Artificial intelligence automatic real-time analysis and diagnosis method for early failure of wind turbine drive chain

Abstract

This invention discloses a diagnostic method for early-stage faults in wind turbine generator drivetrains using AI-based automatic real-time analysis. The method includes: acquiring raw vibration time-domain signals, performing multi-band filtering, and segmenting them into multiple time segments of varying durations; converting each segment into an image, inputting it into a pre-trained impact waveform recognition AI model to identify the location and quantity of impact waveforms; performing time-domain feature analysis on each segment, extracting candidate impact events that meet physical constraints, and evaluating their periodicity parameters, classifying those reaching a threshold as suspected fault impacts; cross-validating the AI ​​recognition results with the suspected fault impacts obtained from the time-domain analysis, and confirming the presence of a suspected fault impact in the segment if a consistency condition is met; and comprehensively outputting an early fault diagnosis conclusion based on the judgment results of single sampling of multiple segments and continuous sampling. This invention achieves early, automatic, and highly reliable diagnosis of drivetrain faults.

Description

Technical Field

[0001] This invention relates to the field of fault identification methods, and more specifically to an AI-based automatic real-time analysis and diagnosis method for early faults in the transmission chain of wind turbine generator sets. Background Technology

[0002] The drivetrain of a wind turbine generator set (mainly including the main bearing, gearbox, and generator bearing) is a core component of the wind turbine, and its operating status directly affects the unit's power generation efficiency, operation and maintenance costs, and overall safety. Therefore, early detection and automatic diagnosis of drivetrain faults are crucial requirements for improving the economic benefits of wind farms and ensuring the safe operation of equipment.

[0003] Currently, the industry commonly uses online vibration monitoring systems (CMS) to monitor the condition of the transmission chain. CMS systems continuously collect vibration time-domain signals by installing accelerometers at key measuring points and perform fault diagnosis based on these signals.

[0004] However, existing mainstream diagnostic methods have serious shortcomings and are unable to meet the two core requirements of "early detection" and "automation," which are specifically manifested in the following three aspects: The first level: Existing methods heavily rely on spectrum-based characteristic frequency analysis combined with the total vibration threshold value specified in the relevant standard GB / T35854-2018 for judgment. This paradigm is only applicable to specific intermediate stages of fault development and has blind spots in its diagnostic capabilities for early-stage faults. Based on the physical process of fault development, it can be divided into four stages: Phase 1: In the very early stages, the vibration frequency induced by the fault is very high, which will cause the sensor to resonate, but it is beyond the effective frequency range of the accelerometer, so characteristic frequency analysis cannot be performed.

[0005] The second stage is the early stage of the fault. Because the fault power is small, the sideband cannot be seen from the spectrum. It can only be seen from the spectrum when the fault develops to a certain extent. At this time, the characteristic frequency cannot play a role.

[0006] The third stage: In this stage, the fault has developed to a considerable extent, and the sideband can be seen from the spectrum, which is the basis for characteristic frequency analysis.

[0007] The final stage: In this stage, the fault has progressed to a severe level, the characteristic frequency has disappeared, and the characteristic frequency analysis method has completely failed. At this point, judgment is basically based on the national standard GB / T 35854.

[0008] The second aspect is that existing methods rely heavily on experienced analysis engineers, which cannot achieve real-time, automated alarms and diagnostics, resulting in poor timeliness.

[0009] The third aspect is that the analysis results are greatly influenced by the subjective experience of the personnel, making it difficult to guarantee consistency. Existing technologies also use methods such as gradient, skewness, and kurtosis to identify impact waveforms from time-domain signals. However, its main drawback is that in the complex vibration signals of the unit, the vast majority of signals that are not impact waveforms are identified as impact waveforms, especially in the early stages of a fault, when impact waveforms cannot be identified at all.

