Wind turbine generator converter health state assessment method and system based on dynamic normalization

Abstract

The embodiment of the invention provides a method and a system for evaluating the health state of a converter of a wind turbine generator based on dynamic regularization. According to the method and the system, the health state evaluation dimension is improved from an original signal layer to a state feature layer. Specifically, firstly, through multi-domain mixed feature extraction, original multi-dimensional data streams (such as three-phase current and direct-current voltage) containing a coupling relation are intelligently condensed into a feature vector sequence capable of comprehensively representing the operation state of the system; then, a dynamic warping algorithm is applied to the high-dimensional feature space, and the health deviation degree is quantified by measuring the warping distance between the real-time state feature sequence and the health benchmark. According to the normal form which is firstly extracted and then structured, the internal coupling relation among the physical quantities is used as core information to be fused into the evaluation process, so that the problem that multi-dimensional data cannot be effectively utilized in a traditional method is fundamentally solved, and the capturing capacity of early weak faults is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of converter condition assessment technology, specifically to a method and system for assessing the health status of wind turbine converters based on dynamic regularization. Background Technology

[0002] Wind turbines are key equipment for converting wind energy into electricity, and the converter, as its core power electronic component, directly affects the grid connection performance, operational reliability, and overall economic benefits of the wind turbine. However, wind turbine converters operate under complex and variable harsh conditions such as wind speed fluctuations and grid disturbances. Key components such as internal power semiconductor devices and DC bus capacitors undergo gradual performance degradation due to cyclical electrothermal stress. If not detected in time, this can develop into sudden failures, leading to unplanned outages and resulting in significant power generation losses and maintenance costs. Therefore, developing technologies that can accurately assess the health status of wind turbine converters, and shifting from post-failure maintenance to predictive maintenance, is of great engineering significance.

[0003] Data-driven health status assessment methods, which analyze large amounts of time-series data (such as three-phase current and DC bus voltage) generated during converter operation to identify early signs of degradation, are currently a hot research topic. Among these, dynamic time warping algorithms are considered a potential technology for condition assessment under varying operating conditions because they can effectively measure the similarity between two time series of different lengths or with time axis scaling or offset. However, when applying dynamic time warping technology to the complex, multi-dimensional system of wind turbine converters, existing methods reveal significant limitations. The core technical problem is that standard dynamic time warping algorithms struggle to effectively handle the multi-dimensional time-series data of converters, neglecting the coupling relationships between various physical quantities. Specifically, changes in the health status of a converter are often the result of the coordinated evolution of multiple physical quantities. For example, an early half-open-circuit fault in an IGBT power module may show slight changes in the amplitude of the three-phase current, but it will significantly disrupt its phase relationship and symmetry, and simultaneously induce ripples of a specific frequency on the DC bus voltage. Existing methods typically employ two approaches: one is to perform DTW calculations on each dimension of the signal (such as three-phase current and DC voltage) separately and independently, and then fuse the results later. This approach completely severs the intrinsic connection between physical quantities and fails to capture the cross-dimensional fault mode characteristics, resulting in the evaluation model being insensitive to early faults. The other approach is to first reduce the multidimensional data to one dimension using methods such as principal component analysis before performing DTW calculations. However, this approach faces the risk of information loss because weak early fault characteristics may be hidden in non-principal components with small variances and are filtered out as noise during the dimensionality reduction process.

[0004] Therefore, an optimized health status assessment scheme for wind turbine converters is desired. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the prior art, and provides a method and system for assessing the health status of wind turbine converters based on dynamic regularization.

[0006] In a first aspect, embodiments of the present invention provide a method for assessing the health status of a wind turbine converter based on dynamic regulation, comprising: Acquire raw three-phase current data and raw DC bus voltage data; A multi-domain hybrid feature extraction based on a sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the state feature vector sequence of the wind turbine converter. The health index is calculated based on the state feature vector sequence of the wind turbine converter and the health benchmark feature sequence. Health status is assessed based on health indices to obtain assessment results.

[0007] Secondly, embodiments of the present invention provide a wind turbine converter health status assessment system based on dynamic regulation, comprising: The data acquisition module is used to acquire raw three-phase current data and raw DC bus voltage data; The multi-domain hybrid feature extraction module is used to perform multi-domain hybrid feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain the state feature vector sequence of the wind turbine converter. The health index calculation module is used to calculate the health index based on the state feature vector sequence of the wind turbine converter and the health benchmark feature sequence. The health assessment module is used to assess health status based on health indices to obtain assessment results.

[0008] Compared with existing technologies, this invention provides a method and system for assessing the health status of wind turbine converters based on dynamic warping, which elevates the dimension of health status assessment from the original signal layer to the state feature layer. It abandons the traditional approach of directly warping and matching multi-dimensional time-series data, which is insensitive to early faults due to the fragmentation or loss of crucial coupling information. This solution constructs an intermediate layer: firstly, through multi-domain hybrid feature extraction, the original multi-dimensional data streams containing coupling relationships (such as three-phase current and DC voltage) are intelligently condensed into a feature vector sequence that comprehensively characterizes the system's operating state. Subsequently, a dynamic warping algorithm is applied to this high-dimensional feature space, quantifying the health deviation by measuring the warping distance between the real-time state feature sequence and the health benchmark. This paradigm of first extracting and then warping incorporates the inherent coupling relationships between various physical quantities as core information into the assessment process, fundamentally solving the problem that traditional methods cannot effectively utilize multi-dimensional data and significantly improving the ability to detect early, subtle faults. Attached Figure Description

