Fault early warning method for cascaded H-bridge frequency converter

Abstract

The invention discloses a cascade H-bridge frequency converter fault early warning method, which comprises the following steps of: acquiring field current signals of each H-bridge unit, performing bidirectional verification on signal zero points by adopting a reverse zero-crossover reconstruction method, and generating a reconstruction timestamp matrix; carrying out phase space differential mapping based on the matrix, and extracting a feature matrix capable of representing internal dynamic characteristics of the system by calculating topological invariants such as track length and curvature; after dimension reduction is conducted on the feature matrix, affinity matching is conducted on the feature matrix and a preset fault mode library through an immune memory prediction mechanism, rapid fault recognition is achieved, and prediction feature vectors are generated through extrapolation; and the prediction features, the residual error and the memory activation degree are combined, a quorum sensing collaborative judgment model is introduced, the diffusion process of abnormal signals is simulated, collaborative and robust judgment of the fault type, the occurrence time and the influence range is achieved, and an early warning instruction is generated. According to the method, the accuracy, reliability and advance of fault early warning are remarkably improved.

Description

Technical Field

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

[0002] As a core equipment of modern power electronic systems, the cascaded H-bridge frequency converter plays an increasingly important role in the fields of new energy power generation, motor drive, and power transmission. With the continuous increase in power levels and the increasing complexity of application scenarios, the reliability issue of the cascaded H-bridge frequency converter has become a key bottleneck restricting its large-scale application. Due to the large number of units and complex mutual coupling in the cascaded H-bridge structure, the fault propagation speed is fast and the influence range is wide. The traditional post-maintenance mode can no longer meet the stringent requirements of modern industry for continuous operation. Therefore, developing high-precision and high-timeliness fault warning technologies to achieve the transformation from passive response to active prevention has great theoretical significance and practical value for ensuring the safe and stable operation of power systems, reducing operation and maintenance costs, and improving equipment utilization rates.

[0003] Current fault warning technologies for cascaded H-bridge frequency converters mainly focus on aspects such as voltage harmonic analysis, current amplitude monitoring, and frequency-domain feature extraction. The voltage harmonic analysis method identifies faults by monitoring the change in harmonic content of the output voltage, but this method is insensitive to minor faults and is easily affected by load changes. The current amplitude monitoring technology uses the statistical characteristics of the current of each unit for fault detection, but its time resolution is limited and it is difficult to capture transient fault characteristics at the millisecond level. The frequency-domain analysis method extracts fault characteristic frequencies through FFT transformation, but the computational complexity is high and the processing effect for non-stationary signals is not good. In recent years, some studies have begun to focus on fault diagnosis methods based on machine learning, such as support vector machines and neural networks. These methods show certain advantages in feature extraction and pattern recognition, but generally have problems such as a large demand for training samples and limited generalization ability. In addition, the model-based fault diagnosis method performs residual analysis by establishing a system mathematical model, but the model complexity is high and parameter identification is difficult, facing great challenges in practical engineering applications.

[0004] However, there are still many problems in the existing technical solutions when dealing with specific technical challenges of the cascaded H-bridge system. For example, the instantaneous paradox problem of zero-crossing detection. That is, the traditional zero-crossing detection assumes the existence of an ideal zero point instant, but there are small oscillations and noise interferences in the actual current signal near the zero point, resulting in multiple false zero-crossing triggers. Although the existing threshold filtering methods can suppress some interferences, the threshold setting depends on manual experience and cannot adapt to the dynamic changes of signal characteristics under different working conditions. This problem of finding discrete time points in continuous signals leads to limited time measurement accuracy, which in turn affects the accuracy of subsequent fault feature extraction. Summary of the Invention

[0005] Objective of the invention: To provide a fault warning method for a cascaded H-bridge frequency converter, aiming to solve the above problems existing in the prior art.

[0006] Technical solution: A fault warning method for a cascaded H-bridge frequency converter includes:

[0007] Obtain the on-site current signals of each cascaded H-bridge unit, apply the reverse zero-crossing reconstruction method to perform two-way verification processing on the zero points of the on-site current signals, and generate a reconstructed timestamp matrix;

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

[0009] Based on the main feature matrix, perform immune memory prediction to generate a prediction feature vector and a memory activation degree;

[0010] Combine the prediction feature vector, the residual matrix and the memory activation degree to perform quorum sensing collaborative decision-making and generate a warning instruction.

[0011] According to one aspect of the present application, applying the reverse zero-crossing reconstruction method to perform two-way verification processing on the zero points of the on-site current signals includes:

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

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

[0014] Quantify the comprehensive matching degree between the ideal positive value trajectory and the actually observed positive value sequence, and only when the comprehensive matching degree exceeds a preset verification threshold, confirm the candidate zero-crossing point as a true zero-crossing point.

[0015] According to one aspect of the present application, applying an interpolation algorithm to the inversion sequence to reconstruct an ideal positive value trajectory includes:

[0016] Set the first-order and second-order derivatives of the inversion sequence as boundary constraint conditions;

[0017] Under the boundary constraint conditions, construct a cubic spline function for the inversion sequence to perform smooth extension in the time domain;

[0018] And perform high-density sampling on the cubic spline function to generate an ideal positive value 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] Calculate the normalized correlation coefficient, root mean square relative error, and signal-to-noise ratio between the ideal positive value trajectory and the positive value sequence respectively from the three dimensions of trajectory shape similarity, numerical deviation amplitude, and random noise level, and perform weighted fusion to generate a comprehensive matching degree.

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

[0022] Perform cross-unit difference operation on the reconstructed timestamp matrix to generate an original difference sequence representing the multi-unit time series relationship;

[0023] Map the original difference sequence to a three-dimensional phase space composed of the current difference value, the difference value at the previous moment, and the difference change rate to shape a dynamically evolving phase space trajectory;

[0024] Analyze the topological invariants of the phase space trajectory, and after temperature rise drift self-calibration, fuse them into a phase space feature matrix.

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

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

[0027] Obtain the difference change rate representing its change trend by performing numerical differentiation operation on the original difference sequence;

[0028] Perform filtering and smoothing processing on the difference change rate, and use the smoothed difference change rate as the third-axis coordinate to construct a phase space trajectory.

[0029] According to one aspect of the present application, analyze the topological invariants of the phase space trajectory, including:

[0030] For the phase space trajectory, sequentially quantify the trajectory length, average curvature, and radius of gyration from the aspects of its overall path length, local bending degree, and spatial distribution compactness, and combine them into topological invariants.

