Fan blade lightning protection conduction test method and system based on machine learning

Abstract

The invention provides a fan blade lightning protection conduction testing method and system based on machine learning, and relates to the technical field of lightning protection conduction testing, and the method comprises the steps: carrying out the current testing of a fan blade lightning protection system through a multi-channel MOS tube control circuit; performing multi-dimensional feature extraction and local feature analysis on the current response data; carrying out feature enhancement and defect identification based on the local feature sequence, and carrying out multistage clustering on an identification result; and performing global health analysis and defect traceability of lightning protection conduction performance on the fan blade lightning protection system according to a defect classification result, generating a defect attribution point location for monitoring, performing feedback optimization on a multi-channel MOS tube control circuit, and constructing a test optimization closed loop. According to the invention, the technical problem that the lightning protection conduction test efficiency is low due to the fact that the static conduction test is difficult to identify early hidden defects in the prior art is solved, the current test is carried out through the multi-channel MOS tube control circuit, and the lightning protection conduction test efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of lightning protection continuity testing technology, and in particular to a lightning protection continuity testing method and system for wind turbine blades based on machine learning. Background Technology

[0002] As the most vulnerable component of a wind turbine, the reliability of its lightning protection system directly affects the safety of the entire unit. With the increasing size of wind turbines and blades, their internal lightning protection systems are becoming increasingly complex.

[0003] Currently, existing lightning protection continuity testing methods mainly rely on static continuity testing, typically assessing the system's continuity by measuring current or resistance. However, latent defects often do not appear immediately, manifesting as minor anomalies in resistance values ​​or changes in transient response characteristics during routine testing. These are easily overlooked, but significantly increase the risk of localized overheating or breakdown during lightning current discharge. This results in low testing efficiency and an inability to provide early warnings, creating blind spots and delays in the safe operation and maintenance of wind turbine lightning protection.

[0004] In summary, existing technologies suffer from the problem that static continuity testing makes it difficult to identify early latent defects, resulting in low efficiency of lightning protection continuity testing and further affecting the safe operation and maintenance level of wind turbine units. Summary of the Invention

[0005] The purpose of this application is to provide a machine learning-based method and system for testing the lightning protection continuity of wind turbine blades, in order to solve the technical problem in the prior art where static continuity testing is difficult to identify early latent defects, resulting in low efficiency of lightning protection continuity testing and further affecting the safe operation and maintenance level of wind turbine units.

[0006] In view of the above problems, this application provides a method and system for testing the lightning protection continuity of wind turbine blades based on machine learning.

[0007] Firstly, this application provides a machine learning-based method for testing the lightning protection continuity of wind turbine blades. This method is implemented through a machine learning-based system for testing the lightning protection continuity of wind turbine blades. The method includes: performing current tests on multiple test points of the wind turbine blade lightning protection system using a multi-channel MOSFET control circuit to obtain current response data; extracting multi-dimensional features from the current response data, constructing a multi-dimensional test feature vector, performing local feature analysis, and obtaining a local feature sequence; enhancing features based on the local feature sequence, calculating feature contribution for defect identification, performing multi-level clustering based on the identification results, and generating defect classification results; performing global health analysis on the wind turbine blade lightning protection system based on the defect classification results, tracing the source of lightning protection continuity performance defects based on the health score results, generating defect attribution points for monitoring, and feeding back the monitoring results to the multi-channel MOSFET control circuit for feedback optimization, thus constructing a test optimization closed loop.

[0008] Optionally, voltage analysis is performed using a programmable voltage source to determine a test voltage sequence, which includes a test voltage signal; a MOSFET switching matrix is ​​constructed based on a multi-channel MOSFET control circuit, which independently controls each channel of the MOSFET control circuit; the test voltage signal is applied to the multiple test points of the wind turbine blade lightning protection system through the MOSFET switching matrix; transient current testing is performed on the multiple test points using a current sensor to obtain transient current response data; steady-state current testing is performed on the multiple test points using a current sensor to obtain steady-state current response data; the transient current response data and the steady-state current response data are integrated to obtain the current response data.

[0009] Optionally, a current signal is extracted based on the current response data; current time-series analysis is performed on the current signal to construct a current-time curve; time-domain analysis is performed on the current-time curve to obtain time-domain waveform features; the current signal is analyzed based on the time-domain waveform features to determine the current amplitude features; the current signal is transformed to obtain the signal frequency for frequency-domain analysis to determine the current frequency-domain features; entropy is calculated based on the current signal to obtain entropy features for nonlinear distortion analysis to determine nonlinear features; the current amplitude features, the current frequency-domain features, and the nonlinear features are grouped to generate multiple feature groups for matching and identification, thus constructing the multi-dimensional test feature vector.

[0010] Optionally, multiple windows are defined according to the current amplitude characteristics, the current frequency domain characteristics, and the nonlinear characteristics; the multi-dimensional test feature vector is segmented based on the multiple windows to generate multiple test features; local analysis is performed on the multiple test features to obtain the feature distribution parameters of the multiple test features; gradient change analysis is performed based on the feature distribution parameters to obtain feature change trend data; the multiple test features are arranged in chronological order according to the feature change trend data to construct the local feature sequence.

[0011] Optionally, the local feature sequence is traversed to normalize the multiple test features, and the processing results are combined to generate composite features; historical defect logs of the wind turbine blade lightning protection system are retrieved, multiple defect types are extracted from the historical defect logs, and multiple feature discrimination scores are generated based on the composite features according to the multiple defect types; multiple key features are obtained by filtering according to the multiple feature discrimination scores; cross-analysis is performed on the multiple key features to generate interaction parameters, and the multiple key features are associated and identified according to the interaction parameters to generate multiple enhanced features; the contribution of the multiple enhanced features combined with the multiple defect types is calculated to obtain the feature contribution.

[0012] Optionally, the multiple defect types are traversed to perform differential analysis on the multiple enhancement features to generate feature distribution difference parameters; the multiple enhancement features are calculated using a normal distribution to obtain a feature central tendency; the multiple enhancement features are calculated for distribution overlap based on the feature central tendency and the feature distribution difference parameters to obtain a distribution overlap degree; the multiple defect types are weighted based on the distribution overlap degree and the multiple enhancement features to obtain a weighted contribution degree for the multiple defect types; the weighted contribution degrees of the multiple defect types are summed to obtain the feature contribution degree.

[0013] Optionally, defect matching is performed according to the feature contribution degree to generate a defect matching degree. When the defect matching degree is greater than a preset matching degree threshold, a target defect dataset is identified and determined. The target defect dataset is coarsely classified according to the physical characteristics of the defects to generate a first clustering result. Based on the first clustering result, defect impact analysis is performed, and the first clustering result is finely classified according to the defect impact factors to generate a second clustering result. Based on the target defect dataset, the defect distribution location is located, and spatial clustering analysis is performed on the second clustering result according to the defect distribution location to generate a third clustering result. Based on the target defect dataset, defect time extrapolation is performed, and temporal clustering analysis is performed on the third clustering result according to the defect evolution data to generate the defect classification result.

[0014] Optionally, based on the defect classification results, the spatial distribution of the wind turbine blade lightning protection system is evaluated to generate defect spatial distribution characteristics; usage prediction is performed based on the defect spatial distribution characteristics to obtain service life prediction results; global health calculation is performed according to the service life prediction results to generate an initial health score; environmental parameters are introduced to correct the initial health score to generate a health rating result; correlation analysis is performed on the multiple test points based on the health rating result to generate data correlation coefficients; defect impact analysis is performed on the health rating result according to the data correlation coefficients; the results are sorted in descending order according to the defect impact coefficients to extract multiple key defect types; defect attribution points are generated based on the multiple key defect types to trace the source of lightning protection conductivity.