[0009] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 A flowchart of a wind turbine converter health status assessment method based on dynamic regularization according to an embodiment of the present invention; Figure 2 This is a data flow diagram illustrating the dynamic regulation-based wind turbine converter health status assessment method according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the process of extracting multi-domain hybrid features from the original three-phase current data and the original DC bus voltage data using a sliding window based method for assessing the health status of wind turbine converters according to an embodiment of the present invention, in order to obtain a sequence of wind turbine converter state feature vectors. Figure 4 This is a flowchart illustrating the calculation of a health index based on a wind turbine converter state feature vector sequence and a health benchmark feature sequence, according to an embodiment of the present invention. Figure 5 This is a block diagram of a wind turbine converter health status assessment system based on dynamic regularization according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0013] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0014] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0015] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0016] To address the technical problem of existing dynamic warping methods being ineffective in handling multi-dimensional coupled data from wind turbine converters, resulting in insensitivity to early faults, this solution proposes a dynamic warping-based health status assessment method for wind turbine converters. This method no longer directly compares the original signals; instead, it elevates the assessment foundation from the signal layer to the state feature layer through a refined feature engineering step. Specifically, the solution first applies a sliding window to the real-time acquired three-phase current and DC bus voltage data, constructing a multi-domain hybrid feature vector within each window. This vector not only includes time-domain statistical features and frequency-domain harmonic features reflecting signal amplitude and waveform distortion, but more importantly, it extracts dq-axis current features through Clarke and Park transforms, directly characterizing the active and reactive power decoupling relationship. This quantifies the physical coupling relationship between multiple variables and integrates it into the vector. Thus, the original data stream is transformed into a more information-dense and physically meaningful sequence of state feature vectors. Subsequently, a dynamic normalization algorithm is applied to this new feature sequence. By calculating the normalized distance between it and the pre-stored health benchmark feature sequence, the degree of deviation from the current operating state is quantified holistically. Finally, this distance is normalized into an intuitive health index to complete the assessment. By first solidifying the coupling relationship into features and then measuring similarity, this scheme fundamentally solves the information fragmentation problem of traditional methods, achieving sensitive and accurate capture of the converter's health status.

[0017] Figure 1 This is a flowchart of a wind turbine converter health status assessment method based on dynamic regularization according to an embodiment of the present invention. Figure 2 This is a data flow diagram illustrating the dynamic regulation-based wind turbine converter health status assessment method according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the wind turbine converter health status assessment method based on dynamic regularization according to an embodiment of the present invention includes the following steps: S100, acquiring raw three-phase current data and raw DC bus voltage data; S200, performing multi-domain hybrid feature extraction based on sliding window on the raw three-phase current data and raw DC bus voltage data to obtain a wind turbine converter state feature vector sequence; S300, calculating a health index based on the wind turbine converter state feature vector sequence and the health benchmark feature sequence; S400, performing a health status assessment based on the health index to obtain an assessment result.

[0018] Specifically, in step S100, raw three-phase current data and raw DC bus voltage data are acquired. It should be understood that since the health status of a wind turbine converter is an external manifestation of its internal physical processes, any performance degradation or early fault symptoms will inevitably be reflected in the dynamic behavior of core electrical quantities, particularly the symmetry of the three-phase current, harmonic content, and the stability and ripple characteristics of the DC bus voltage. Therefore, by acquiring raw three-phase current data and raw DC bus voltage data, raw electrical signals that characterize the converter's power conversion process and the status of key components can be captured comprehensively and in real time. This provides the necessary high-fidelity raw data input for subsequent multi-domain hybrid feature extraction, thus laying a solid data foundation for the accuracy and sensitivity of the entire health status assessment model.

[0019] More specifically, in a concrete example of the invention, the data acquisition process is implemented through a data acquisition module deployed within the wind turbine converter control system. First, for the measurement of the three-phase current, a high-precision Hall effect current sensor is configured for each phase on the AC output bus of the converter. This sensor linearly converts the large current flowing through the bus into a low-voltage analog electrical signal. Second, for the measurement of the DC bus voltage, a high-impedance voltage divider network is connected in parallel between the positive and negative terminals of the DC bus to proportionally reduce the high voltage of several hundred volts to a suitable low voltage range for acquisition, and the signal is conditioned by an isolation amplifier. Then, the analog signals output by the sensors are sent to a multi-channel synchronous data acquisition card. The built-in analog-to-digital converter of this card synchronously samples and quantizes all channels at a preset high sampling frequency, converting the continuous analog signal into a discrete digital sequence. Finally, these digital raw data streams with precise timestamps are transmitted to the processing unit as input for subsequent health status assessment.

[0020] Specifically, in step S200, multi-domain hybrid feature extraction based on a sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state feature vector sequence. It should be understood that because the original time-series data has high dimensionality, contains a large amount of redundant information, and the degradation characteristics directly reflecting the health state are implicit, directly comparing the original data for similarity would not only incur a huge computational burden but would also be easily overwhelmed by operating condition fluctuations and measurement noise, resulting in insensitive and poorly robust evaluation results. Therefore, in the technical solution of this invention, multi-domain hybrid feature extraction based on a sliding window is further performed on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state feature vector sequence. This maps the high-dimensional, redundant original signal space to a lower-dimensional, more information-dense feature space, in which the comprehensive operating state of the converter in the time domain, frequency domain, and multivariate coupling relationships can be explicitly and quantitatively characterized. In this way, a sequence of state feature vectors that is sensitive to performance degradation but insensitive to fluctuations in normal operating conditions can be generated, providing high-quality, high-information-content input for subsequent accurate dynamic regularization matching and health status assessment.