[0010] Therefore, how to provide an AI diagnostic method that can capture early faults in the transmission chain of wind turbine generators and perform automatic and reliable diagnostic analysis is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0011] In view of the above problems, the present invention is proposed to provide an AI-based automatic real-time analysis and diagnosis method for early faults in the drivetrain of wind turbine generator sets, which overcomes or at least partially solves the above problems.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an AI-based automatic real-time analysis and diagnosis method for early-stage faults in the drivetrain of wind turbine generator sets, comprising the following steps: S1: Collect the original vibration time-domain signal of the wind turbine generator drive chain; perform multi-band filtering on the original vibration time-domain signal to obtain multiple sets of vibration time-domain signals corresponding to different frequency bands; for each set of vibration time-domain signals, extract multiple time segments according to multiple preset durations. S2: Convert the vibration time-domain signal corresponding to each time segment into image data and input it into the pre-trained impact waveform recognition AI model. The impact waveform recognition AI model will identify and output the position and number of impact waveforms in each time segment image. S3: Perform time-domain feature analysis on the vibration signals corresponding to each time segment in S1, extract candidate impact events that meet the preset physical constraints, and evaluate the periodicity parameters of the candidate impact events. Candidate impact events that reach the periodicity parameter threshold are judged as suspected fault impacts. S4: The recognition results based on the impact waveform recognition AI model are cross-validated with the suspected fault impact. If the two meet the preset consistency conditions in terms of impact location and quantity, it is determined that there is a suspected fault impact in this time segment. S5: Based on the judgment results of multiple time segments in a single sampling, and the judgment results of multiple consecutive samplings at the same measuring point, a comprehensive diagnostic conclusion for early faults in the transmission chain is output.

[0013] Preferably, S1 includes: performing bandpass filtering on the original vibration time-domain signal in multiple different target frequency ranges with multiple center frequencies having overlapping steps, so as to extract the filtered signal covering different frequency bands.

[0014] Preferably, step S1 further includes signal preprocessing to remove abnormal acquisition signals before performing multi-band filtering on the original vibration time-domain signal: The original vibration time-domain signal is input into the abnormal acquisition signal elimination AI model to identify abnormal acquisition signals of the original vibration time-domain signal and discard the abnormal acquisition signals.

[0015] Preferably, the preset duration in S1 is configured as: a preset multiple of the theoretical impact period caused by a single fault source when the operating speed of the fan exceeds a specified percentage multiple of the rated speed.

[0016] Preferably, step S2 further includes the following step: converting the time-domain vibration signal of the time segment into a grayscale image, with signals of the same filtering parameter and the same length of time segment forming a group.

[0017] Preferably, the impact waveform recognition AI model in S2 is a deep learning model based on image target localization and recognition. The positive samples of vibration signal images used for pre-training are impact waveform regions with impact waveform labels. The impact waveform regions include: rising edge waveform regions that meet steepness requirements and waveform regions with oscillation decay characteristics after the rising edge.

[0018] Preferably, the impact waveform recognition AI model is any deep learning model used to achieve target localization and recognition in an image.

[0019] Preferably, the step of performing time-domain feature analysis on the vibration signals corresponding to each time segment in S3 includes: S31: Vertically project the time segment image to extract the contour boundary of the signal and obtain a time domain contour image; S32: From the time-domain contour image, candidate peaks are selected based on the preset theoretical minimum time interval and the first amplitude ratio threshold. The theoretical minimum time interval is determined based on the physical dimensions of the transmission chain components and the real-time speed of the fan. S33: Calculate the gradient of the temporal contour image corresponding to the time segment, locate the gradient extremum point, and associate and match the gradient extremum point with the position of the candidate peak to filter out the gradient peak that simultaneously meets the time interval requirement and the gradient requirement. S34: Verify in sequence whether the waveform of the gradient peak has an oscillation decay pattern, and filter out gradient peaks with signal amplitude lower than the second amplitude ratio threshold to obtain the final filtered gradient peaks. S35: Calculate the coefficient of variation (CV) of the time interval between adjacent peaks based on the final screening gradient peak. If the coefficient of variation (CV) is less than or equal to a preset CV threshold, then the vibration signal corresponding to the final screening gradient peak is determined to be a suspected fault impact.

[0020] Preferably, before step S32, the method further includes: filtering out preliminary peaks from the temporal contour image based on a preset minimum pixel interval and amplitude ratio threshold; determining whether the preliminary peaks exceed a physical upper limit threshold, wherein the physical upper limit threshold is determined based on the physical dimensions of the transmission chain components and the real-time rotational speed of the fan; if the number of preliminary peaks does not exceed the physical upper limit threshold, then proceeding to step S32.