[0015] Optionally, based on the multiple key defect types, the multiple test points are traced according to a time sequence to identify the anomaly initiation point; causal reasoning is performed based on the anomaly initiation point to construct a defect cause classification list for attribution, determining the defect attribution point; monitoring enhancement is performed based on the defect attribution point to construct a key monitoring scheme; the key monitoring scheme is executed to monitor and identify the multiple test points, determining multiple key test monitoring points; the monitoring results of the multiple key test monitoring points are transmitted to the multi-channel MOSFET control circuit in real time to generate control feedback parameters; the multi-channel MOSFET control circuit is dynamically adjusted and optimized based on the control feedback parameters, a self-learning mechanism is constructed for iterative improvement, and a test optimization closed loop is constructed.

[0016] Secondly, this application also provides a machine learning-based wind turbine blade lightning protection continuity testing system for performing the machine learning-based wind turbine blade lightning protection continuity testing method described in the first aspect. The machine learning-based wind turbine blade lightning protection continuity testing system includes: a current testing module for performing current tests on multiple test points of the wind turbine blade lightning protection system using a multi-channel MOS transistor control circuit to obtain current response data; a local feature analysis module for extracting multi-dimensional features from the current response data, constructing a multi-dimensional test feature vector for local feature analysis, and obtaining a local feature sequence; a defect clustering module for feature enhancement based on the local feature sequence, calculating feature contribution for defect identification, performing multi-level clustering based on the identification results, and generating defect classification results; and a defect tracing module for performing global health analysis of the wind turbine blade lightning protection system based on the defect classification results, tracing the defects in lightning protection continuity performance based on the health score results, generating defect attribution points for monitoring, and feeding back the monitoring results to the multi-channel MOS transistor control circuit for feedback optimization, thus constructing a test optimization closed loop.

[0017] One or more technical solutions provided in this application have at least the following beneficial effects: Current tests are performed on multiple test points of the wind turbine blade lightning protection system using a multi-channel MOSFET control circuit to obtain current response data. Multi-dimensional feature extraction is performed on the current response data to construct a multi-dimensional test feature vector for local feature analysis, resulting in a local feature sequence. Feature enhancement is performed based on the local feature sequence, and feature contribution is calculated for defect identification. Multi-level clustering is then performed based on the identification results to generate defect classification results. A global health analysis is conducted on the wind turbine blade lightning protection system based on the defect classification results. Defects in lightning protection conduction performance are traced based on the health score results, generating defect attribution points for monitoring. The monitoring results are fed back to the multi-channel MOSFET control circuit for feedback optimization, constructing a test optimization closed loop. In other words, by using a multi-channel MOSFET control circuit to perform current tests on multiple test points, the collected data is used to identify defects. Based on the defect classification results, a global health analysis of the wind turbine blade lightning protection system is performed to determine the defect attribution points, monitor them, and feed back the monitoring results to the multi-channel MOSFET control circuit for feedback optimization. This achieves a shift from passive maintenance to predictive maintenance, significantly improving testing efficiency and thus enhancing the safe operation and maintenance level of the wind turbine lightning protection system. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the lightning protection continuity testing method for wind turbine blades based on machine learning, as described in this application.

[0019] Figure 2 This is a schematic diagram of the structure of the wind turbine blade lightning protection continuity test system based on machine learning in this application.

[0020] Figure labeling: Current testing module 11, Local feature analysis module 12, Defect clustering module 13, Defect tracing module 14. Detailed Implementation

[0021] This application provides a machine learning-based method and system for testing the lightning protection continuity of wind turbine blades. It addresses the technical problem in existing technologies where static continuity testing struggles to identify early, latent defects, leading to low efficiency and negatively impacting the safe operation and maintenance of wind turbine units. By using a multi-channel MOSFET control circuit to perform current tests at multiple test points, the collected data is used for defect identification. Based on the defect classification results, a global health analysis of the wind turbine blade lightning protection system is performed to determine the defect attribution points, which are then monitored. The monitoring results are fed back to the multi-channel MOSFET control circuit for optimization, achieving a shift from passive inspection to predictive maintenance. This significantly improves testing efficiency and enhances the safe operation and maintenance level of the wind turbine lightning protection system.

[0022] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0023] Example 1, please refer to the appendix. Figure 1 This application provides a machine learning-based method for testing the lightning protection continuity of wind turbine blades. The method is executed by a machine learning-based system and specifically includes the following steps: Current response data was obtained by performing current tests at multiple test points of the wind turbine blade lightning protection system using a multi-channel MOSFET control circuit.

[0024] Furthermore, this application also includes the following steps: performing voltage analysis using a programmable voltage source to determine a test voltage sequence, the test voltage sequence including a test voltage signal; constructing a MOSFET switching matrix based on a multi-channel MOSFET control circuit, the MOSFET switching matrix independently controlling each channel MOSFET control circuit; applying the test voltage signal to the plurality of test points of the wind turbine blade lightning protection system through the MOSFET switching matrix; performing transient current testing on the plurality of test points using a current sensor to obtain transient current response data; performing steady-state current testing on the plurality of test points using a current sensor to obtain steady-state current response data; and integrating the transient current response data with the steady-state current response data to obtain the current response data.

[0025] Specifically, a programmable voltage source is used to set a sequence of voltage signals covering different voltage ranges and waveforms to simulate the voltage conditions that a wind turbine blade lightning protection system may encounter under various operating conditions. A programmable voltage source is a power supply device capable of outputting different voltage signals according to changes in set parameters. Through program control, the output voltage can be precisely adjusted, and different waveforms or sequences of voltage signals can be generated.

[0026] By designing a multi-channel MOSFET control circuit, a MOSFET switching matrix is ​​constructed. The switching of each channel is controlled by an independent MOSFET. This matrix allows for precise control of the voltage signal at each test point. The MOSFET switching matrix is ​​a circuit composed of multiple MOSFETs integrated in an array. It can be understood as a highly automated, electrically controlled multi-channel terminal block. The microprocessor can flexibly switch the input signal to any output channel by controlling the on / off state of each MOSFET.

[0027] A multi-channel MOSFET control circuit is a circuit composed of multiple metal-oxide-semiconductor field-effect transistors (MOSFETs). Each MOSFET can be regarded as a high-speed, precision electronic switch, capable of independently controlling the on / off state of a test channel. Multi-channel means that dozens or even hundreds of such switches can be controlled simultaneously, thereby enabling automated, sequential scanning tests of multiple points in the wind turbine blade lightning protection network.

[0028] A MOSFET switching matrix is ​​used to apply a test voltage sequence to multiple test points of the wind turbine blade lightning protection system. Each test point corresponds to a different part of the system, allowing for parallel testing of multiple areas. The wind turbine blade lightning protection system is a collective term for the devices installed on the wind turbine blades to prevent damage from lightning strikes. It includes lightning arresters at the blade tips, downleads that guide current from the blade tips to the blade roots, and connection terminals at the blade roots to conduct lightning current to the tower and ultimately to the ground.

[0029] At the instant the voltage signal changes, a current sensor records the transient current response data at each test point. While the voltage signal remains stable, the current sensor continues to monitor the current response data at each test point. The transient current test is performed at the instant the voltage signal changes, designed to capture the ability to respond to rapid voltage changes. The steady-state current test, performed after the voltage signal has stabilized, measures the current response and is used to test the long-term performance of the wind turbine blade lightning protection system under constant voltage. Integrating the transient and steady-state current response data yields complete current response data, allowing for the evaluation of its response characteristics under different voltage conditions.