[0021] Figure 3 This is a flowchart illustrating the process of extracting multi-domain hybrid features from raw three-phase current data and raw DC bus voltage data using a sliding window method to obtain a sequence of state feature vectors for the wind turbine converter, according to an embodiment of the present invention. Figure 3 As shown, step S200 includes: S210, performing time-domain statistical feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state time-domain feature sequence; S220, performing frequency-domain harmonic feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state frequency-domain feature sequence; S230, performing coupling relationship feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state coupling relationship feature sequence; S240, performing multi-domain feature mixing on the wind turbine converter state time-domain feature sequence, the wind turbine converter state frequency-domain feature sequence, and the wind turbine converter state coupling relationship feature sequence to obtain the wind turbine converter state feature vector sequence.

[0022] Accordingly, in step S210, time-domain statistical feature extraction based on a sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the time-domain feature sequence of the wind turbine converter state. It should be understood that performance degradation or early failures of key components in the converter directly alter the macroscopic statistical characteristics of its electrical signals in the time domain. For example, three-phase current imbalance disrupts amplitude symmetry, and aging of the DC bus capacitor leads to increased voltage ripple. These changes are subtle in the original data point sequence and difficult to quantify directly. Therefore, in the technical solution of this invention, time-domain statistical feature extraction based on a sliding window is further performed on the original three-phase current data and the original DC bus voltage data to obtain the time-domain feature sequence of the wind turbine converter state. This extracts the waveform information within a short data segment into a set of quantitative indicators that can stably characterize the signal energy, dispersion, and distribution pattern. This generates a feature sequence highly sensitive to macroscopic changes in the signal, effectively highlighting the weak but continuous statistical characteristic shifts caused by physical degradation from background noise.

[0023] Specifically, in one example of the present invention, the time-domain statistical feature extraction is performed within the processing unit for each data segment of a sliding window. The processing unit first calculates the root mean square (RMS) value of the three-phase current data within the window to accurately quantify the energy magnitude of each phase current. Simultaneously, the processing unit calculates the standard deviation of the DC bus voltage data within the window to quantify the voltage ripple fluctuation amplitude and calculates its kurtosis to capture spikes or impulsive disturbances that may be caused by switching anomalies. Subsequently, the processing unit combines these calculated statistics—the three current RMS values, the voltage standard deviation, and the voltage kurtosis—in a preset order into a time-domain feature vector. As the sliding window moves along the time axis with a fixed step size, the above calculation and combination process is repeated continuously, and the resulting series of time-domain feature vectors constitutes the time-domain feature sequence of the wind turbine converter state.

[0024] Accordingly, in step S220, frequency domain harmonic feature extraction based on a sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the frequency domain feature sequence of the wind turbine converter state. It should be understood that due to problems such as the switching non-ideality of the power semiconductor devices inside the converter, aging, and performance degradation of the DC bus capacitor, nonlinear harmonic distortion is directly introduced into the current and voltage waveforms. The specific frequencies and amplitudes of these harmonic components are sensitive indicators characterizing specific fault modes, but they are difficult to directly observe and quantify in the time domain. Therefore, in the technical solution of this invention, frequency domain harmonic feature extraction based on a sliding window is further performed on the original three-phase current data and the original DC bus voltage data to obtain the frequency domain feature sequence of the wind turbine converter state. This transforms the time-domain signal to the frequency domain, thereby enabling accurate separation and quantification of the energy of specific harmonic components closely related to the health state. In this way, a feature sequence directly reflecting the nonlinearity and imbalance of the system can be generated, greatly enhancing the detection capability of subtle spectral changes caused by early faults.

[0025] Specifically, in one example of the present invention, the frequency domain harmonic feature extraction process is performed by the processing unit for each data segment within a sliding window. The processing unit first applies a Fast Fourier Transform (FFT) algorithm to the time-series data of the three-phase current and DC bus voltage within the window, transforming them from the time domain to the frequency domain to obtain the amplitude spectrum of each signal. Subsequently, based on a preset fault characteristic frequency, the processing unit locates and extracts the amplitudes of key harmonic components from the calculated amplitude spectrum. For example, for the three-phase current signal, the processing unit extracts the amplitudes of the fifth and seventh harmonics to monitor distortion caused by dead-time effects or switching device inconsistencies; for the DC bus voltage signal, the processing unit extracts the harmonic amplitudes at twice the fundamental frequency to quantify voltage ripple caused by capacitor degradation or three-phase imbalance. Finally, the processing unit combines all extracted harmonic amplitudes into a frequency domain feature vector and repeats this process as the sliding window moves, thereby constructing a complete frequency domain feature sequence of the wind turbine converter state.

[0026] Accordingly, in step S230, coupling relationship feature extraction based on a sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain a state coupling relationship feature sequence of the wind turbine converter. It should be understood that since the original three-phase AC current is physically highly coupled and varies sinusoidally with time, directly analyzing its time-domain waveform makes it difficult to intuitively assess the converter's performance and stability in active and reactive power control, which are precisely the key indicators reflecting its health status. Therefore, in the technical solution of this invention, coupling relationship feature extraction based on a sliding window is further performed on the original three-phase current data and the original DC bus voltage data to obtain a state coupling relationship feature sequence of the wind turbine converter. This allows for the decoupling and transformation of complex AC quantities into quasi-DC components directly corresponding to active and reactive power, based on vector control theory, and the stability of these components is quantified. This generates a feature sequence that directly characterizes the health of the converter's core control functions, making it highly sensitive to early fault symptoms such as control loop parameter drift and power device asymmetry.