[0031] According to one aspect of the present application, jointly predict the feature vector, residual matrix, and memory activation degree, and perform quorum sensing collaborative decision-making to generate a warning instruction, including:

[0032] Abstract each cascaded H-bridge unit as a biological individual, and fuse the predicted feature vector and the residual matrix to calculate the initial abnormal concentration for each individual;

[0033] Construct a diffusion propagation model based on the physical connection relationships among individuals to simulate the spatio-temporal evolution process of the initial abnormal concentration among adjacent individuals, and form a diffusion concentration field;

[0034] When the concentration in a local area of the diffusion concentration field exceeds the threshold, trigger a collaborative warning, thereby generating a warning instruction.

[0035] According to one aspect of the present application, construct a diffusion propagation model based on the physical connection relationships among individuals to simulate the spatio-temporal evolution process of the initial abnormal concentration among adjacent individuals, including:

[0036] Abstract the physical connection relationships among individuals into a weighted network graph, and construct a weighted graph Laplacian matrix for the network graph to characterize the coupling strength and topological structure among individuals;

[0037] Based on the weighted graph Laplacian matrix, construct a diffusion partial differential equation, and perform numerical solution with the initial abnormal concentration as the initial condition, thereby obtaining a diffusion concentration field.

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

[0039] Match the affinity between the principal feature matrix and a preset fault mode memory bank, calculate the similarity degree between the current system state and each historical fault mode, and obtain an affinity matrix;

[0040] Based on the affinity matrix, when its value exceeds the activation threshold, trigger the corresponding memory mode, and perform feature extrapolation on the principal feature matrix according to the activated memory mode to generate a prediction feature vector;

[0041] At the same time, calculate the memory activation degree according to the inherent weights of the affinity matrix and the memory mode.

[0042] According to one aspect of the present application, match the affinity between the principal feature matrix and a preset fault mode memory bank to obtain an affinity matrix, including:

[0043] By fusing the Euclidean distance, Mahalanobis distance, and cosine similarity distance, calculate the comprehensive distance between the principal feature matrix and each memory mode in the fault mode memory bank;

[0044] Estimate an adaptive variance parameter that takes into account both intra-class compactness and inter-class separation for each memory mode;

[0045] Based on the comprehensive distance and the adaptive variance parameter, generate an affinity matrix through exponential function transformation.

[0046] According to one aspect of the present application, perform feature extrapolation on the principal feature matrix according to the activated memory mode to generate a prediction feature vector, including:

[0047] Taking the memory activation degree as the weight, perform a weighted sum of multiple direction vectors of the current system state pointing to each activated memory pattern to determine the comprehensive weighted prediction direction;

[0048] Adaptively adjust the prediction step size according to the intensity of the memory activation degree;

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

[0050] Beneficial effects: Fundamentally solve the instantaneous paradox problem of zero-crossing detection in the prior art through the reverse zero-crossing reconstruction theory. 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 of process reconstruction. It establishes a time-reversal model, uses the negative value sequence with less interference in the second half period of the zero point, and reversely interpolates and reconstructs the theoretically perfect positive value change process in the first half period. Then, calculate 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 10 ps level of the traditional method to the 2 ps level, enabling the system to capture the weak temporal 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 analyses, thereby significantly improving the sensitivity and reliability of fault warning as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flowchart of the present invention.

[0052] Figure 2 is a flowchart of the present invention for performing two-way verification processing on the zero point of the on-site current signal.

[0053] Figure 3 is a flowchart of the present invention for reconstructing an ideal positive value trajectory.

[0054] Figure 4 is a flowchart of the present invention for obtaining a phase space feature matrix. DETAILED DESCRIPTION OF THE INVENTION

[0055] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the following will be combined with Figures 1 to 4 Specific embodiments of the present invention will be 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 protection scope of the present invention.

[0056] In addition to the above problems, there is mainly the problem of loss of phase structure 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 inherent dynamic structure of the time series, converts 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. Especially for the fault modes of cascade systems with complex spatio-temporal coupling characteristics, traditional linear differential methods cannot effectively extract their non-linear dynamic characteristics, limiting the accuracy and lead time of fault warning.

[0057] Example 1: Describe the process of obtaining high-precision timestamps based on reverse zero-crossing reconstruction. Through a two-way verification method, accurately determine the zero-crossing point of the current signal to provide a high-precision time reference for fault feature extraction.

[0058] This example mainly includes the following steps:

[0059] Step S101, apply the reverse zero-crossing reconstruction method to perform two-way verification processing on the zero point of the on-site current signal to generate a reconstructed timestamp matrix.

[0060] In this example, to solve the problem that the traditional zero-crossing detection method is prone to generate false trigger points and has low accuracy under noise interference. This method utilizes the symmetry of the sine alternating current in the first half cycle before and after the zero-crossing point, uses the waveform of one half cycle to reverse reconstruct the ideal waveform of the other half cycle, and then compares the reconstructed waveform with the actual observed waveform. Only when they match highly is it confirmed as the true zero-crossing point.

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

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

[0063] Among them, the candidate zero-crossing point is the signal polarity change point preliminarily detected by the traditional comparator method. Time reversal and numerical inversion are to reverse the negative value sequence after the zero point in the time axis and take the absolute value of all its values, so as to obtain a sequence that is similar in shape to the positive value sequence before the zero point.

[0064] Obtain the synchronized current signal I sync (t) from the current transformers of each H-bridge unit. For a candidate zero-crossing point detected at time t i , separate the positive value sequence I pos of its first half cycle (for example, the first 5 sampling points) =I(t i -5:t i-1) and the negative value sequence I in the second half cycle (e.g., 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] Since an ideal noise-free signal should have good symmetry around the zero-crossing point. Through the inversion operation, an ideal first half cycle sequence based on the second half cycle information is created. This provides a benchmark for subsequent cross-validation and lays a foundation for the identification of noise and small perturbations.

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

[0067] Preferably, set the first and second derivatives of the inversion sequence as boundary constraint conditions; under the boundary constraint conditions, construct a cubic spline function for the inversion sequence to perform smooth extension in the time domain; and perform high-density sampling on the cubic spline function to generate the ideal positive value trajectory.

[0068] In this embodiment, first, calculate the first and second derivatives of the inversion sequence I rev at both ends as the boundary conditions for spline interpolation. Then, use these boundary conditions to construct a piecewise cubic spline function S rev (t). Finally, within the time interval corresponding to the original positive value sequence, perform high-density sampling (e.g., 10 times the original sampling rate) on S rev (t) to obtain the reconstructed ideal positive value 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 order of interpolation and the sampling density can be adjusted according to the requirements for calculation accuracy and efficiency.