[0030] For example, the programmable voltage source, under the instruction of the control algorithm, generates a preset test voltage sequence, such as 12V, 24V, and 36V. The low voltage is used for safe detection of minor defects, while the high voltage is used to simulate lightning strike stress and detect potential breakdown risks. Taking the medium-voltage 24V test as an example, a 24V DC step signal is output. Then, driven by the controller, the MOSFET switching matrix turns on the channel connected to one of the test points, such as the lightning arrester 1 in the leaf, while keeping the other channels closed, thus accurately applying the 24V test voltage to the loop at that point. Almost simultaneously, a high-precision current sensor begins to capture the instantaneous changes in the current in the loop at a high sampling rate, recording the initial transient current response data. The current sensor observes a sharp rise in current, with an overshoot peak of 450 mA, which then decays and stabilizes at 400 mA within 5 milliseconds. The transient data clearly records the overshoot and oscillation process, while the steady-state data shows that the loop resistance is R = 24V / 0.4A = 60Ω, within the normal range. However, the overshoot phenomenon in the transient response suggests that this point may have slight inductive characteristics, possibly due to the lead spacing being too close. The entire 10-point 24V test was completed automatically within 5 seconds. The MOSFET switching matrix cut off the first channel and turned on the second channel, applying the same test voltage to the next point, repeating the above process until all points were tested. The transient and steady-state data of each point were integrated according to the point number to form a complete current response dataset, which includes the record for that point: Point ID is A, Test Voltage 24V, Transient Waveform: [Time Array, Current Array], Steady-State Current 0.400A.

[0031] By precisely applying test voltage and recording transient and steady-state current response data in real time, a comprehensive test of the wind turbine blade lightning protection system was effectively achieved. By integrating transient and steady-state data, the current response capability of the wind turbine blade lightning protection system can be comprehensively evaluated, and potential problems such as poor contact and excessive resistance in the system can be detected in a timely manner.

[0032] Multidimensional feature extraction is performed on the current response data to construct a multidimensional test feature vector for local feature analysis, thereby obtaining a local feature sequence.

[0033] Furthermore, this application also includes the following steps: extracting a current signal based on the current response data; performing current time-series analysis on the current signal to construct a current-time curve; performing time-domain analysis on the current-time curve to obtain time-domain waveform features; analyzing the current signal based on the time-domain waveform features to determine current amplitude features; transforming the current signal to obtain the signal frequency and performing frequency-domain analysis to determine current frequency-domain features; calculating entropy based on the current signal to obtain entropy features and performing nonlinear distortion analysis to determine nonlinear features; grouping the current amplitude features, the current frequency-domain features, and the nonlinear features to generate multiple feature groups for matching and identification, and constructing the multi-dimensional test feature vector.

[0034] Specifically, current response data refers to the current change data output by the wind turbine blade lightning protection system after a voltage signal is applied. Based on the current response data, the current signal is extracted and current time-series analysis is performed, that is, analyzing the process of current signal change over time to construct a current-time curve. The current-time curve allows observation of current trends, such as sudden current changes, rise time, and steady-state.

[0035] Time-domain analysis is performed on the current-time curve to analyze the timing characteristics of the current signal and extract time-domain waveform features. Time-domain waveform features are extracted from the current-time curve and include current amplitude, waveform shape, pulse width, etc., reflecting the timing characteristics of the current signal. For example, if the current-time curve exhibits an abrupt change, such as the current suddenly increasing from 0A to 50mA in 10ms, then the amplitude and rise time of this abrupt change will be extracted as time-domain waveform features. Time-domain features include peak value, average current, rise time, and fall time.

[0036] The current signal is analyzed based on its time-domain waveform characteristics to extract its amplitude features, including steady-state current, peak current, rise time, overshoot, and settling time. The current amplitude feature is the maximum current value at a specific moment, reflecting instantaneous conductivity and transient response characteristics. The average value is calculated in the smooth region at the end of the curve to determine the steady-state current value; the maximum current value is found in the transient phase, i.e., the peak current; the time required for the current to rise from 10% to 90% of the steady-state value is calculated, i.e., the rise time; the degree to which the transient response exceeds the steady-state value is assessed to determine the overshoot rate. For example, a 28V DC test voltage is applied to point A of the lightning arrester at the tip of a wind turbine blade, and time-domain analysis is performed on the acquired current-time curve. The average value of the current data in the last 100 milliseconds is calculated, yielding 350mA. Based on this, the DC resistance of the circuit can be calculated as R = U / I = 28V / 0.35A = 80Ω, which is the steady-state current value. The highest point of the curve in the initial stage was identified, with a peak current of 395mA. The time it took for the current to rise from 35mA (10% of 350mA) to 315mA (90% of 350mA) was calculated, i.e., a rise time of 4.2ms. Further calculation showed an overshoot of (395-350) / 350*100%=12.9%. The current amplitude characteristics extracted for this point can be expressed as: steady-state current 350mA, peak current 395mA, rise time 4.2ms, and overshoot rate 12.9%. The 12.9% overshoot and 4.2ms rise time may indicate a certain distributed inductance in the lead wire of this circuit, perhaps due to a slight bend in the installation path. If the overshoot at other similar points on the same blade is less than 5%, the anomaly at this point will be more apparent.

[0037] A Fast Fourier Transform (FFT) is performed on the current signal to transform it from the time domain to the frequency domain, yielding a spectrum. Frequency domain features of the current, such as the dominant resonant frequency, bandwidth, and energy in specific frequency bands, are extracted from the spectrum. These frequency domain features are obtained after frequency domain analysis; the dominant resonant frequency is the frequency point where energy is most concentrated in the signal, reflecting the periodicity and frequency characteristics of the current signal. Entropy is calculated on the current signal to measure its complexity and information entropy. A higher entropy value indicates a more complex signal and greater uncertainty. Through entropy calculation and nonlinear distortion analysis, nonlinear distortion in the current signal is identified, especially signal distortion or errors caused by nonlinear components in the system. Entropy, a concept from information theory, is used to measure the disorder, randomness, or unpredictability of a signal. Calculating the sample entropy or approximate entropy of the current signal can quantify its waveform complexity and degree of nonlinear distortion. A pure, ideal conducting loop has a smooth response, while defects such as looseness or corrosion make the response signal more complex and irregular, and the entropy value increases accordingly.

[0038] The current amplitude characteristics, current frequency domain characteristics, and nonlinear characteristics are grouped and standardized, and each characteristic is assigned a unique identifier. These are then combined into a high-dimensional numerical array, namely the multi-dimensional test feature vector. The multi-dimensional test feature vector integrates the extracted features into a vector by combining the results of time-domain, frequency-domain, and nonlinear analysis, and contains the signal's characteristics in multiple dimensions.

[0039] Furthermore, this application also includes the following steps: defining multiple windows according to the current amplitude characteristics, the current frequency domain characteristics, and the nonlinear characteristics; segmenting the multi-dimensional test feature vector based on the multiple windows to generate multiple test features; traversing the multiple test features for local analysis to obtain feature distribution parameters of the multiple test features; performing gradient change analysis based on the feature distribution parameters to obtain feature change trend data; and arranging the multiple test features in chronological order according to the feature change trend data to construct the local feature sequence.

[0040] Specifically, multiple windows are set based on the extracted current amplitude characteristics, current frequency domain characteristics, and nonlinear characteristics. In signal processing, a window is a subset of data used for local analysis. Multiple windows divide the entire signal data into multiple segments, each representing the signal's characteristics within a certain time interval. Windows typically overlap to ensure the continuity and completeness of the analysis. The multi-dimensional test feature vector is segmented according to multiple windows, with each window containing a consistent number of features. Segmentation helps extract local features from different time periods or signal states, facilitating independent analysis of the signal in each time period. Segmentation refers to the process of using multiple windows to cut a complete, high-dimensional test feature vector into a series of shorter subsequences. For example, assuming a multi-dimensional test feature vector of length N, containing amplitude, frequency domain, and nonlinear characteristics, the window size is set to w, the step size to s, and the overlapping area to ws. This results in multiple windows, each containing w features.

[0041] Based on the window size and overlapping area of ​​multiple windows, the multi-dimensional test feature vector is segmented to generate multiple shorter test feature segments. Each test feature segment is traversed, and its internal feature distribution parameters, such as mean, variance, kurtosis, and skewness, are calculated to reflect the distribution characteristics of the signal features within that window.