[0027] Specifically, in this embodiment of the invention, coupling relationship feature extraction based on a sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the state coupling relationship feature sequence of the wind turbine converter. This includes: performing Clarke transform and Park transform on the original three-phase current data to obtain the d-axis current and q-axis current; calculating the mean and standard deviation of the d-axis current and q-axis current within the sliding window to obtain the state coupling relationship features of the wind turbine converter. In other words, specifically, this coupling relationship feature extraction process is executed by the processing unit within each sliding window. The processing unit first applies a Clarke transform to the sequence of three-phase current data points within the window, mapping it from a three-phase stationary coordinate system (abc) to a two-phase stationary coordinate system (α-β), obtaining the α-axis and β-axis current components. Subsequently, the processing unit uses the grid voltage synchronization phase angle tracked in real time by a phase-locked loop to perform a Park transform on the α-axis and β-axis current components, rotating them from the two-phase stationary coordinate system to a rotating coordinate system synchronized with the grid voltage, i.e., the dq coordinates, thereby obtaining the time series of the d-axis current and q-axis current. Under steady-state conditions, these two components are quasi-DC signals. Finally, the processing unit calculates the time mean and standard deviation of the d-axis and q-axis current sequences within the window. These four calculated statistical values—the d-axis current mean, q-axis current mean, d-axis current standard deviation, and q-axis current standard deviation—together constitute the state coupling characteristics of the wind turbine converter at that moment. This process is repeated as the sliding window moves, ultimately forming a complete sequence of wind turbine converter state coupling characteristics.

[0028] Accordingly, in step S240, the wind turbine converter state time-domain feature sequence, wind turbine converter state frequency-domain feature sequence, and wind turbine converter state coupling relationship feature sequence are subjected to multi-domain feature mixing to obtain the wind turbine converter state feature vector sequence. It should be understood that since the feature sequences extracted from the time domain, frequency domain, and coupling relationship domain are independent of each other, each can only reflect the converter's operating state from a single dimension, failing to form a unified view that comprehensively describes the system's health status. This limits the ability of subsequent evaluation models to capture the coordinated changes of multi-dimensional features. Therefore, in the technical solution of this invention, the wind turbine converter state time-domain feature sequence, wind turbine converter state frequency-domain feature sequence, and wind turbine converter state coupling relationship feature sequence are further subjected to multi-domain feature mixing to obtain the wind turbine converter state feature vector sequence. This integrates complementary feature information from different analysis dimensions, constructing a unified high-dimensional state descriptor containing all key information for each time window. In this way, a feature vector sequence that can fully characterize the overall health status of the converter at any time can be generated, providing a complete and structured input for the subsequent dynamic warping algorithm. This allows the similarity measurement to be based on global state information, significantly improving the accuracy and comprehensiveness of the assessment.

[0029] Specifically, in a specific example of the present invention, the multi-domain feature mixing process is executed by the processing unit for each synchronous time window. After completing the parallel calculation of the time domain, frequency domain, and coupling relationship features of the data within a specific window, the processing unit obtains a time domain feature vector, a frequency domain feature vector, and a coupling relationship feature vector. Subsequently, the processing unit performs a vector concatenation operation, that is, according to a pre-set fixed order, concatenates the three independent feature vectors end to end to form a single feature vector with higher dimension and more complete information. For example, the time domain feature vector is used as the beginning of the new vector, the frequency domain feature vector as the middle, and the coupling relationship feature vector as the end. This concatenation operation is performed once at each sliding window position, and each high-dimensional feature vector generated corresponds to a state snapshot of a time window. These vectors are arranged in chronological order, ultimately forming the wind turbine converter state feature vector sequence that can dynamically reflect the evolution trajectory of the converter's health state.

[0030] Specifically, in step S300, a health index is calculated based on the wind turbine converter state feature vector sequence and the health benchmark feature sequence. It should be understood that since the wind turbine converter state feature vector sequence itself is a dynamically changing high-dimensional time series, its deviation from the health benchmark cannot be effectively measured by simple point-by-point comparison, especially under varying operating conditions where the time axis undergoes nonlinear scaling. Furthermore, the resulting similarity metric lacks intuitive physical meaning and a unified evaluation standard. Therefore, in the technical solution of this invention, a health index is further calculated based on the wind turbine converter state feature vector sequence and the health benchmark feature sequence. This allows the use of a dynamic time warping algorithm to find the optimal matching path between the two feature sequences, thereby obtaining a quantitative coefficient that comprehensively reflects the overall similarity. This coefficient is then mapped to a standardized health metric within a specific range. In this way, the complex sequence alignment problem can be transformed into the calculation of a single, intuitive, and physically meaningful health index, providing a stable and reliable quantitative basis for subsequent health status assessment and early warning decisions.

[0031] Figure 4 This is a flowchart illustrating the calculation of a health index based on a wind turbine converter state feature vector sequence and a health benchmark feature sequence, according to an embodiment of the present invention. Figure 4 As shown, step S300 includes: S310, calculating the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence; S320, performing dynamic programming to solve the cost matrix to obtain the cumulative cost matrix; S330, extracting the upper right corner element of the cumulative cost matrix as the health quantification coefficient; S340, normalizing the health quantification coefficient to obtain the health index.