[0070] Since directly comparing using the discrete inversion sequence will have problems with inaccurate time alignment. Through interpolation for smooth extension and high-density sampling, an ideal trajectory in the continuous time domain can be generated, enabling precise comparison with the original, discrete positive value sequence at any time point. This improves the accuracy of subsequent matching degree calculations.

[0071] In the third step, quantify the comprehensive matching degree between the ideal positive value trajectory and the actually observed positive value sequence, and only when this comprehensive matching degree exceeds the preset verification threshold, confirm the candidate zero-crossing point as a true zero-crossing point.

[0072] Preferably, the normalized correlation coefficient, root mean square relative error, and signal-to-noise ratio between the ideal positive value trajectory and the positive value sequence are calculated respectively from three dimensions: trajectory shape similarity, numerical deviation magnitude, and random noise level, and weighted and fused to generate a comprehensive matching degree.

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

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

[0075] It is found that a single index (such as only using the correlation coefficient) is easily affected by specific types of noise. For example, in-phase DC bias will seriously affect the root mean square error, but has little effect on the correlation coefficient. Combining indicators from multiple dimensions can more robustly and comprehensively evaluate the similarity of two sequences, effectively exclude pseudo zero crossing points caused by glitches, oscillations, etc., and improve the accuracy of the time stamp from the 10 ps level of traditional methods to the 2 ps level.

[0076] Through the above steps, the current signals of all H-bridge units are processed, and finally a high-precision, verified reconstructed time stamp matrix T recon [N×K] is obtained, 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 techniques are based on the principle of instantaneous comparison, that is, finding 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: the theoretically zero crossing should be an infinitesimal time point, while in actual signals, there must be noise, glitches, and tiny oscillations near zero, resulting in multiple pseudo-crossings in the zero neighborhood. Existing technologies suppress interference by setting hysteresis thresholds or digital filtering, but these methods essentially forcefully determine a discrete time point in a continuous and noise-filled signal, and their accuracy is limited by the noise level and the subjectivity of threshold setting. The present invention proposes a method of process reconstruction verification: instead of trying to find the ideal zero moment in noise, it utilizes the time-reversal invariance of the ideal sine signal (i.e., the mathematical property of I(t) = -I(-t)), and through the waveform data in the second half period of the zero point, strictly reconstructs the ideal noise-free trajectory in the first half period of the zero point mathematically, and then conducts multi-dimensional similarity verification between this reconstructed trajectory and the actual observed trajectory. This self-referencing and two-way verification mechanism transforms the time point detection problem into a waveform matching problem, eliminates the influence of noise on the time measurement accuracy, and realizes an accuracy improvement from the 10 ps level to the 2 ps level, laying a technical foundation for the precise capture of microsecond-level timing jitter between units in a cascaded H-bridge system.

[0078] In addition, the field of power system fault warning has long faced the technical challenge that early fault characteristics are weak and easily masked. Traditional methods based on frequency-domain analysis and statistical feature extraction can often only capture the significant changes after a fault occurs, and lack sensitivity to the weak signs in the fault germination period. The present invention introduces phase space topology analysis technology, realizing the transformation from surface numerical analysis to internal dynamics structure analysis. Specifically, traditional differential jitter analysis converts the time series Δt into a scalar difference matrix, which can eliminate common-mode interference, but 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 of [Δt(k), Δt(k - 1), dΔt / dk], expands the one-dimensional time series information into a high-dimensional geometric object, that is, the phase space trajectory, and the geometric form of this trajectory completely retains all the dynamic characteristics of the interaction between multiple units of the system. More importantly, the topological invariants (trajectory length L traj , average curvature k avg , radius of gyration R gyrationIt has invariance to geometric transformations and robustness to noise, and can identify subtle changes in the system's dynamic behavior from seemingly normal operating data. For example, the slight phase lag caused by the early aging of IGBTs can hardly be detected in traditional amplitude or harmonic analysis, but in the phase space, it will be significantly manifested as the distortion of the trajectory shape and the increase of 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 limitation of traditional methods, realizes the early identification of gradual and hidden faults, advances the warning time window from the ex-post response after the fault occurs to the ex-ante prediction in the budding period of the fault, and provides a new technical path for the preventive maintenance of power systems.

[0079] Embodiment 2: Based on the reconstructed timestamp matrix Trecon generated in Embodiment 1, this embodiment describes the process of extracting dynamic features based on phase space mapping. The core of this step is to map the one-dimensional timestamp difference sequence to a high-dimensional phase space, and extract the system dynamic features that cannot be obtained by traditional difference methods by analyzing the topological structure of its trajectory.

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

[0081] In this embodiment, this step mainly solves the problem that the 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] The first step: Perform cross-cell difference operation on the reconstructed timestamp matrix to generate an original difference sequence representing the multi-cell time series 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 cells, in-phase cells), perform cross-cell subtraction: Δt ij =T recon (i)-T recon (j). For example, for a system with N cells, N - 1 adjacent cell pairs and N - 2 cross-phase cell pairs can be selected, and a total of P = 2N - 3 difference sequences are obtained. Finally, output the original difference sequence matrix Δt raw [P×K].

[0085] In this embodiment, the common-mode noise and system-level time-base drift are eliminated through the difference operation, thus highlighting the relative time series differences between cells. These differences often contain signs of early faults.

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

[0087] Preferably, assign the current difference value and the difference value at the previous moment in the original difference sequence as the first-axis coordinate and the second-axis coordinate of the three-dimensional phase space respectively; obtain the difference change rate characterizing its change trend through numerical differentiation operation on the original difference sequence; perform filtering and smoothing processing on the difference change rate, and use the smoothed difference change rate as the third-axis coordinate to construct a phase space trajectory.

[0088] In one embodiment, the X-axis is the current difference value Δt(k). The Y-axis is the difference value Δt(k - 1) at the previous moment. The Z-axis is the difference change rate dΔt / dk. The difference change rate can be calculated by numerical methods such as central difference, and a moving average filter is applied for smoothing 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 can unfold the dynamic information hidden in the one-dimensional time series in a higher-dimensional space. The X-axis and Y-axis characterize the state of the system, and the Z-axis characterizes the change trend of the state. The trajectory jointly formed by the three can intuitively reflect the change of the system's dynamic behavior, such as from stable (the trajectory converges to a point or a small area) to abnormal (the trajectory diverges or the form changes drastically).

[0091] In the third step, analyze the topological invariants of the phase space trajectory, and after temperature rise drift self-calibration, fuse them into a phase space feature matrix.

[0092] Preferably, the calculation process can be as follows: for the phase space trajectory, sequentially quantify the trajectory length, average curvature, and radius of gyration from aspects of its overall path length, local bending degree, and spatial distribution compactness, and combine them into topological invariants.