[0042] Based on the characteristic distribution parameters, gradient change analysis is performed. This involves calculating the difference between the distribution parameters of the current window and the corresponding parameters of the previous window to obtain characteristic change trend data. For example, the mean steady-state current increased by 2% from window 1 to window 2. Gradient change analysis is used to calculate the differences in characteristic distribution parameters between adjacent windows, thereby quantifying the trend of characteristic changes as the test progresses or the order of test points changes. For example, if the current amplitude gradually increases over a certain period of time, gradient change analysis will yield positive trend data.

[0043] Based on the characteristic change trend data obtained from gradient change analysis, multiple test features are arranged in chronological order, and a local feature sequence is constructed to show the characteristic change trend of the current signal in different time periods. For example, a wind turbine blade equipped with 20 lightning rods is tested, and feature data from 20 points are obtained. The window size is set to 5 points, with an overlapping area of ​​2 points. The data from the 20 points are then divided as follows: Window 1: points 1, 2, 3, 4, 5; Window 2: points 4, 5, 6, 7, 8; Window 3: points 7, 8, 9, 10, 11, and so on. The mean and standard deviation within each window are calculated: Window 1: mean 385mA, standard deviation 8mA, data is relatively concentrated; Window 2: mean 378mA, standard deviation 15mA, mean decreases, dispersion increases; Window 3: mean 365mA, standard deviation 25mA, mean continues to decrease, dispersion increases significantly. The trend from window 1 to window 2: the mean change gradient = (378-385) = -7mA, showing a downward trend; the trend from window 2 to window 3: the mean change gradient = (365-378) = -13mA, showing an accelerated downward trend. From the blade tip (window 1) to the middle of the blade (window 3), the steady-state current shows an accelerated downward trend, and the data dispersion is getting larger and larger, suggesting that from the middle of the blade, the connection quality of the lightning protection conductor or the conductor itself is systematically degraded, such as increased oxidation and an increase in loose connection points, rather than just random faults at a single point.

[0044] By segmenting multi-dimensional test features, performing local analysis, and constructing local feature sequences, the dynamic changes of current signals over different time periods can be captured. This is particularly effective when system faults or anomalies occur, allowing for timely reflection of feature trends. Gradient change analysis improves the accuracy and safety of fault diagnosis in wind turbine blade lightning protection systems.

[0045] Feature enhancement is performed based on the local feature sequence, feature contribution is calculated for defect identification, and multi-level clustering is performed based on the identification results to generate defect classification results.

[0046] Furthermore, this application also includes the following steps: traversing the local feature sequence to normalize the multiple test features, combining the processing results to generate composite features; retrieving historical defect record logs of the wind turbine blade lightning protection system, extracting multiple defect types based on the historical defect record logs, calculating multiple feature discrimination scores based on the composite features according to the multiple defect types, filtering according to the multiple feature discrimination scores to obtain multiple key features; performing cross-analysis on the multiple key features to generate interaction parameters, associating and identifying the multiple key features according to the interaction parameters to generate multiple enhanced features; calculating the contribution of the multiple enhanced features in combination with the multiple defect types to obtain the feature contribution score.

[0047] Furthermore, this application also includes the following steps: traversing the multiple defect types to perform difference analysis on the multiple enhancement features, generating feature distribution difference parameters; performing normal distribution calculation on the multiple enhancement features to obtain feature central tendency; calculating the distribution overlap of the multiple enhancement features based on the feature central tendency and the feature distribution difference parameters to obtain distribution overlap degree; performing weighted calculation on the multiple defect types based on the distribution overlap degree and the multiple enhancement features to obtain weighted contribution degree of multiple defect types; summing the weighted contribution degrees of the multiple defect types to obtain the feature contribution degree.

[0048] Specifically, the local feature sequence is traversed to normalize multiple test features, unifying their scale and range, eliminating dimensional differences between features, and ensuring fair comparison of each feature in subsequent analysis. The normalized features are then combined to generate composite features, fusing key information from multiple features to obtain a more representative feature for easier subsequent analysis. Composite features are new features generated by combining, weighting, or transforming multiple features, capable of fusing information across multiple dimensions and enhancing feature representativeness.

[0049] Retrieve the historical defect log of the wind turbine blade lightning protection system and extract the different defect types. The historical defect log refers to the various defect information recorded during the past use of the wind turbine blade lightning protection system, including the occurrence, severity, and time of each defect. Defect types refer to the different fault modes or defect types that may exist in the wind turbine blade lightning protection system, such as abnormal current, poor conductivity, short circuit, and poor contact.

[0050] Based on the composite feature and defect type, the discriminant power of each feature against different defect types is calculated. A higher discriminant power indicates a better ability to diagnose different defect types. The calculated discriminant power will be used to select the key features that best distinguish defect types. Feature discriminant power measures a feature's ability to differentiate between different defect types. A higher discriminant power indicates a stronger ability to identify fault types. For example, there are 50 cases of loose connections in the historical log. The discriminant power of composite feature 1 between this defect type and the normal type is calculated to be 45.2. Meanwhile, the discriminant power of the single feature with normalized steady-state current is 20.1, and the discriminant power of the single feature with normalized rise time is 15.5. Clearly, composite feature 1 has a higher discriminant power. From all features, the 10 with the highest discriminant power are selected, and composite feature 1 is chosen as the key feature due to its high discriminant power value.

[0051] Cross-analysis is performed on multiple key features to generate interaction parameters, analyzing whether there are interactions, such as synergy or cancellation, between two or more key features. The interaction parameters quantify the strength and magnitude of this interaction. Multiple key features are associated based on the interaction parameters, and enhanced features are created accordingly. For example, composite feature 1 interacts with another key feature, the frequency domain peak amplitude. When composite feature 1 > 0.7 and the frequency domain peak amplitude < 0.3, the accuracy of diagnosing loose connections improves from 85% to 96%. This combined effect is assigned a high interaction parameter of 0.85. An enhanced feature is generated, prioritizing this combination condition in the decision rule. Enhanced features are new feature representations formed based on key features, considering the interactions between features. They can be combinations of the original key features, such as products, or specially labeled pairs of related features in the model, aiming to better capture complex defect patterns.

[0052] The algorithm iterates through multiple defect types, performs differential analysis on each enhanced feature, and calculates the feature's distribution range across different defect types. The difference in feature distribution range reflects the magnitude of feature variation across different defect types. The feature distribution difference parameter is a quantitative parameter that measures the difference in the distribution range of a feature across different defect types, reflecting the magnitude and stability of feature variation under different defect states. It is typically used to determine the feature's ability to distinguish between defect types.

[0053] The enhanced features are calculated using a normal distribution to estimate their distribution under normal conditions. The mean and standard deviation of the features are obtained through the normal distribution model, revealing the central tendency of the features. The central tendency refers to the central value of the feature's distribution under normal conditions, usually the mean, reflecting the main location of concentration of the feature under normal circumstances. For example, assuming that the current amplitude feature follows a normal distribution under normal conditions, with a mean of 0.6 mA and a standard deviation of 0.05 mA, then the central tendency of this feature is 0.6 mA, indicating that the main values ​​of the current amplitude under normal conditions are concentrated around 0.6 mA.

[0054] Based on the feature central tendency and feature distribution difference parameters, the distribution overlap of multiple enhanced features is calculated, and the degree of overlap of feature distribution under normal and various defect states is calculated. A higher degree of overlap indicates a smaller difference between normal and defect states, making fault identification more difficult. Conversely, a lower degree of overlap indicates a larger difference in features under different states, making fault identification easier.