[0032] Accordingly, in step S310, the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence is calculated. It should be understood that existing dynamic time warping methods typically use Euclidean distance to measure the difference between real-time feature vectors and health baseline feature vectors when constructing the cost matrix. The inherent flaw of this approach is that it assumes the multi-dimensional feature space to be an isotropic standard Euclidean space, which does not conform to the physical reality of wind turbine converter systems. Specifically, this flaw manifests on two levels: First, it ignores the strong coupling relationships between various physical features. For example, features such as the dq-axis current component, DC bus voltage ripple, and output current harmonics are not independent but change collaboratively according to specific physical laws. The practice of independently calculating the differences in each dimension using Euclidean distance severs this inherent connection and cannot effectively identify early fault modes composed of collaborative offsets of multiple features. Second, it assigns equal weights to all features, ignoring the significant differences in the inherent fluctuation range (i.e., variance) of different features in a healthy state. Some key features (such as d-axis current) should be highly stable under healthy operating conditions, and even slight deviations may indicate serious problems; while other features (such as RMS current) will fluctuate normally with operating conditions. Euclidean distance cannot distinguish between these two types of deviations, making the evaluation results easily interfered with by normal fluctuations in high-variance features, thus masking weak abnormal signals on low-variance features with diagnostic value. Therefore, to address the above technical shortcomings, a feature space cost calculation method based on probability measures is proposed. Its core lies in replacing Euclidean distance with Mahalanobis distance, which contains the inherent statistical structure of the data. Specifically, in the technical solution of this invention, the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence is further calculated. This abandons the traditional Euclidean distance metric and instead constructs a metric tensor that reflects the correlation and variance between features by learning the covariance structure of the health state feature space. Based on this metric tensor, the local cost between feature vectors is defined. In this way, a more statistically accurate cost matrix can be generated, in which each cost value reflects the true degree of deviation of the current state from the healthy statistical distribution, rather than a simple geometric distance. This cost matrix can effectively amplify the multi-feature collaborative offset that indicates early failures, while suppressing the interference of high variance feature fluctuations under normal operating conditions, providing a high signal-to-noise ratio and high sensitivity foundation for subsequent dynamic programming solutions.

[0033] Specifically, in this embodiment of the invention, calculating the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence includes: learning the health state feature space covariance structure of the health baseline feature sequence to obtain a global covariance matrix; regularizing the global covariance matrix to obtain a metric tensor; and calculating the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence based on the metric tensor.

[0034] More specifically, the health benchmark feature sequence is subjected to health state feature space covariance structure learning to obtain a global covariance matrix. It should be understood that, since the multidimensional features of a converter during healthy operation are not uniformly and randomly distributed in space, but rather follow inherent physical laws, forming a high-dimensional probability distribution cloud with a specific shape and orientation and anisotropy, without establishing a mathematical model of this statistical structure, any distance metric will be unable to accurately assess state deviation because it ignores the strong coupling relationships between features and the significant differences in their inherent fluctuation ranges. Therefore, in the technical solution of this invention, the health benchmark feature sequence is subjected to health state feature space covariance structure learning to obtain a global covariance matrix, thereby learning from a large number of health benchmark samples and establishing a mathematical model that can accurately describe the inherent statistical structure of the feature space under healthy conditions. That is, specifically, the collected health benchmark feature sequence... Calculate its sample covariance matrix It can be expressed by the following formula:

[0035] in, It is a sequence of health benchmark features within a health benchmark feature sequence. The global covariance matrix is ​​a structured and mathematical expression of the converter's healthy operating mode. This allows us to obtain a global covariance matrix that serves as a structured and mathematical expression of the converter's healthy operating mode. This matrix quantitatively captures the fluctuation amplitude of each feature (represented by the variance of the diagonal elements) and the linear correlation between features (represented by the covariance of the off-diagonal elements), laying a core foundation for subsequently constructing a cost metric sensitive to physical degradation and based on statistical probability.

[0036] More specifically, the global covariance matrix is ​​regularized to obtain the metric tensor. It should be understood that an effective distance metric must be able to adaptively weight features based on their stability and correlation. The inverse of the covariance matrix (i.e., the precision matrix) naturally possesses this capability. However, directly inverting the covariance matrix learned from actual engineering data can lead to matrix singularity or ill-conditioned behavior due to feature collinearity, resulting in numerical instability and preventing the acquisition of a reliable metric. Therefore, in the technical solution of this invention, the global covariance matrix is ​​further regularized to obtain the metric tensor. This allows the construction of a mathematically robust and physically meaningful intelligent metric based on the learned covariance structure. This metric is no longer uniform but is precisely curved according to the inherent distribution characteristics of health data. Specifically, in this embodiment of the invention, regularizing the global covariance matrix to obtain the metric tensor includes: regularizing the global covariance matrix using the following formula:

[0037] in, The global covariance matrix, The regularization coefficient is . It is the identity matrix. For measuring tensors.

[0038] This allows the generation of a metric tensor that assigns extremely high weights when the direction of feature change aligns with the low-variance principal axis direction in a healthy state, and lower weights in the high-variance direction, thus fundamentally achieving sensitive amplification of weak anomalous signals indicating early faults. Here, a regularization term is introduced. This is to ensure good numerical stability of the matrix inversion process when dealing with eigencollinearity issues that may exist in real-world engineering data. This allows for the generation of an intelligent metric that is no longer uniform but curved according to the distribution characteristics of the health data. This enables the generation of a metric tensor that assigns extremely high weights to directions of characteristic variation that align with the low-variance principal axis of the healthy state, and lower weights to directions with high variance, thereby fundamentally achieving sensitive amplification of weak anomaly signals indicating early faults.