[0093] Among them, the trajectory length L traj is the integral of the lengths of each small line segment along the trajectory, used to reflect the cumulative total amount of difference jitter. The average curvature κ avg is the average value of calculating the local curvature of each point on the trajectory, used to reflect the degree of sharp bending of the trajectory. The radius of gyration R gyration is the average distance from all points on the trajectory to its geometric centroid, used to reflect the distribution range and compactness of the trajectory in space.

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

[0095] Average curvature κ avg = mean(κ loca l[k]), where the local curvature κ local is calculated by the cross product of the 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] Combine these three invariants into a feature vector [L traj , κ avg , R gyration , and obtain the topological feature matrix Topo[P×3].

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

[0099] Since topological invariants are a high-level generalization of the geometric shape of the trajectory, compared with the trajectory itself, they are less sensitive to noise and can better reflect the inherent and essential dynamic properties of the system. For example, an incipient insulation aging fault may not significantly change the mean value of the difference, but will cause an increase in its volatility, which will be directly reflected in the increase in the trajectory length and the radius of gyration.

[0100] Before fusion, utilize the symmetry of the phase space trajectory to calculate the drift vector of the trajectory centroid and perform translational correction on the coordinates to eliminate the systematic measurement drift caused by temperature changes. Finally, output the calibrated phase space feature matrix Φ phase [N×M×3].

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

[0102] Example 3: Describe the process of performing fault mode matching based on the principal feature matrix P obtained after feature dimensionality reduction. By borrowing the memory and recognition mechanisms of the biological immune system, rapid matching of fault symptoms and early prediction of future states are achieved.

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

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

[0105] First step: Perform affinity matching between the main feature matrix and a preset fault mode memory bank, calculate the similarity between the current system state and each historical fault mode, and obtain an affinity matrix.

[0106] The fault mode memory bank is a pre-constructed database that stores the typical feature vectors of various known fault types, similar to antibodies in the immune system. Affinity is used to quantify the matching degree between the current system state (antigen) and these antibodies.

[0107] Preferably, by fusing the Euclidean distance, Mahalanobis distance, and cosine similarity distance, calculate the comprehensive distance between the main feature matrix and each memory mode in the fault mode memory bank; estimate an adaptive variance parameter that takes into account both intra-class compactness and inter-class separation for each memory mode; based on the comprehensive distance and the adaptive variance parameter, generate the affinity matrix through exponential function transformation.

[0108] First, for each fault mode k, calculate the adaptive variance σ by analyzing the intra-class dispersion of its historical samples and the inter-class distance from other modes k . This makes the affinity calculation adaptive to the feature distributions of different fault modes.

[0109] Calculate the comprehensive distance D combined =w d1 ·d eucl +w d2 ·d maha +w d3 ·d cos ; The multi-scale measurement method is more robust than a single distance.

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

[0111] Through this step, the immune recognition process is simulated. Among them, the affinity function can accurately quantify the possibility of the current system state evolving into various known fault modes, providing a decision-making 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 feature extrapolation is performed on the main feature matrix according to the activated memory pattern to generate a predicted feature vector.

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

[0114] When the maximum value of the affinity A(P, m k ) of a certain memory pattern k exceeds an adaptively adjusted activation threshold θ act , this pattern is activated. The activation degree α k is calculated according to the affinity size and the inherent weight of this pattern.

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

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

[0117] In this embodiment, once a familiar antigen (fault symptom) is recognized, the system can quickly mobilize memory to predict its future development direction and speed. It is more targeted and has a faster prediction speed than a pure autoregressive model based on time series, and is especially suitable for fault types with existing empirical knowledge.

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

[0119] As above, the memory activation degree α[K×1] is a vector calculated by comprehensively considering factors such as the affinity size and the inherent weight W mem of the memory pattern. Each element αk represents the intensity at which the k-th fault mode is activated. This activation degree 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 features in a future period (e.g., 300 ms) and quantify the credibility of this prediction (memory activation degree), achieving a leap from diagnosis to prediction.

[0121] Example 4: Describe the collaborative decision-making process based on quorum sensing using the above-mentioned predicted feature vector Ppred, residual matrix R, and memory activation degree α. In this implementation, multiple H-bridge units are regarded as a whole, and by simulating the diffusion process of abnormal signals, collaborative and robust decision-making on faults is achieved.

[0122] Step S401: Combine the predicted feature vector, residual matrix, and memory activation degree to perform quorum sensing collaborative decision-making and generate a warning instruction.

[0123] In this embodiment, this step aims to solve the problem that the decision-making of a single unit is vulnerable to noise interference and cannot accurately evaluate the scope of the 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] First step: Abstract each cascaded H-bridge unit as a biological individual, and fuse the predicted feature vector and the residual matrix to calculate the initial abnormal concentration for each individual. A biological individual is a bionic abstraction of each H-bridge unit. The initial abnormal concentration is a quantitative index used to represent the degree of abnormality of each individual at the current moment.

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

[0127] By fusing two different types of abnormal information. Ppred represents an interpretable anomaly similar to historical failures, while R represents a newly emerging anomaly that is different from all known patterns. Combining the two can more comprehensively evaluate the health status of each unit.

[0128] In the second step, a diffusion propagation model is constructed based on the physical connection relationships between individuals to simulate the spatio-temporal evolution process of the initial anomaly concentration among adjacent individuals, forming a diffusion concentration field.

[0129] Preferably, the physical connection relationships between individuals are 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; a diffusion partial differential equation is constructed based on the weighted graph Laplacian matrix, and numerical solutions are carried out with the initial anomaly concentration as the initial condition, so as to obtain the diffusion concentration field.

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

[0131] First, an adjacency matrix is established according to the physical adjacency relationships of the H-bridge units. Then, considering the proximity of 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 . It is used to describe the difficulty and path of the abnormal signal propagation between different units.

[0132] Construct a partial differential equation in the form of dc / dt = D▽ 2 c + S(x, t) - γc. Wherein, the Laplace operator ▽ 2 c is approximated by the graph Laplacian matrix L weighted , the source term S is the initial anomaly concentration C local , D is the diffusion coefficient, and γ is the attenuation coefficient. Stable numerical methods such as Crank-Nicolson are used to solve this equation to obtain the spatio-temporal evolution matrix C of the anomaly concentration in the entire H-bridge system over a future period of time diff [N×T pred .

[0133] By this step, the diffusion process of signal molecules in the bacterial population is simulated. The abnormal signal generated by a certain unit will diffuse to its neighbors, forming a concentration field. This mechanism has the functions of spatial smoothing and filtering. The abnormal concentration generated by the instantaneous noise perturbation of a single unit will be averaged out due to diffusion, while the abnormal concentration generated by persistent and real fault sources will continuously accumulate and diffuse, forming obvious local high-concentration regions, enhancing the robustness of the decision.