[0055] Based on distribution overlap and central tendency, the contributions of multiple enhancement features to different defect types are weighted and calculated. The weighted contribution of each defect type reflects the magnitude of that defect type's contribution to fault diagnosis. The higher the feature contribution, the greater its importance in fault diagnosis. The weighted contributions of all defect types are summed to obtain the final feature contribution. The feature contribution is the overall importance score of a particular enhancement feature in identifying all defect types; it is the weighted sum or direct sum of its weighted contributions to each defect type. For example, taking the enhancement feature overshoot-entropy product and the defect type "loose downlead connection" as an example, the overshoot-entropy product has a mean of 15 and a standard deviation of 3. 40 defect samples with loose downlead connections were found, with a mean of 55 and a standard deviation of 10. After calculating and combining the standard deviations, a Cohen d-value of 5.2 is obtained. According to Cohen's standard, d > 0.8 is considered a large effect size, indicating that the overshoot-entropy product has a very strong ability to distinguish loose downlead connections. Under normal conditions, N(15,3) and the defect distribution N(55,10) are confirmed. The overlap area of ​​the two distributions is calculated, yielding an overlap of 0.008, indicating that the two distributions are almost completely separated. Assuming the weight of the defect type "loose downlead connection" is 1.0, the weighted contribution of the overshoot-entropy product to this defect is (1-0.008)*1.0=0.992. Assuming another defect, conductor corrosion, has a lower weighted contribution of 0.3, the global feature contribution of the overshoot-entropy product is 0.992+0.3=1.292. Similarly, the global feature contribution values ​​of all enhanced features are calculated and sorted. Features with high global feature contribution values ​​are considered core features and play a decisive role in defect identification.

[0056] This study comprehensively analyzes the characteristics of wind turbine blade lightning protection systems from multiple dimensions, identifies key features related to defect types, and enhances the discriminative power of features through methods such as feature discrimination and cross-analysis. Using a data-driven approach, the most discriminative key features are automatically selected from a large number of potential features, and further explored to uncover deep correlations between features to form enhanced features. Finally, rigorous statistical calculations quantify the importance of each feature, i.e., its feature contribution.

[0057] Furthermore, this application also includes the following steps: performing defect matching according to the feature contribution degree to generate a defect matching degree; when the defect matching degree is greater than a preset matching degree threshold, identifying and determining a target defect dataset; coarsely classifying the target defect dataset according to the physical characteristics of the defects to generate a first clustering result; performing defect impact analysis based on the first clustering result, and further classifying the first clustering result according to the defect impact factor to generate a second clustering result; locating the defect distribution position based on the target defect dataset, and performing spatial clustering analysis on the second clustering result according to the defect distribution position to generate a third clustering result; performing defect time extrapolation based on the target defect dataset, and performing temporal clustering analysis on the third clustering result according to the defect evolution data to generate the defect classification result.

[0058] Specifically, defect matching is performed based on feature contribution to determine the matching degree of each defect. When the matching degree is greater than a preset matching degree threshold, the target defect dataset is identified, and the most likely defect dataset is selected to reduce the complexity of the analysis. The target defect dataset is coarsely classified according to the physical characteristics of the defects. Based on the physical characteristics of the defects, the defect data is divided into different categories, generating the first clustering result. Coarse classification is a preliminary classification process for the target defect dataset, mainly based on the physical characteristics of the defects. At this stage, the classification is relatively broad and is usually based on simple feature distinctions. The first clustering result is the first clustering analysis result formed based on the data after coarse classification. The first clustering result obtained from coarse classification is similar to answering the question of what physical properties the defect has.

[0059] Defect impact analysis is performed on the coarse classification results to determine defect impact factors. Defect impact factors are standards for assessing the degree of impact of defects on the system, such as the magnitude of current anomalies or the duration of defect persistence. The impact factors of defects in each cluster are analyzed, and this serves as the basis for further refined classification. Defect impact factors typically consider the severity of the defect and its potential risks to the system, such as the magnitude and duration of current anomalies.

[0060] The first clustering results are further subdivided based on the defect impact factors, breaking down the defect data into smaller categories. This finer subdivision allows for more precise defect grouping, thereby improving the accuracy of fault diagnosis. For example, high-resistivity defects can be subdivided into mild corrosion (resistance deviation <30%) and severe corrosion / early-stage open circuit (resistance deviation >30%). The second clustering results obtained from this finer subdivision are similar to answering the question of defect severity.

[0061] Based on the target defect dataset, the location of defect distribution is determined, i.e., the location of defects in the wind turbine blade lightning protection system is identified. The location of defects may be related to specific parts of the equipment or external environmental conditions. After determining the defect distribution location, spatial clustering analysis is used to aggregate defects in similar locations from the second clustering results, forming spatial groups. Defects that are spatially close are grouped into the same cluster, resulting in the third clustering result, which essentially answers where the defects are concentrated.

[0062] Defect time-lapse is performed on the target defect dataset. Based on the defect evolution data, development trends are analyzed to obtain defect evolution data, including the evolution process, frequency of occurrence, and severity of defects. Using data from the past 3-5 detections, defect time-lapse is performed on each defect cluster in the third clustering result. The curves of key parameters changing over time are fitted to obtain defect evolution data. The third clustering result is then further clustered based on the evolution rate to obtain defect classification results, clearly showing the state, severity, and future evolution trend of each defect. Through multi-level clustering analysis, a comprehensive and refined understanding of defects is achieved, from qualitative to quantitative, from one-dimensional to multi-dimensional, and from current state to evolution. This not only accurately identifies the physical type of defects but also assesses their severity, locates their spatial distribution patterns, and predicts their development trends.

[0063] Based on the defect classification results, a global health analysis is performed on the wind turbine blade lightning protection system. Based on the health score results, the defects in lightning protection conduction performance are traced, defect attribution points are generated for monitoring, and the monitoring results are fed back to the multi-channel MOS transistor control circuit for feedback optimization, thus constructing a test optimization closed loop.

[0064] Furthermore, this application also includes the following steps: spatial distribution assessment of the wind turbine blade lightning protection system based on the defect classification results, generating defect spatial distribution characteristics; usage prediction based on the defect spatial distribution characteristics, obtaining service life prediction results; global health calculation based on the service life prediction results, generating an initial health score; introducing environmental parameters to correct the initial health score, generating a health rating result; correlation analysis of the multiple test points based on the health rating result, generating data correlation coefficients; defect impact analysis of the health rating result based on the data correlation coefficients, sorting the results in descending order according to the defect impact coefficients, extracting multiple key defect types; defect tracing of lightning protection conductivity based on the multiple key defect types, generating defect attribution points.

[0065] Specifically, the spatial distribution of defects in the wind turbine blade lightning protection system is assessed based on the defect classification results. The spatial distribution patterns of defects within the system are analyzed to obtain the spatial distribution characteristics of the defects. For example, some defects may be concentrated at the joints of the wind turbine blades, while others may be distributed across the blade surface. The spatial distribution characteristics describe the spatial distribution of defects, showing the areas, locations, and potential clustering trends of defects, which helps in identifying medium- and high-risk areas.

[0066] Based on the spatial distribution characteristics of defects, usage prediction is performed. The current status is analyzed to predict the remaining service life of the wind turbine blade lightning protection system, considering defect severity, system load, and environmental factors. After obtaining the service life prediction results, a global health calculation is performed, comprehensively considering all defects, spatial distribution, and environmental conditions of the wind turbine blade lightning protection system to obtain a score representing its health status. The calculation formula includes a weighted average of various factors, integrating various influencing factors to obtain the current health score of the system. The initial health score is a quantitative score that integrates the number, severity, spatial distribution, and evolution trend of defects, used to intuitively reflect the overall status of the wind turbine blade lightning protection system, typically a value from 0 to 100.

[0067] Environmental parameters are introduced for correction to obtain the health score. These parameters are external environmental data of the wind turbine's location, including annual average wind speed, air salinity, temperature and humidity, and the number of thunderstorm days, which significantly affect the rate of defect development. For example, if the wind turbine is located in a coastal area with high salinity and frequent thunderstorms, the corrosion and lightning strike risk factors are multiplied by the defect deterioration rate based on historical data models, thereby lowering the health score to accurately reflect the risks in harsh environments.