[0039] More specifically, based on the metric tensor, a cost matrix is ​​calculated between the wind turbine converter state feature vector sequence and the health baseline feature sequence. It should be understood that while the preceding steps have constructed a metric tensor capable of representing the inherent statistical structure of the health state feature space, if it is not applied to the core of the dynamic time warping algorithm to replace the flawed Euclidean distance, the valuable information about feature stability and correlation contained in this metric tensor will be unable to play a role in similarity measurement, thus failing to fundamentally transform the traditional DTW algorithm. Therefore, in the technical solution of this invention, a cost matrix is ​​further calculated between the wind turbine converter state feature vector sequence and the health baseline feature sequence based on the metric tensor, thereby injecting a new, probability-based distance metric paradigm into the core of the DTW algorithm, enabling it to perceive and utilize the inherent statistical structure of the feature space. Specifically, in this embodiment of the invention, for each pair of wind turbine converter state feature vectors and health baseline features between the wind turbine converter state feature vector sequence and the health baseline feature sequence, the square root of the Mahalanobis distance between them is calculated through quadratic form operations and used as an element of the cost matrix. That is, based on the metric tensor, the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence is calculated using the following formula:

[0040] in, To measure tensors, This refers to the state feature vectors of each wind turbine converter in the sequence of wind turbine converter state feature vectors. For each health benchmark feature in the health benchmark feature sequence, These are the elements in the cost matrix. This allows the generation of a completely new cost matrix where each element no longer merely reflects the geometric distance, but rather more profoundly measures the probability of real-time features deviating from a healthy baseline. This makes the entire DTW algorithm highly sensitive to subtle, multi-feature collaborative change patterns that foreshadow early failures.

[0041] In particular, this optimized technical solution significantly improves the accuracy and physical interpretability of the dynamic time warping algorithm in similarity measurement. By introducing a cost function based on Mahalanobis distance, the DTW distance can effectively suppress noise interference caused by fluctuations in normal operating conditions, while amplifying early fault characteristics that are inconsistent with the healthy mode, thereby obtaining a more sensitive and robust quantitative indicator for changes in the true health state of the converter.

[0042] Accordingly, in step S320, dynamic programming is used to solve the cost matrix to obtain the cumulative cost matrix. It should be understood that the cost matrix obtained in the previous step only provides a local, static similarity measure for all point pairs between two feature vector sequences. It does not address the nonlinear scaling of the time axis caused by varying operating conditions or small disturbances, nor can it directly provide a single global index to measure the overall similarity between the two sequences. Therefore, in the technical solution of this invention, dynamic programming is further used to solve the cost matrix to obtain the cumulative cost matrix. This allows for a systematic search and determination of an optimal regular path from the starting point to the ending point in the two-dimensional space defined by the cost matrix, where the cumulative cost of this path is minimized among all possible paths. In this way, the complex, high-dimensional sequence alignment problem can be transformed into an optimization problem with optimal substructure and overlapping subproblems, which can be efficiently solved through dynamic programming, ultimately generating a cumulative cost matrix whose each element contains the minimum global cost information to reach that point.

[0043] Specifically, in a specific example of the present invention, the dynamic programming solution process is executed by the processing unit after obtaining the cost matrix. The processing unit first creates a cumulative cost matrix with the same size as the cost matrix and initializes its first element to the value of the first element of the cost matrix. Then, the processing unit iterates through all the remaining cells of the cumulative cost matrix from left to right and from top to bottom using a nested loop. For each cell to be calculated, the processing unit examines the cumulative cost values ​​of its three adjacent cells (left, top, and top-left) that have already been calculated, and selects the minimum value. Then, the processing unit adds this minimum value to the local cost value corresponding to the current position in the original cost matrix, and uses this sum as the cumulative cost value of the current cell. This process strictly follows the recursive formula of dynamic programming until the last cell of the cumulative cost matrix is ​​filled, thus obtaining the complete cumulative cost matrix.

[0044] Accordingly, in step S330, the upper right element of the cumulative cost matrix is ​​extracted as a health quantification coefficient. It should be understood that since the cumulative cost matrix itself is a complete data field containing the optimal path cost to all possible matching points, it is not a final scalar result directly used for evaluation. The core information for measuring the global similarity between two complete sequences is uniquely encoded at the endpoint of this matrix. Therefore, in the technical solution of this invention, the upper right element of the cumulative cost matrix is ​​further extracted as a health quantification coefficient to complete the final step of the dynamic time warping algorithm, that is, to separate the globally optimal matching cost of the entire sequence alignment process from the intermediate calculation results, as a single, comprehensive quantification index. In this way, the overall difference between two high-dimensional time series after nonlinear alignment can be condensed into a single, unnormalized scalar value. This value directly quantifies the overall deviation of the current operating state from the health benchmark, providing direct numerical input for subsequent normalization processing to generate a standardized health index.

[0045] Specifically, in a specific example of the present invention, the extraction process is executed immediately by the processing unit after all calculations of the cumulative cost matrix are completed. The processing unit first determines the dimensions of the cumulative cost matrix, with the number of rows and columns corresponding to the lengths of the health baseline feature sequence and the wind turbine converter state feature vector sequence, respectively. Subsequently, the processing unit directly locates and reads the element located in the upper right corner of the matrix, i.e., the element value of the last column in the first row, or, according to a specific implementation convention, the element value at the intersection of the last row and the last column. This read value, physically representing the total cost of the optimal normalized path between the two complete sequences, is directly designated as the health quantification coefficient and passed to the subsequent normalization processing module.

[0046] Accordingly, in step S340, the health quantification coefficient is normalized to obtain the health index. It should be understood that since the health quantification coefficient obtained in the previous step is a raw, unscaled distance value, its absolute size not only lacks intuitive physical meaning, but its value range also varies with the length of the feature sequence and the dimension of the feature vector. This makes setting a unified, fixed threshold to judge the health status extremely difficult and unreliable. Therefore, in the technical solution of this invention, the health quantification coefficient is further normalized to obtain the health index, thereby mapping this dimensionless distance value, which is related to specific operating conditions and data length, to a fixed, standardized interval with clear physical meaning, such as between 0 and 1. This generates an intuitive, stable, and cross-operating condition-comparable final evaluation index, which directly corresponds to the health level of the equipment, providing a clear and consistent quantitative basis for subsequent automated status assessment, trend prediction, and maintenance decisions.