[0134] In some embodiments, the calculation formula is the abnormal concentration c of the i-th H-bridge unit at the t + 1 moment 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 abnormal concentration of the i-th H-bridge unit at time t. Δτ is the discrete time step in numerical solution. D is the diffusion coefficient. L ij is the element in the weighted graph Laplacian matrix that characterizes the connection relationship between units i and j. N(i) is the set of neighbor units of unit i. S i t is the abnormal concentration source term (i.e., the initial abnormal concentration) of unit i at time t. 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 collaborative warning is triggered, thereby generating a warning instruction.

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

[0137] Once the warning is triggered, the system will integrate information and generate a structured warning instruction. The instruction content includes: the type of fault (determined according to the activated memory pattern α), the predicted occurrence time (current time + predicted lead time), and the list of affected units (i.e., the area where the concentration exceeds the threshold). This instruction is finally written into the upper-layer monitoring system such as SCADA for the operation personnel to make decisions.

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

[0139] Example 5: Describe the complete process of the fault warning method for the cascaded H-bridge inverter. Show an end-to-end application from data collection to warning output. It should be noted that the content already described in detail in the above Examples 1 to 4 will not be elaborated here.

[0140] In a typical application scenario, such as a photovoltaic inverter system containing 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 the zero-crossing candidate points in each signal cycle, the reverse zero-crossing reconstruction method is applied for two-way verification. For example, for a candidate point, cubic spline interpolation is used to reconstruct the ideal trajectory, and the comprehensive matching degree M between it and the actual trajectory is calculated total . Only when M total > 0.95, a 10-ps-level TDC is used to record its exact moment. Finally, a reconstructed timestamp matrix T recon [30×K] is formed.

[0142] Read T recon , and calculate the difference sequence Δt between adjacent units raw . Map each Δt raw to the three-dimensional phase space of [Δt(k), Δt(k - 1), dΔt / dk] to obtain a dynamic trajectory. Calculate the three topological invariants of the length, average curvature, and radius of gyration of each trajectory, and perform temperature rise calibration to form a phase space feature matrix Φ phase .

[0143] Perform sparsification processing and sliding window SVD decomposition on Φ phase to obtain a principal feature matrix P[dxM] and a residual matrix R[NxM]. Input P into the immune memory prediction module and perform affinity matching with the fault memory library. Assume that the current P has the highest affinity with the IGBT open-circuit fault memory pattern and exceeds the activation threshold, then this pattern is activated. The system then extrapolates the features according to this activated pattern to generate a predicted feature vector P pred for the next 300 ms, and at the same time outputs a relatively high memory activation degree α k for the IGBT open-circuit fault.

[0144] The system fuses P pred and R, and calculates the initial abnormal concentration C of each unit local . Based on the physical connection relationship between units, a graph Laplacian matrix is constructed, and the diffusion process of C local in the entire system is simulated by solving the diffusion equation to obtain a diffusion concentration field C diff . If it is found that the concentrations of the 5th, 6th, and 7th units and their adjacent regions continuously exceed the population threshold θ group during the prediction time period, then a collaborative early warning is triggered.

[0145] The system generates an early warning instruction W, the content of which is: Fault type: IGBT open circuit fault, Predicted occurrence time: [Current time + 300ms], Affected units: Unit 5, 6, and 7. This instruction is sent to the monitoring center to guide the maintenance personnel to check or isolate the relevant units in advance, thus avoiding an unplanned shutdown.

[0146] Through the above complete implementation process, it is possible to issue an early warning with high reliability and high precision hundreds of milliseconds before the fault occurs, and specify the nature and scope of the fault, demonstrating great engineering application value.

[0147] Example 6. Describe the process of reducing the dimension of the high-dimensional phase space feature matrix, extracting its main components and quantifying the residual information.

[0148] This process continues with the phase space feature matrix Φ generated in Example 2 phase , and specifically includes the following steps:

[0149] Step S301, perform sparse dimensionality reduction on the phase space feature matrix. Among them, the sparse processing means that on the premise of retaining key information, redundant or less informative entries in the high-dimensional feature matrix are discarded, thereby reducing the dimension and computational complexity of the matrix.

[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 physically adjacent or electrically phase-identical units, so in this step, according to the physical topology of the H-bridge, only the feature entries corresponding to these strongly correlated units are retained. Subsequently, the three-dimensional phase space features [L traj , κ avg , R gyration retained for each unit are flattened and combined into a large two-dimensional matrix.

[0151] For a system with N units, if all pairwise interactions between units are considered, the feature dimension will reach O(N2) or even higher, resulting in a huge computational burden. Through the sparse processing based on the physical topology, the feature dimension can be significantly reduced from the O(N2) or O(N3) level to the O(N) level, while retaining most of the key fault information, greatly improving the efficiency of subsequent processing and making it applicable to 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 can decompose a matrix into the product of three matrices and identify the most prominent patterns or components in the data. The sliding window moves over the data in a window of fixed length and analyzes it segment by segment.

[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 within the first window, i.e., Φ sparse_window1 =U0Σ0V0 T . Among them, the diagonal elements of the matrix Σ0 are the singular values, arranged from largest to smallest, representing the importance of different patterns; U0 and V0 are the left and right singular vector matrices, representing the specific forms of these patterns.

[0156] When the window slides forward by one time step, to avoid re - performing the computationally expensive complete SVD decomposition, preferably, the Rank - 1 correction algorithm is used to quickly update the SVD decomposition result of the previous moment. This algorithm slightly corrects the original U, Σ, and V matrices using the data newly entered into the window and the data removed from the window, and the computational efficiency is much higher than the full - scale decomposition.

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

[0158] SVD can decouple the mixed system dynamic signals into a series of orthogonal principal modes. Usually, a small number of principal modes (corresponding to the largest several singular values) are sufficient to capture more than 99% of the dynamic behavior of the system. This method not only effectively compresses the data dimension, but more importantly, it separates the information into the following matrices:

[0159] The main feature matrix P[d×M], which is composed of the first d largest singular values and their corresponding singular vectors, is used to capture the current most prominent and energy - concentrated 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], which is composed of the remaining smaller singular values and their corresponding components, is used to represent the weak or newly emerging signal components ignored by the principal modes. This matrix is of great value for detecting early faults of unknown types or subtle changes in signals.