[0068] Based on the health score results, correlation analysis is performed on multiple test points to calculate data correlation coefficients, identify the correlations between different test points, and thus discover potential sources of failure. Defect impact analysis is then performed on the health score results according to the defect impact coefficient to screen out the most influential defect types. Points with high correlation coefficients are the cause of lower overall health scores; therefore, the defect impact coefficient for each defect type is calculated, such as the average correlation coefficient of a certain type of defect multiplied by the severity of that type of defect. The data correlation coefficient is a metric obtained by performing correlation analysis on data from multiple test points; the higher the correlation coefficient, the greater the impact of the condition of that point on the overall health. The defect impact coefficient is an indicator that combines the data correlation coefficient and the severity of the defect itself, used to rank which defect types pose the greatest threat to the health of the wind turbine blade lightning protection system.

[0069] Based on the defect impact coefficient, defects are sorted in descending order to identify the most priority defect categories and extract multiple key defect types. Defect source tracing is then performed based on these key defect types. By analyzing their spatial distribution and electrical characteristics, the root cause is inferred, and defect attribution points are generated. By analyzing the distribution and characteristics of key defect types, the root cause leading to that type of defect is inferred in reverse, and one or more most likely specific locations are pinpointed. For example, source tracing reveals that severe corrosion at all blade root downlead connections occurs on the same type of connector, thus locking the attribution point to the blade root A-type connector process interface and recommending an inspection of the batch's installation process or material quality. For instance, for a wind turbine, the initial health score calculated based on defects is 72 points. Given that the area has an average of 90 thunderstorm days per year, high air chloride ion concentration, and an environmental coefficient of 0.85, the health score is 72 * 0.85 = 61.2 points, triggering a medium-risk alarm. Correlation analysis revealed that the health score was primarily lowered by anomalies at three locations, with correlation coefficients of -0.90, -0.88, and -0.85 respectively. A negative correlation indicates that the worse the data at a particular location, the lower the health score. All three locations were classified as accelerated corrosion at the leading edge of the blade. This defect type had the highest calculated impact coefficient and was therefore listed as a critical defect type. All three locations are situated in the same area of ​​the lightning protection strip at the leading edge of the blade. Protective paint was applied during the last maintenance. It is speculated that the accelerated corrosion may be related to defects or peeling of the protective paint coating, leading to direct exposure of the metal in this area to salt spray. The defect attribution point is determined to be the integrity of the protective coating on the lightning protection strip at the leading edge of the blade (35-40 meters from the blade root). Maintenance recommendations are: prioritize high-resolution image inspection and coating repair in this area.

[0070] By accurately assessing the overall health of the wind turbine blade lightning protection system and predicting potential failures, environmental parameters are introduced to correct the initial health score, making the health assessment more consistent with the actual environment. Data correlation coefficient analysis and defect impact analysis help to discover the correlation between different test points, thereby identifying key defect types.

[0071] Furthermore, this application also includes the following steps: tracing the multiple test points according to a time sequence based on the multiple key defect types to identify the abnormal starting point; performing causal reasoning based on the abnormal starting point to construct a defect cause classification list for attribution and determine the defect attribution point; enhancing monitoring based on the defect attribution point to construct a key monitoring scheme; executing the key monitoring scheme to monitor and identify the multiple test points, determining multiple key test monitoring points; transmitting the monitoring results of the multiple key test monitoring points to the multi-channel MOS transistor control circuit in real time to generate control feedback parameters; dynamically adjusting and optimizing the test of the multi-channel MOS transistor control circuit based on the control feedback parameters, constructing a self-learning mechanism for iterative improvement, and constructing a test optimization closed loop.

[0072] Specifically, based on the extracted key defect types, corresponding test points are selected. Time series data collected from these test points are then traced back to plot the resistance value over time at each point, identifying the specific time when the resistance begins to rise abnormally—that is, the anomaly initiation point of the defect—and determining the time of defect occurrence and its early signs. Causal reasoning is then performed on the anomaly identification points. Using known data and patterns, the cause of a phenomenon or defect is inferred, determining the root cause of the defect. The defect cause classification list is a list derived from causal reasoning, listing the possible causes of a particular defect or anomaly, with each cause associated with a specific defect type or failure mode.

[0073] Based on the defect cause classification list, the cause of the defect is determined by identifying the defect attribution point, which is the specific location where the defect occurred and a key monitoring area for subsequent processing. More targeted monitoring strategies are developed for the vulnerabilities exposed by the defect attribution point, including increasing testing frequency, using more precise testing parameters, and adding auxiliary sensors for joint monitoring. The key monitoring plan is a key monitoring plan developed based on the defect attribution point and other key factors, targeting high-risk areas and important test points. This plan will strengthen monitoring of key monitoring points, i.e., the defect attribution point, to promptly identify potential problems. For example, if the log shows that the blade underwent routine maintenance involving the blade root area, a defect cause classification list is constructed. It is inferred that the accelerated corrosion is likely related to improper torque or incomplete curing of the sealant during the reinstallation of the connector after maintenance. Therefore, the root cause is attributed to a flaw in the post-maintenance installation process, and the connector is identified as the defect attribution point. Develop a key monitoring plan: Enhance monitoring of the defect attribution point and three adjacent connectors of the same type. The plan includes: increasing the testing frequency from quarterly to monthly; and adding a 15V low-voltage micro-current test to more sensitively monitor early oxidation of the contact surface.

[0074] Through a focused monitoring scheme, multiple test points are identified and marked to determine those requiring intensive monitoring. These key test monitoring points are specifically marked in the scheme and require enhanced monitoring; they are typically the defect attribution point itself and its surrounding physically related points. The monitoring results from these key points are transmitted in real-time to the multi-channel MOSFET control circuit. After analyzing this data, control feedback parameters are generated: if the resistance increment exceeds 5% for two consecutive months, an automatic command is triggered to temporarily increase the test voltage from the standard 24V to 30V for more stringent stress testing to confirm the defect development rate. The control feedback parameters are a set of instructions generated from the real-time monitoring results of the key test points, directly guiding adjustments in the multi-channel MOSFET control circuit.

[0075] Based on control feedback parameters, the multi-channel MOSFET control circuit is dynamically adjusted and optimized. By comparing the effectiveness of monitoring data before and after adjustment, a self-learning mechanism is constructed. Each feedback message causes the multi-channel MOSFET control circuit to self-correct, thereby continuously optimizing the testing process and gradually improving testing accuracy and efficiency. This endows the entire continuity test with dynamic adaptability and self-evolution capabilities, realizing the ultimate intelligence of operation and maintenance strategies. It is no longer a static, one-way detection-diagnosis tool, but a system that can learn the root causes from historical faults, dynamically adjust monitoring priorities, and continuously optimize its own detection strategies. By constructing a test optimization closed loop, it proactively focuses on the most vulnerable and critical links, capturing risk changes with the highest efficiency. Ultimately, it achieves a leap from preventing faults to predicting and preventing faults from escalating, maximizing the safety and reliability of the wind turbine blade lightning protection system while optimizing the operation and maintenance costs throughout the entire life cycle.

[0076] In summary, the machine learning-based lightning protection continuity testing method for wind turbine blades provided in this application has the following beneficial effects: Current tests are performed on multiple test points of the wind turbine blade lightning protection system using a multi-channel MOSFET control circuit to obtain current response data. Multi-dimensional feature extraction is performed on the current response data to construct a multi-dimensional test feature vector for local feature analysis, resulting in a local feature sequence. Feature enhancement is performed based on the local feature sequence, and feature contribution is calculated for defect identification. Multi-level clustering is then performed based on the identification results to generate defect classification results. A global health analysis is conducted on the wind turbine blade lightning protection system based on the defect classification results. Defects in lightning protection conduction performance are traced based on the health score results, generating defect attribution points for monitoring. The monitoring results are fed back to the multi-channel MOSFET control circuit for feedback optimization, constructing a test optimization closed loop. In other words, by using a multi-channel MOSFET control circuit to perform current tests on multiple test points, the collected data is used to identify defects. Based on the defect classification results, a global health analysis of the wind turbine blade lightning protection system is performed to determine the defect attribution points, monitor them, and feed back the monitoring results to the multi-channel MOSFET control circuit for feedback optimization. This achieves a shift from passive maintenance to predictive maintenance, significantly improving testing efficiency and thus enhancing the safe operation and maintenance level of the wind turbine lightning protection system.