[0047] Specifically, in this embodiment of the invention, normalizing the health quantification coefficient to obtain the health index includes: normalizing the health quantification coefficient using the following formula:

[0048] in, A health quantification coefficient, The warning distance threshold, The normalization process is performed by the processing unit, specifically, to determine the fault distance threshold. The processing unit first calls two pre-set and stored threshold parameters: a warning distance threshold and a fault distance threshold. These two thresholds are determined through statistical analysis of health quantification coefficients from a large amount of historical health data and known fault data. Subsequently, when the processing unit receives a new health quantification coefficient, it performs a piecewise linear mapping. If the health quantification coefficient is less than or equal to the warning distance threshold, the processing unit directly assigns a health index of 1, indicating that the device is in a fully healthy state. If the coefficient is greater than or equal to the fault distance threshold, the health index is assigned a value of 0, indicating that the device has experienced a definite fault. When the value of the coefficient is between the warning distance threshold and the fault distance threshold, the processing unit calculates a value between 0 and 1 using a linear interpolation formula as the health index. This calculation result is the normalized final health index and is output for subsequent evaluation.

[0049] Specifically, in step S400, a health status assessment is performed based on the health index to obtain an assessment result. It should be understood that since the health index generated in the previous step is a continuous quantitative value, it does not directly provide a discrete qualitative conclusion that is easy for maintenance personnel to understand and implement. Furthermore, the health index at a single point in time may fluctuate due to instantaneous disturbances. Directly judging based on a single-point value would reduce the stability and reliability of the assessment system. Therefore, in the technical solution of this invention, a health status assessment is further performed based on the health index to obtain an assessment result. This establishes a mapping rule from continuous quantitative indicators to discrete health status levels, and combines this with trend analysis of the health index time series to make a final, robust health status judgment. In this way, the dynamic, continuous numerical assessment can be transformed into a clear, stable, and instructive assessment result, effectively avoiding misjudgments caused by instantaneous data fluctuations, and providing direct and reliable decision input for the predictive maintenance strategy of wind turbine units.

[0050] More specifically, in a concrete example of the present invention, the health status assessment process is triggered by the processing unit calculating a new health index in each assessment cycle. First, the processing unit initially classifies the current health index according to preset health status partitioning thresholds. For example, the system presets two thresholds: a health threshold (e.g., 0.9) and a warning threshold (e.g., 0.7). If the health index is greater than the health threshold, it is initially determined to be healthy; if it is between the warning threshold and the health threshold, it is considered sub-healthy; if it is lower than the warning threshold, it is considered faulty. Next, to enhance the robustness of the assessment, the processing unit stores the health index of the current and several past assessment cycles in a fixed-length first-in-first-out queue and performs a moving average calculation on the data in the queue to smooth out short-term noise. Finally, the processing unit applies the aforementioned partitioning thresholds to the smoothed health index mean and combines it with a continuous judgment logic. For example, only when the smoothed health index is lower than the health threshold for three consecutive assessment cycles will the system officially update the assessment result to sub-healthy and generate a corresponding warning signal. The final assessment result is recorded and displayed on the monitoring interface.

[0051] In summary, the dynamic warping-based health status assessment method for wind turbine converters according to embodiments of the present invention is explained, which elevates the dimension of health status assessment from the original signal layer to the state feature layer. It abandons the traditional approach of directly warping and matching multidimensional time-series data, which is insensitive to early faults due to the fragmentation or loss of key coupling information. This scheme constructs an intermediate layer: firstly, through multi-domain hybrid feature extraction, the original multidimensional data streams containing coupling relationships (such as three-phase current and DC voltage) are intelligently condensed into a feature vector sequence that comprehensively characterizes the system's operating state. Subsequently, the dynamic warping algorithm is applied to this high-dimensional feature space, quantifying the health deviation by measuring the warping distance between the real-time state feature sequence and the health benchmark. This paradigm of first extracting and then warping incorporates the inherent coupling relationships between various physical quantities as core information into the assessment process, thereby fundamentally solving the problem that traditional methods cannot effectively utilize multidimensional data and significantly improving the ability to detect early, subtle faults.

[0052] Furthermore, a health status assessment system for wind turbine converters based on dynamic regularization is also provided.

[0053] Figure 5 This is a block diagram of a wind turbine converter health status assessment system based on dynamic regulation according to an embodiment of the present invention. Figure 5 As shown, the wind turbine converter health status assessment system 100 based on dynamic regularization according to an embodiment of the present invention includes: a data acquisition module 110 for acquiring raw three-phase current data and raw DC bus voltage data; a multi-domain hybrid feature extraction module 120 for performing multi-domain hybrid feature extraction based on a sliding window on the raw three-phase current data and raw DC bus voltage data to obtain a wind turbine converter state feature vector sequence; a health index calculation module 130 for calculating a health index based on the wind turbine converter state feature vector sequence and a health benchmark feature sequence; and a health assessment module 140 for performing a health status assessment based on the health index to obtain an assessment result.

[0054] Furthermore, the multi-domain hybrid feature extraction module 120 is specifically used for: performing time-domain statistical feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state time-domain feature sequence; performing frequency-domain harmonic feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state frequency-domain feature sequence; performing coupling relationship feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain a wind turbine converter state coupling relationship feature sequence; and performing multi-domain feature mixing on the wind turbine converter state time-domain feature sequence, the wind turbine converter state frequency-domain feature sequence, and the wind turbine converter state coupling relationship feature sequence to obtain the wind turbine converter state feature vector sequence.