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

[0162] Embodiment 7 is used to describe in detail how the fault pattern memory bank performs adaptive weight update and decay, so that the entire early warning system has the ability of continuous learning and self-optimization. On the basis of Embodiment 3, this solution further includes:

[0163] Step S404: Perform weight update and decay on the fault feature memory bank. Memory update and decay is a computational mechanism that simulates the proliferation and apoptosis of memory B cells in the biological immune system. It strengthens the memories that have been proven effective, while gradually forgetting the memories that have been ineffective or incorrect for a long time.

[0164] It is executed after the system completes an early warning judgment and obtains the feedback of the actual operation (for example, confirming the occurrence of a fault through the SCADA system or on-site verification by maintenance personnel). The input is the memory activation degree α[K×1] calculated in Embodiment 3 and the verification result of the current fault.

[0165] For the memory pattern k activated in this early warning (i.e., α k >0), if the predicted fault is finally verified to be true, 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]; where W mem_new [k] is the updated weight. γ is the memory decay coefficient, and its value is between 0 and 1. For example, it can be set to 0.9. The meaning of this formula is that the new weight is the weighted average of the old weight and the new evidence (activation degree α k ), and γ controls the degree of retention of historical memories.

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

[0167] The decay coefficient γ can be adjusted according to the stability of the system and the speed of environmental changes. In a relatively stable system, γ can take a larger value (such as 0.95) to place more trust in historical experience. In a system with drastic changes in working conditions, γ can be appropriately reduced (such as 0.85) to adapt to new patterns faster.

[0168] This step introduces a feedback learning closed-loop into the early warning system, enabling high-quality memory patterns that can accurately predict faults to have a greater say in future decision-making, thereby improving the early warning sensitivity and accuracy for similar faults. Weight decay can gradually eliminate those memory patterns that have become outdated (e.g., the fault pattern has changed after equipment maintenance) or incorrect, preventing them from interfering with future judgments and keeping the memory bank clean and efficient.

[0169] Through a dynamic and adaptive update mechanism, the fault memory bank is no longer a static rule set, but a living knowledge base that can grow and evolve with the equipment. This enhances the robustness and adaptive ability of the entire fault early warning method during long-term operation, ensuring its continuous effectiveness under various complex working conditions.

[0170] Example 8: Describe the relevant 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 sine signal. For an ideal current signal I ideal (t)=Asin(ωt), it has odd symmetry near the zero point (t = 0), that is, it satisfies: I ideal (-t)=-I ideal (t). This property is the time-reversal invariance. It shows that reversing the signal on the time axis (t → -t) is equivalent to taking the opposite of its amplitude. Therefore, the ideal waveform on the other side can be strictly deduced using the waveform information on one side of the zero point through this invariant relationship.

[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 deviation degree between the real signal and this ideal model, the pseudo zero-crossings caused by noise are identified and removed in reverse, and the moments of real zero-crossings are reconstructed with high precision.

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

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

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

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

[0178] According to the time - reversal invariance model, use the negative value sequence N actual after the zero point to construct an ideal positive value sequence that should theoretically appear before the zero point. For each point (t actual , I j , I j ) in N

[0179] Perform the following transformations: j Time reversal t j ' = t0 - (t j - t0) = 2t0 - t

[0180] Numerical inversion I j ' = -I j Thus, a new discrete point set is obtained, namely the inversion sequence: Prev = {(t j ', I j ')}. In the ideal case, P rev should coincide exactly with P actual .

[0181] To achieve an accurate comparison, the discrete inversion sequence Prev needs to be converted into a continuous function. In this embodiment, the cubic spline interpolation method is used to construct this ideal trajectory. Take Prev as the interpolation nodes and construct a piece - wise cubic polynomial function S ideal (t). For any interval [t j ', t j+1 '], the function form is: S ideal (t) = a j (t - t j ')3 + b j (t - t j ')2 + c j (t - t j ') + d j ; where the coefficients {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 the receptive field width.

[0190] If σ kfault is too large, it will result in an overly wide receptive field, causing this fault mode to have a high affinity for dissimilar features and thus triggering false alarms.

[0191] If σ kfault is too small, it will result in an overly narrow receptive field, making this fault mode insensitive to some of its own normal variants and thus triggering missed alarms.

[0192] Therefore, it is crucial to estimate an optimal and adaptive σ fault for each fault type k kfault .

[0193] The goal of the algorithm is to make the estimated variance parameter balance the within-class compactness and between-class separability, as follows:

[0194] Fault mode memory bank: M lib ={m1, m2,..., m Kfault}, where m kfault is the central feature vector (i.e., antibody) of the kfault-th class of faults.

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

[0196] Calculate the within-class compactness σ 2 intra to measure the dispersion of the samples of a certain class of faults in their own distribution.

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

[0198] The smaller this value, the more stable and concentrated the feature representation of this class of faults.

[0199] Calculate the between-class separability σ2 inter , which is 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 and the distance from the centers m j of all other fault categories (where j = kfault), and take the minimum value. σ 2 inter (kfault) = min j≠kfault (||m kfault - m j || 2 ). The larger this value is, the higher the feature recognition ability of this type of fault, and the less likely it is to be confused with other faults.

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

[0201] Adopt a weighted fusion method to combine the within-class compactness and between-class separability.

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

[0203] When the characteristic variation range of the fault category itself is very large (σ 2 intra is large), α bal can be appropriately increased (such as 0.7) so that σ kfaul t can better cover these samples.

[0204] When the distance between the fault category and other classes in the feature space is very close (σ 2 inter is small), α bal needs to be appropriately reduced (such as 0.3) so that σ kfault becomes smaller, forming a sharper receptive field to avoid misidentifying samples of neighboring classes.

[0205] Through the above steps, instead of using a globally unified variance, a unique variance parameter σ kfaultThis adaptive parameter estimation method can significantly improve the accuracy and robustness of fault affinity matching.

[0206] Example Ten describes the specific process of temperature rise drift detection, which is used to eliminate systematic measurement drift caused by temperature changes in sensors or acquisition circuits before performing topological invariant analysis.

[0207] The intrinsic dynamic characteristics of the system (i.e., the shape of the trajectory) should not change with temperature under steady-state operating conditions, and temperature changes mainly cause the entire trajectory to translate in the 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] Execute this step when the frequency converter system is in a healthy and stable operating state and the operating temperature is at a known reference base temperature (e.g., 25°C, or the thermal equilibrium temperature after long-term stable operation).

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

[0210] For each trajectory i, calculate its geometric centroid C base [i, :]. The centroid is the coordinate mean of all points on the trajectory.

[0211] The calculation formula for the drift vector (benchmark centroid 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) are stored in the non-volatile memory as the reference for subsequent calibration.

[0217] This step is periodically executed during the online operation of the frequency converter system.

[0218] Obtain the real-time phase space trajectory data X current [i,k time ,:].

