decision reliability, effectively avoiding false alarms like "crying wolf." Furthermore, it can precisely locate the core unit affected by the fault and its surrounding area by identifying the spatial distribution of the concentration field, providing maintenance personnel with accurate fault location information and significantly improving operation and maintenance efficiency.

[0238] In this embodiment, the meanings of common related characters are explained as follows: fault is the category index of the failure mode, for example, k fault ∈{1,2,...,K}; k time is the index of the discrete time point of the time series, for example, k time ∈{1,2,...,M};k sens is the sensitivity coefficient, for example, used to calculate the population threshold; N unit is the total number of units in the cascaded H-bridge; N(i) is the set of neighboring units of unit i, and the set is represented by cursive letters; α act is the memory activation, such as α k ; α bal is the balancing factor used to calculate the adaptive variance. coeff is a general coefficient (or named according to its specific meaning, such as α update ). R res is the residual matrix. gyr is the radius of gyration of the phase space trajectory. corr is the normalized correlation coefficient.

[0239] In this document, terms such as "larger," "smaller," "too large," and "too small" mean larger or smaller than the corresponding threshold value. The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details of the aforementioned embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solution of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A cascade H-bridge inverter fault warning method, characterized in that: include: Obtain the field current signal of each level of the H-bridge unit, apply the reverse zero-crossing reconstruction method to perform bidirectional verification processing on the zero point of the field current signal, and generate a reconstructed timestamp matrix; Based on the reconstructed timestamp matrix, phase space mapping and dynamic feature extraction are performed to obtain the phase space feature matrix and perform dimensionality reduction processing to obtain the main feature matrix and residual matrix; Predict immune memory based on the main feature matrix to generate prediction feature vectors and memory activation; Combined prediction of eigenvectors, residual matrices, and memory activation to conduct group sensing collaborative judgment and generate early warning instructions; The reverse zero-crossing reconstruction method is used to perform bidirectional verification processing on the zero point of the field current signal, including: For each candidate zero-crossing point, the positive value sequence before zero and the negative value sequence after zero are deconstructed, and the negative value sequence is subjected to time reversal and numerical inversion to shape the inversion sequence of the ideal positive value trajectory; Apply interpolation algorithm to the inversion sequence to reconstruct the ideal positive trajectory that is precisely aligned with the positive sequence in time; The comprehensive matching degree between the ideal positive trajectory and the actually observed positive sequence is quantified, and the candidate zero-crossing point is confirmed as the real zero-crossing point only when the comprehensive matching degree exceeds the preset verification threshold.

2. The method according to claim 1, characterized in that Apply interpolation algorithms to the inversion sequence to reconstruct the ideal positive trajectory, including: Set first-order and second-order derivatives as boundary constraints for the inversion sequence; Under boundary constraints, a cubic spline function is constructed for the inversion sequence to perform smooth extension in the time domain. The cubic spline function is sampled at a high density to generate an ideal positive trajectory.

3. The method according to claim 1, characterized in that Quantify the comprehensive match between the ideal positive trajectory and the actual observed positive sequence, including: From the three dimensions of trajectory morphological similarity, numerical deviation amplitude and random noise level, the normalized correlation coefficient, root mean square relative error and signal-to-noise ratio between the ideal positive trajectory and the positive sequence are calculated respectively, and weighted fusion is performed to generate a comprehensive matching degree.

4. The method according to claim 1, wherein Phase space mapping and dynamic feature extraction are performed based on the reconstructed timestamp matrix to obtain the phase space feature matrix, including: Applying cross-unit differential operations to the reconstructed timestamp matrix to generate the original differential sequence that represents the multi-unit temporal relationship; The original difference sequence is mapped to a three-dimensional phase space consisting of the current difference value, the difference value at the previous moment, and the difference change rate, to shape the phase space trajectory of dynamic evolution; The topological invariants of the analytical phase space trajectory are fused into the phase space characteristic matrix after temperature rise drift self-calibration.

5. The method according to claim 4, characterized in that Mapping the original difference sequence into a three-dimensional phase space to shape the phase space trajectory of the dynamic evolution, including: Assign the current difference value and the previous moment difference value in the original difference sequence as the first axis coordinate and the second axis coordinate of the three-dimensional phase space respectively; By performing numerical differentiation operations on the original difference sequence, the difference change rate that represents its change trend is obtained; The differential change rate is filtered and smoothed, and the smoothed differential change rate is used as the third axis coordinate to construct a phase space trajectory.

6. The method according to claim 4, characterized in that Analyze topological invariants of phase space trajectories, including: For the phase space trajectory, the trajectory length, average curvature and gyration radius are quantified respectively from the aspects of its overall path length, local curvature and spatial distribution compactness, and combined into topological invariants.

7. The method according to claim 1, characterized in that Combined prediction of eigenvectors, residual matrices, and memory activation levels to conduct group sensing collaborative judgments and generate early warning instructions, including: Each cascade H-bridge unit is abstracted as a biological individual, and the predicted feature vector and residual matrix are fused to calculate the initial abnormal concentration for each individual; A diffusion propagation model is constructed based on the physical connection between individuals to simulate the spatiotemporal evolution of the initial abnormal concentration between adjacent individuals and form a diffusion concentration field. When the concentration in a local area of ​​the diffusion concentration field exceeds a threshold, a coordinated warning is triggered, thereby generating a warning instruction.

8. The method according to claim 7, characterized in that A diffusion propagation model is constructed based on the physical connection between individuals to simulate the spatiotemporal evolution of the initial abnormal concentration between adjacent individuals, including: The physical connection relationship between individuals is abstracted into a weighted network graph, and a weighted graph Laplacian matrix is ​​constructed for the network graph to characterize the coupling strength and topological structure between individuals. The diffusion partial differential equation is constructed based on the weighted graph Laplace matrix and numerically solved with the initial anomalous concentration as the initial condition to obtain the diffusion concentration field.

9. The method according to claim 1, characterized in that Immune memory prediction is performed based on the main feature matrix to generate predicted feature vectors and memory activation, including: Perform affinity matching between the main feature matrix and the preset fault mode memory library, calculate the similarity between the current system state and each historical fault mode, and obtain the affinity matrix; Based on the affinity matrix, when its value exceeds the activation threshold, the corresponding memory mode is triggered, and the main feature matrix is ​​extrapolated according to the activated memory mode to generate a predicted feature vector; At the same time, the memory activation degree is calculated based on the affinity matrix and the inherent weight of the memory pattern.

Images