[0055] As described above, the wind turbine converter health status assessment system 100 based on dynamic regularization according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with a wind turbine converter health status assessment algorithm based on dynamic regularization. In one possible implementation, the wind turbine converter health status assessment system 100 based on dynamic regularization according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the wind turbine converter health status assessment system 100 based on dynamic regularization can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the wind turbine converter health status assessment system 100 based on dynamic regularization can also be one of many hardware modules of the wireless terminal.

[0056] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for assessing the health status of wind turbine converters based on dynamic regularization, characterized in that, include: Acquire raw three-phase current data and raw DC bus voltage data; A multi-domain hybrid feature extraction based on a sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the state feature vector sequence of the wind turbine converter. The health index is calculated based on the state feature vector sequence of the wind turbine converter and the health benchmark feature sequence. Health status is assessed based on health indices to obtain assessment results.

2. The method for assessing the health status of wind turbine converters based on dynamic regularization according to claim 1, characterized in that, Multi-domain hybrid feature extraction based on a sliding window is performed on the raw three-phase current data and raw DC bus voltage data to obtain the state feature vector sequence of the wind turbine converter, including: Time-domain statistical feature extraction based on sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the time-domain feature sequence of the wind turbine converter state; Frequency domain harmonic feature extraction based on sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the frequency domain feature sequence of the wind turbine converter state; The coupling relationship feature is extracted based on the sliding window from the original three-phase current data and the original DC bus voltage data to obtain the state coupling relationship feature sequence of the wind turbine converter; The wind turbine converter state time-domain feature sequence, wind turbine converter state frequency-domain feature sequence, and wind turbine converter state coupling relationship feature sequence are mixed with multi-domain features to obtain the wind turbine converter state feature vector sequence.

3. The method for assessing the health status of wind turbine converters based on dynamic regularization according to claim 2, characterized in that, The original three-phase current data and original DC bus voltage data are subjected to coupling relationship feature extraction based on a sliding window to obtain the state coupling relationship feature sequence of the wind turbine converter, including: The original three-phase current data were subjected to Clarke and Park transformations to obtain the d-axis and q-axis currents. The mean and standard deviation of the d-axis current and q-axis current within the sliding window are calculated to obtain the state coupling characteristics of the wind turbine converter.

4. The method for assessing the health status of wind turbine converters based on dynamic regularization according to claim 1, characterized in that, Based on the state characteristic vector sequence of the wind turbine converter and the health benchmark characteristic sequence, a health index is calculated, including: Calculate the cost matrix between the state feature vector sequence of the wind turbine converter and the health baseline feature sequence; The cumulative cost matrix is ​​obtained by solving the cost matrix using dynamic programming. Extract the top right element of the cumulative cost matrix as the health quantification coefficient; The health quantification coefficients are normalized to obtain the health index.

5. The method for assessing the health status of wind turbine converters based on dynamic regularization according to claim 4, characterized in that, Calculate the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence, including: The global covariance matrix is ​​obtained by learning the covariance structure of the health status feature space from the health baseline feature sequence. Regularize the global covariance matrix to obtain the metric tensor; Based on the metric tensor, the cost matrix between the state feature vector sequence of the wind turbine converter and the health baseline feature sequence is calculated.

6. The method for assessing the health status of wind turbine converters based on dynamic regularization according to claim 5, characterized in that, Regularizing the global covariance matrix to obtain the metric tensor includes: regularizing the global covariance matrix using the following formula: in, The global covariance matrix, The regularization coefficient is . It is the identity matrix. For measuring tensors.

7. The method for assessing the health status of wind turbine converters based on dynamic regularization according to claim 5, characterized in that, Based on the metric tensor, the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence is calculated, including: based on the metric tensor, the cost matrix between the wind turbine converter state feature vector sequence and the health baseline feature sequence is calculated using the following formula: in, To measure tensors, This refers to the state feature vectors of each wind turbine converter in the sequence of wind turbine converter state feature vectors. For each health benchmark feature in the health benchmark feature sequence, These are the elements in the cost matrix.

8. The method for assessing the health status of wind turbine converters based on dynamic regularization according to claim 4, characterized in that, The health index is obtained by normalizing the health quantification coefficients, including: normalizing the health quantification coefficients using the following formula: in, A health quantification coefficient, The warning distance threshold, This is the fault distance threshold.

9. A health status assessment system for wind turbine converters based on dynamic regularization, characterized in that, include: The data acquisition module is used to acquire raw three-phase current data and raw DC bus voltage data; The multi-domain hybrid feature extraction module is used to perform multi-domain hybrid feature extraction based on a sliding window on the original three-phase current data and the original DC bus voltage data to obtain the state feature vector sequence of the wind turbine converter. The health index calculation module is used to calculate the health index based on the state feature vector sequence of the wind turbine converter and the health benchmark feature sequence. The health assessment module is used to assess health status based on health indices to obtain assessment results.

10. The wind turbine converter health status assessment system based on dynamic regularization according to claim 9, characterized in that, The multi-domain hybrid feature extraction module is further used for: Time-domain statistical feature extraction based on sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the time-domain feature sequence of the wind turbine converter state; Frequency domain harmonic feature extraction based on sliding window is performed on the original three-phase current data and the original DC bus voltage data to obtain the frequency domain feature sequence of the wind turbine converter state; The coupling relationship feature is extracted based on the sliding window from the original three-phase current data and the original DC bus voltage data to obtain the state coupling relationship feature sequence of the wind turbine converter; The wind turbine converter state time-domain feature sequence, wind turbine converter state frequency-domain feature sequence, and wind turbine converter state coupling relationship feature sequence are mixed with multi-domain features to obtain the wind turbine converter state feature vector sequence.

Images