[0219] Using exactly the same method as the first step, calculate the real-time centroid C current [i,:].

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

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

[0222] After calculating the drift vector, immediately correct the current trajectory data.

[0223] Subtract the drift vector corresponding to the current 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 trajectory data X corrected after correction is output to the subsequent topological invariant calculation module.

[0226] It can dynamically and 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 the sensitivity to real faults.

[0227] According to one aspect of the present application, achieving multi-channel synchronization at the nanosecond or picosecond level is a mature technology in fields such as test and measurement, communication, and power. Its implementation typically uses a commercial GPSDO (GPS-disciplined oven-controlled crystal oscillator) module, which receives the atomic clock signal of GPS satellites and outputs a high-precision 1PPS (pulses per second) signal and a 10MHz frequency reference. The acquisition units of each H-bridge receive these synchronization signals through optical fibers or dedicated coaxial cables and use this signal as a reference for triggering or phase-locked loops, thereby achieving time-base synchronization of each channel. This belongs to the conventional engineering practice of those skilled in the art.

[0228] According to one aspect of the present application, managing and distributing the input reference clock inside the FPGA is a standard process in digital logic design. Mainstream FPGA manufacturers (such as Xilinx, Intel / Altera) all provide powerful and silicon-verified PLL (phase-locked loop) or MMCM (mixed-mode clock manager) hard IP cores. Designers can easily configure parameters such as frequency multiplication, frequency division, and phase shift of the IP core through a graphical configuration interface or instantiation code to generate various clock signals required by the system. The FPGA development environment will automatically complete the subsequent clock tree synthesis and routing, which also belongs to the well-known technology in this field.

[0229] According to one aspect of the present application, implementing a high-precision time-to-digital converter (TDC) and its calibration in the FPGA has a large number of publicly available literatures and mature methods. The Vernier delay line mentioned in the article is a classic TDC architecture. Its calibration method usually uses code density testing, that is, by inputting a signal with a known random distribution, counting the hit times of each time quantization step (bin), thereby measuring and correcting the differential nonlinearity (DNL) and integral nonlinearity (INL) of the TDC. Accuracy verification can be completed through a comparison test 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, the setting and tuning of threshold parameters are standard steps. The methods mainly include:

[0231] Statistical analysis method: Collect a large amount of data under normal operating conditions, calculate the mean (μ) and standard deviation (σ) of relevant features, and set 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, plot the relationship curve (ROC curve) of the true positive rate (sensitivity) and false positive rate (1 - specificity) at different thresholds, and select the best operating point on the curve as the final threshold according to the tolerance of false alarms and missed alarms.

[0233] Empirical setting and online fine-tuning. At the initial stage of system deployment, a conservative initial value is set according to experience, and then during actual operation, the threshold is fine-tuned online and continuously based on the early warning results and the feedback from operation and maintenance personnel. Those skilled in the art are fully capable of using one or more of the above methods to find appropriate setting values for each threshold in the present invention.

[0234] In summary, through the reverse zero-crossing reconstruction algorithm, unprecedented accuracy and reliability are achieved at the source of fault early warning, the time measurement link. By establishing a time-reversal invariance model: that is, using the current waveform data in the second half period after zero point (usually less affected by disturbances such as arcs), Ineg through time reversal and numerical inversion, a noise-free and theoretically ideal trajectory of the first half period is reconstructed mathematically. Ipos_ideal . Subsequently, the multi-dimensional matching degree between this reconstructed ideal trajectory and the actually observed first half period trajectory I pos is calculated. This self-referencing and two-way verification algorithm transforms the problem of finding discrete time points into the problem of evaluating the similarity between two curves, thus fundamentally solving the technical pain point that the traditional comparator method is vulnerable to noise and glitches and generates false zero crossings. In the specific scenario of a cascaded H-bridge frequency converter with high-frequency switching and strong electromagnetic interference, this ability to obtain high-precision timestamps (up to picosecond level) enables the system to capture the microsecond-level or even nanosecond-level timing anomalies between units caused by early aging of IGBTs, weak jitter of drive signals, etc., which were previously difficult to detect, providing a high-quality data basis for subsequent accurate fault feature extraction, which is an advantage that cannot be compared with traditional amplitude or harmonic analysis methods.

[0235] By mapping the timestamp difference sequence Δt to the three-dimensional phase space of [Δt(k), Δt(k - 1), dΔt / dk] and analyzing the topological invariants of its trajectory, the present invention solves the core problem that traditional differential analysis methods will destroy the internal dynamic structure of the time series and cause information loss. The acquisition of this technical effect stems from the creative transformation of one-dimensional and discrete time series difference information into a geometric object in a high-dimensional space - the phase space trajectory. The shape, length, curvature, and spatial distribution range of this trajectory intuitively and completely retain all the dynamic information of the interaction between system units. For example, in the application scenario of cascaded H-bridge, the slow aging of the capacitor of a certain power unit may not immediately cause a significant change in the amplitude of the output current, but will cause a slight phase lag and irregular fluctuations in its dynamic response. This change in dynamic characteristics may be submerged in traditional differential mean or variance statistics, but in the phase space, it will clearly be manifested as the distortion of the trajectory shape, the increase in the trajectory length (Ltraj), and the 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 characterizing the deterioration of the system health state from seemingly normal operation data, realizing the accurate identification of early and slow-changing faults, and greatly improving the sensitivity and depth of early warning.

[0236] By introducing an immune memory prediction mechanism that draws on the principles of biological immunology, the present invention realizes the leap from passive diagnosis to active prediction and obtains a valuable early warning lead for faults. The realization of this technical effect is the deep integration of business rules (constructing an expert knowledge base using historical fault data) and specific algorithms (affinity matching and feature extrapolation). Specifically, the system stores various fault characteristics that have occurred and been verified in history (such as IGBT open circuit, bus capacitor failure) as antibodies in the fault mode memory bank. During real-time monitoring, by calculating the affinity between the main eigenvector P of the current system and each antibody m_k in the bank, millisecond-level fast pattern matching of fault symptoms is achieved. Once the affinity exceeds the activation threshold, the system can immediately identify the potential fault type and, based on the fault evolution direction information contained in the activated memory pattern, perform forward extrapolation on the current feature to predict the state in the next few hundred milliseconds (for example, 300 ms). In scenarios such as new energy power generation and high-end equipment drive where continuous operation requirements are extremely stringent, this lead of a few hundred milliseconds is decisive. It is sufficient for the upper-level control system to execute protective actions, such as starting standby units, safely isolating faulty units, or performing an orderly shutdown, thus transforming a potentially destructive fault that may cause production line interruption or system downtime into a planned preventive maintenance. Its application advantage lies in greatly reducing the huge economic losses caused by unplanned downtime.