[0077] Example 2: Based on the same inventive concept as the machine learning-based wind turbine blade lightning protection continuity testing method in Example 1, this application also provides a machine learning-based wind turbine blade lightning protection continuity testing system. Please refer to the appendix. Figure 2 The machine learning-based wind turbine blade lightning protection continuity testing system includes: The current testing module 11 is used to perform current tests on multiple test points of the wind turbine blade lightning protection system through a multi-channel MOS transistor control circuit to obtain current response data; the local feature analysis module 12 is used to extract multi-dimensional features from the current response data, construct a multi-dimensional test feature vector for local feature analysis, and obtain a local feature sequence; the defect clustering module 13 is used to perform feature enhancement based on the local feature sequence, calculate the feature contribution degree for defect identification, perform multi-level clustering based on the identification results, and generate defect classification results; the defect tracing module 14 is used to perform global health analysis on the wind turbine blade lightning protection system based on the defect classification results, trace the defects in lightning protection conduction performance based on the health score results, generate defect attribution points for monitoring, and feed back the monitoring results to the multi-channel MOS transistor control circuit for feedback optimization to construct a test optimization closed loop.

[0078] Furthermore, the current testing module 11 in the machine learning-based wind turbine blade lightning protection continuity testing system is also used for: performing voltage analysis using a programmable voltage source to determine a test voltage sequence, the test voltage sequence including a test voltage signal; constructing a MOSFET switching matrix based on a multi-channel MOSFET control circuit, the MOSFET switching matrix independently controlling each channel MOSFET control circuit; applying the test voltage signal to the multiple test points of the wind turbine blade lightning protection system through the MOSFET switching matrix; performing transient current testing on the multiple test points using a current sensor to obtain transient current response data; performing steady-state current testing on the multiple test points using a current sensor to obtain steady-state current response data; and integrating the transient current response data with the steady-state current response data to obtain the current response data.

[0079] Furthermore, the local feature analysis module 12 in the machine learning-based wind turbine blade lightning protection continuity test system is also used for: extracting current signals based on the current response data; performing current time-series analysis based on the current signals to construct a current-time curve; performing time-domain analysis based on the current-time curve to obtain time-domain waveform features; analyzing the current signals based on the time-domain waveform features to determine current amplitude features; transforming the current signals to obtain signal frequencies for frequency-domain analysis to determine current frequency-domain features; calculating entropy values ​​based on the current signals to obtain entropy value features for nonlinear distortion analysis to determine nonlinear features; grouping the current amplitude features, current frequency-domain features, and nonlinear features to generate multiple feature groups for matching and identification, and constructing the multi-dimensional test feature vector.

[0080] Furthermore, the local feature analysis module 12 in the machine learning-based wind turbine blade lightning protection continuity test system is also used for: defining multiple windows according to the current amplitude feature, the current frequency domain feature, and the nonlinear feature; segmenting the multi-dimensional test feature vector based on the multiple windows to generate multiple test features; traversing the multiple test features for local analysis to obtain the feature distribution parameters of the multiple test features; performing gradient change analysis based on the feature distribution parameters to obtain feature change trend data; and arranging the multiple test features in chronological order according to the feature change trend data to construct the local feature sequence.

[0081] Furthermore, the defect clustering module 13 in the machine learning-based wind turbine blade lightning protection continuity testing system is also used for: traversing the local feature sequence to normalize the multiple test features, combining the processing results to generate composite features; retrieving the historical defect record log of the wind turbine blade lightning protection system, extracting multiple defect types based on the historical defect record log, calculating multiple feature discrimination scores based on the composite features according to the multiple defect types, filtering according to the multiple feature discrimination scores to obtain multiple key features; performing cross-analysis on the multiple key features to generate interaction parameters, associating and identifying the multiple key features according to the interaction parameters to generate multiple enhanced features; and calculating the contribution based on the multiple enhanced features combined with the multiple defect types to obtain the feature contribution score.

[0082] Furthermore, the defect clustering module 13 in the machine learning-based wind turbine blade lightning protection continuity testing system is also used for: traversing the multiple defect types to perform difference analysis on the multiple enhancement features and generating feature distribution difference parameters; performing normal distribution calculation on the multiple enhancement features to obtain feature central tendency; calculating the distribution overlap of the multiple enhancement features according to the feature central tendency and the feature distribution difference parameters to obtain distribution overlap; performing weighted calculation on the multiple defect types based on the distribution overlap and the multiple enhancement features to obtain the weighted contribution of the multiple defect types; and summing the weighted contribution of the multiple defect types to obtain the feature contribution.

[0083] Furthermore, the defect clustering module 13 in the machine learning-based wind turbine blade lightning protection continuity testing system is also used for: performing defect matching according to the feature contribution degree to generate a defect matching degree; when the defect matching degree is greater than a preset matching degree threshold, identifying and determining a target defect dataset; coarsely classifying the target defect dataset according to the physical characteristics of the defects to generate a first clustering result; performing defect impact analysis based on the first clustering result, and finely classifying the first clustering result according to the defect impact factor to generate a second clustering result; locating the defect distribution location based on the target defect dataset, and performing spatial clustering analysis on the second clustering result according to the defect distribution location to generate a third clustering result; performing defect time extrapolation based on the target defect dataset, and performing temporal clustering analysis on the third clustering result according to the defect evolution data to generate the defect classification result.

[0084] Furthermore, the defect tracing module 14 in the machine learning-based wind turbine blade lightning protection continuity testing system is also used for: spatial distribution evaluation of the wind turbine blade lightning protection system based on the defect classification results, generating defect spatial distribution characteristics; usage prediction based on the defect spatial distribution characteristics, obtaining service life prediction results, global health calculation based on the service life prediction results, generating an initial health score; introducing environmental parameters to correct the initial health score, generating a health rating result; performing correlation analysis on the multiple test points based on the health rating result, generating data correlation coefficients; performing defect impact analysis on the health rating result according to the data correlation coefficients, sorting the results in descending order based on the defect impact coefficients, and extracting multiple key defect types; and tracing the defects in lightning protection continuity performance based on the multiple key defect types, generating the defect attribution points.

[0085] Furthermore, the defect tracing module 14 in the machine learning-based wind turbine blade lightning protection continuity testing system is also used for: tracing the multiple test points according to a time sequence based on the multiple key defect types to identify the abnormal starting point; performing causal reasoning based on the abnormal starting point to construct a defect cause classification list for attribution and determine the defect attribution point; enhancing monitoring based on the defect attribution point to construct a key monitoring scheme; executing the key monitoring scheme to monitor and identify the multiple test points, determining multiple key test monitoring points; transmitting the monitoring results of the multiple key test monitoring points to the multi-channel MOS transistor control circuit in real time to generate control feedback parameters; and dynamically adjusting and optimizing the multi-channel MOS transistor control circuit based on the control feedback parameters, constructing a self-learning mechanism for iterative improvement, and constructing a test optimization closed loop.

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Figure 1 The machine learning-based wind turbine blade lightning protection continuity test method and specific examples in Example 1 are also applicable to the machine learning-based wind turbine blade lightning protection continuity test system in this example. Through the foregoing detailed description of the machine learning-based wind turbine blade lightning protection continuity test method, those skilled in the art can clearly understand the machine learning-based wind turbine blade lightning protection continuity test system in this example. Therefore, for the sake of brevity, it will not be described in detail here.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0088] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A machine learning-based method for testing the lightning protection continuity of wind turbine blades, characterized in that, include: Current response data was obtained by performing current tests at multiple test points of the wind turbine blade lightning protection system using a multi-channel MOSFET control circuit. Multidimensional feature extraction is performed on the current response data to construct a multidimensional test feature vector for local feature analysis, thereby obtaining a local feature sequence. Based on the local feature sequence, feature enhancement is performed, feature contribution is calculated for defect identification, and multi-level clustering is performed based on the identification results to generate defect classification results. Based on the defect classification results, a global health analysis is performed on the wind turbine blade lightning protection system. Based on the health score results, the defects in lightning protection conduction performance are traced, defect attribution points are generated for monitoring, and the monitoring results are fed back to the multi-channel MOS transistor control circuit for feedback optimization, thus constructing a test optimization closed loop.

2. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 1, characterized in that, Current response data was obtained by performing current tests at multiple test points of the wind turbine blade lightning protection system using a multi-channel MOSFET control circuit. The method included: Voltage analysis is performed using a programmable voltage source to determine a test voltage sequence, which includes a test voltage signal. A MOSFET switching matrix is ​​constructed based on a multi-channel MOSFET control circuit, wherein the MOSFET switching matrix independently controls the MOSFET control circuit of each channel. The test voltage signal is applied to the multiple test points of the wind turbine blade lightning protection system through the MOS transistor switching matrix; Current transient test is performed on the multiple test points using a current sensor to obtain current transient response data; The current steady-state test is performed on the multiple test points using a current sensor to obtain current steady-state response data. The transient current response data and the steady-state current response data are integrated to obtain the current response data.

3. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 1, characterized in that, The method for extracting multidimensional features from the current response data and constructing a multidimensional test feature vector includes: Based on the current response data, the current signal is extracted, and current time series analysis is performed on the current signal to construct a current-time curve. Time-domain analysis is performed based on the current-time curve to obtain time-domain waveform characteristics. The current signal is then analyzed based on these time-domain waveform characteristics to determine the current amplitude characteristics. Based on the current signal, the signal frequency is transformed and frequency domain analysis is performed to determine the current frequency domain characteristics. Entropy is calculated based on the current signal, and nonlinear distortion analysis is performed to determine the nonlinear characteristics. Based on the current amplitude characteristics, current frequency domain characteristics, and nonlinear characteristics, multiple feature groups are generated for matching and identification, and the multi-dimensional test feature vector is constructed.

4. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 3, characterized in that, Constructing multi-dimensional test feature vectors for local feature analysis to obtain local feature sequences, methods include: Multiple windows are defined based on the current amplitude characteristics, the current frequency domain characteristics, and the nonlinear characteristics; The multi-dimensional test feature vector is segmented using multiple windows to generate multiple test features. By traversing the multiple test features and performing local analysis, the feature distribution parameters of the multiple test features are obtained. Gradient change analysis is performed based on the aforementioned feature distribution parameters to obtain feature change trend data; Based on the characteristic change trend data, the multiple test features are arranged in chronological order to construct the local feature sequence.

5. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 4, characterized in that, Feature enhancement is performed based on the local feature sequence, and the feature contribution is calculated. The method includes: The local feature sequence is traversed to normalize the multiple test features, and the processing results are combined to generate composite features. Retrieve historical defect logs of the wind turbine blade lightning protection system, extract multiple defect types from the historical defect logs, calculate multiple feature discrimination scores based on the composite features according to the multiple defect types, and filter according to the multiple feature discrimination scores to obtain multiple key features; Cross-analysis is performed on the multiple key features to generate interaction parameters. The multiple key features are then associated and identified according to the interaction parameters to generate multiple enhanced features. The contribution of each feature is calculated by combining the multiple enhancement features with the multiple defect types.

6. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 5, characterized in that, The contribution of a feature is calculated based on the combination of the multiple enhancement features and the multiple defect types to obtain the feature contribution. The method includes: Perform differential analysis on the multiple enhancement features by traversing the multiple defect types, and generate feature distribution difference parameters; Normal distribution calculation is performed on the multiple enhanced features to obtain the feature central tendency; Based on the feature central tendency and the feature distribution difference parameter, the distribution overlap of the multiple enhanced features is calculated to obtain the distribution overlap degree; Based on the distribution overlap and the multiple enhancement features, the multiple defect types are weighted and calculated to obtain the weighted contribution of the multiple defect types. The weighted contributions of the multiple defect types are summed to obtain the feature contribution.

7. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 1, characterized in that, Defect identification is performed by calculating feature contribution, and multi-level clustering is conducted based on the identification results to generate defect classification results. The methods include: Defect matching is performed according to the feature contribution degree to generate a defect matching degree. When the defect matching degree is greater than a preset matching degree threshold, the target defect dataset is identified and determined. The target defect dataset is coarsely classified according to the physical characteristics of the defects to generate the first clustering result; Based on the first clustering result, a defect impact analysis is performed. The first clustering result is further subdivided according to the defect impact factor to generate a second clustering result. Based on the target defect dataset, the defect distribution locations are located, and spatial clustering analysis is performed on the second clustering results according to the defect distribution locations to generate a third clustering result; Based on the target defect dataset, defect time extrapolation is performed, and temporal clustering analysis is conducted on the third clustering result according to the defect evolution data to generate the defect classification result.

8. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 1, characterized in that, Based on the defect classification results, a global health analysis is performed on the wind turbine blade lightning protection system. Based on the health score results, defect sources in the lightning protection conductivity are traced, and defect attribution points are generated. The method includes: Based on the defect classification results, the spatial distribution of the wind turbine blade lightning protection system is evaluated, and the spatial distribution characteristics of defects are generated. Based on the spatial distribution characteristics of the defects, a usage prediction is performed to obtain a service life prediction result. A global health calculation is then performed according to the service life prediction result to generate an initial health score. Environmental parameters are introduced to correct the initial health score, generating a health score result; Based on the health score results, a correlation analysis is performed on the multiple test points to generate data correlation coefficients; The health score results are analyzed for defects based on the data correlation coefficients, and then sorted in descending order according to the defect impact coefficients to extract multiple key defect types. Based on the aforementioned multiple key defect types, defect attribution points are generated to trace the source of lightning protection conductivity defects.

9. The machine learning-based lightning protection continuity testing method for wind turbine blades as described in claim 8, characterized in that, Based on the health score results, defects in lightning protection conductivity are traced, defect attribution points are generated for monitoring, and the monitoring results are fed back to the multi-channel MOSFET control circuit for feedback optimization, thus constructing a test optimization closed loop. The method includes: Based on the aforementioned multiple key defect types, the multiple test points are traced according to time sequence to identify the starting point of the anomaly; Based on the anomaly initiation point, causal reasoning is performed to construct a defect cause classification list for attribution and to determine the defect attribution point. Based on the aforementioned defect attribution points, monitoring is enhanced, and a key monitoring scheme is constructed. The key monitoring plan is implemented to monitor and identify the multiple test points, thereby identifying multiple key test monitoring points; The monitoring results of the multiple key test monitoring points are transmitted to the multi-channel MOS transistor control circuit in real time to generate control feedback parameters; The control feedback parameters are used to dynamically adjust and optimize the control circuit of the multi-channel MOS transistor, and a self-learning mechanism is constructed for iterative improvement to build a test optimization closed loop.

10. A machine learning-based lightning protection continuity testing system for wind turbine blades, characterized in that, The steps for implementing the machine learning-based wind turbine blade lightning protection continuity testing method according to any one of claims 1 to 9, wherein the machine learning-based wind turbine blade lightning protection continuity testing system comprises: The current testing module is used to perform current testing on multiple test points of the wind turbine blade lightning protection system through a multi-channel MOSFET control circuit to obtain current response data. The local feature analysis module is used to extract multi-dimensional features from the current response data, construct a multi-dimensional test feature vector for local feature analysis, and obtain a local feature sequence. The defect clustering module is used to enhance features based on the local feature sequence, calculate feature contribution for defect identification, perform multi-level clustering based on the identification results, and generate defect classification results. The defect tracing module is used to perform a global health analysis of the wind turbine blade lightning protection system based on the defect classification results, trace the defects in lightning protection conduction performance based on the health score results, generate defect attribution points for monitoring, and feed back the monitoring results to the multi-channel MOS transistor control circuit for feedback optimization, thus constructing a test optimization closed loop.

Images