[0237] Through the original quorum sensing collaborative decision-making model, the present invention solves the technical problem that in a large multi-unit system, the decision of a single unit is unreliable and prone to false alarms due to local noise or individual differences. This effect is achieved by abstracting each H-bridge unit into a biological individual and using the diffusion equation based on the graph Laplacian matrix to simulate the propagation process of abnormal concentration among physically connected unit groups. The input parameter abnormal concentration c is jointly determined by the predicted feature P_pred (representing known risks) and the residual R_res (representing unknown risks), ensuring the comprehensiveness of information input. The core of the algorithm is that the abnormal concentration generated by a real and continuous fault source will form an accumulating and spreading concentration field around it; while the instantaneous abnormality of a single unit caused by random noise will be quickly averaged and diluted due to diffusion. Therefore, only when the concentration in a local area exceeds the adaptive group threshold θ_group obtained from the normal background noise statistics, the system will trigger the final warning. In a cascaded H-bridge system with a large number of units and tight mutual coupling, the advantage of this mechanism is to achieve extremely high decision reliability and effectively avoid false alarms like the boy who cried wolf; at the same time, it can also accurately locate the core unit affected by the fault and its adjacent range by identifying the spatial distribution of the concentration field, providing accurate fault location information for maintenance personnel and significantly improving the operation and maintenance efficiency.

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

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

Claims

1. A fault warning method for a cascaded H-bridge frequency converter, characterized in that, Including: Obtain the on-site current signals of each cascaded H-bridge unit, apply the reverse zero-crossing reconstruction method to perform two-way verification processing on the zero points of the on-site current signals, and generate a reconstructed timestamp matrix; Based on the reconstructed timestamp matrix, perform phase space mapping and dynamic feature extraction, obtain a phase space feature matrix and perform dimensionality reduction processing to obtain a main feature matrix and a residual matrix; Based on the main feature matrix, perform immune memory prediction to generate a predicted feature vector and a memory activation degree; Combine the predicted feature vector, the residual matrix and the memory activation degree to perform quorum sensing collaborative decision-making and generate a warning instruction.

2. The method according to claim 1, wherein Applying the reverse zero-crossing reconstruction method to perform two-way verification processing on the zero points of the on-site current signals includes: For each candidate zero-crossing point, deconstruct the positive value sequence before the zero point and the negative value sequence after the zero point, and apply time reversal and numerical inversion to the negative value sequence to shape an inversion sequence of the ideal positive value trajectory; Apply an interpolation algorithm to the inversion sequence to reconstruct an ideal positive value trajectory that is precisely aligned in time with the positive value sequence; Quantify the comprehensive matching degree between the ideal positive value trajectory and the actually observed positive value sequence, and only when the comprehensive matching degree exceeds the preset verification threshold, confirm the candidate zero-crossing point as a true zero-crossing point.

3. The method according to claim 2, characterized in that, Applying an interpolation algorithm to the inversion sequence to reconstruct an ideal positive value trajectory includes: Set the first-order and second-order derivatives of the inversion sequence as boundary constraint conditions; Under the boundary constraint conditions, construct a cubic spline function for the inversion sequence to perform smooth extension in the time domain; And perform high-density sampling on the cubic spline function to generate an ideal positive value trajectory.

4. The method according to claim 2, characterized in that Quantify the comprehensive matching degree between the ideal positive value trajectory and the actually observed positive value sequence, including: Calculate the normalized correlation coefficient, root mean square relative error and signal-to-noise ratio between the ideal positive value trajectory and the positive value sequence respectively from three dimensions of trajectory shape similarity, numerical deviation magnitude and random noise level, and perform weighted fusion to generate the comprehensive matching degree.

5. The method according to claim 1, characterized in that, Based on the reconstructed timestamp matrix, perform phase space mapping and dynamic feature extraction to obtain a phase space feature matrix, including: Perform cross-unit difference operation on the reconstructed timestamp matrix to generate an original difference sequence representing the multi-unit time sequence relationship; Map the original difference sequence to a three-dimensional phase space composed of the current difference value, the previous moment difference value and the difference change rate to shape a dynamically evolving phase space trajectory; Analyze the topological invariants of the phase space trajectory, and after temperature rise drift self-calibration, fuse them into a phase space feature matrix.

6. The method according to claim 5, wherein Map the original difference sequence to a three-dimensional phase space to shape a dynamically evolving phase space trajectory, 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 differential operation on the original difference sequence, obtain the difference change rate representing its change trend; Perform filtering and smoothing processing on the difference change rate, and use the smoothed difference change rate as the third-axis coordinate to construct a phase space trajectory.

7. The method according to claim 5, wherein Analyze the topological invariants of the phase space trajectory, including: For the phase space trajectory, the trajectory length, average curvature, and radius of gyration are respectively quantified in terms of its overall path length, local bending degree, and spatial distribution compactness, and combined into topological invariants.

8. The method according to claim 1, characterized in that, Combined with the prediction feature vector, residual matrix, and memory activation degree, quorum sensing collaborative decision-making is carried out to generate early warning instructions, including: Each cascaded H-bridge unit is abstracted as a biological individual, and the prediction feature vector and residual matrix are fused to calculate the initial abnormal concentration for each individual; Based on the physical connection relationship between individuals, a diffusion propagation model is constructed to simulate the spatio-temporal evolution process of the initial abnormal concentration between adjacent individuals, forming a diffusion concentration field; When the concentration in a local area of the diffusion concentration field exceeds the threshold, collaborative early warning is triggered, thus generating early warning instructions.

9. The method according to claim 8, wherein Based on the physical connection relationship between individuals, a diffusion propagation model is constructed to simulate the spatio-temporal evolution process of the initial abnormal concentration between adjacent individuals, including: The physical connection relationship between individuals is abstracted as 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; Based on the weighted graph Laplacian matrix, a diffusion partial differential equation is constructed and numerically solved with the initial abnormal concentration as the initial condition, thereby obtaining the diffusion concentration field.

10. The method according to claim 1, characterized in that, Based on the principal eigenmatrix, immune memory prediction is carried out to generate a prediction feature vector and memory activation degree, including: The principal eigenmatrix is matched with a preset fault mode memory library to calculate the similarity between the current system state and each historical fault mode, obtaining an affinity matrix; Based on the affinity matrix, when its value exceeds the activation threshold, the corresponding memory mode is triggered, and the principal eigenmatrix is extrapolated according to the activated memory mode to generate a prediction feature vector; At the same time, the memory activation degree is calculated based on the affinity matrix and the inherent weights of the memory modes.

Images