[0021] Preferably, S35 further includes: calculating the time interval between adjacent peaks based on the final screening gradient peak to determine the period of the final screening gradient peak, searching for the candidate peak located in the period to update it as the final screening gradient peak, and recalculating the coefficient of variation CV based on the set of the final screening gradient peaks.

[0022] Preferably, step S5 includes the following steps: If the proportion of time segments of the same length under the vibration time domain signal corresponding to the same frequency band that are determined to contain suspected fault impacts exceeds the first set proportion threshold, then the vibration time domain signal is marked as a suspected fault; if the number of vibration time domain signals sampled each time within a continuous period at the same measuring point that are marked as suspected faults exceeds the second set proportion threshold, then it is determined to be a real fault and a maintenance alarm is triggered.

[0023] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: This invention directly identifies the impact waveform from the time-domain signal. Therefore, it can provide early warning in the first and second stages of fault development and extend stable and effective diagnostic capabilities to the third stage of fault development, achieving advanced and reliable monitoring of the entire life cycle of transmission chain faults.

[0024] This invention automatically identifies impacts using a deep learning AI model, ensuring high sensitivity to subtle anomalies. Subsequently, through multi-layered verification based on physical laws and temporal characteristics, it accurately distinguishes between fault impacts and non-fault disturbances, solving the problem of high false alarm rates in traditional time-domain statistical methods. The entire process requires no human intervention, achieving real-time, intelligent diagnosis and improving the timeliness and consistency of operations and maintenance.

[0025] The signal of this invention undergoes multi-band, multi-duration slice preprocessing, which highlights the impact characteristics within the optimal analysis window. The diagnostic conclusion, based on the judgment logic from time segments to continuous duration, greatly improves the reliability of alarms and strongly supports predictive maintenance decisions. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1 This is a flowchart of the AI-based automatic real-time analysis and diagnosis method for early faults in the wind turbine generator drivetrain according to the present invention. Figure 2 This is a schematic diagram illustrating the use of an abnormal acquisition signal removal AI model to remove common abnormal signals during signal preprocessing, as provided in an embodiment of the present invention. Figure 3 Time-domain waveform diagram of early gear failure provided in an embodiment of the present invention; Figure 4 A spectrum diagram of the time-domain waveform of an early gear failure provided in an embodiment of the present invention; Figure 5 The early gear fault time-domain waveform diagram provided in the embodiments of the present invention is a filtered early fault impact waveform. Figure 6 The early fault impulse waveform automatically identified using this AI diagnostic method is provided in the embodiments of the present invention; Figure 7 This is an early fault impulse waveform verified using the time-domain characteristics of the AI ​​diagnostic method provided in an embodiment of the present invention. Figure 8 The time-domain waveform diagram of the third-stage gear fault provided in the embodiment of the present invention; Figure 9 The spectrum diagram of the time-domain waveform of the third-stage gear fault provided in the embodiment of the present invention; Figure 10 The AI ​​diagnostic method provided in this embodiment of the invention automatically identifies the third-stage fault impact waveform. Figure 11 The time-domain waveform diagram of the fourth stage gear fault provided in the embodiment of the present invention; Figure 12 A spectrum diagram of the fourth stage gear failure provided in an embodiment of the present invention; Figure 13 This is a non-fault-related impact waveform diagram caused by unit yaw provided in an embodiment of the present invention; Figure 14 The waveform diagram provided in this embodiment of the invention is automatically excluded from being modulated but not impulse using this AI diagnostic method; Figure 15 This is a schematic diagram of the first working condition for using time-domain features to assist in correcting periodic assessments, provided in an embodiment of the present invention. Figure 16 This is a schematic diagram of the second working condition provided by an embodiment of the present invention, which utilizes time-domain features to assist in correcting periodic assessments. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] This invention discloses an AI-based automatic real-time analysis and diagnosis method for early-stage faults in the transmission chain of wind turbine generators, based on multi-scale considerations of the data center, server rack, and server. Figure 1 As shown, it includes the following steps: S1: Collect the original vibration time-domain signal of the wind turbine generator drive chain; perform multi-band filtering on the original vibration time-domain signal to obtain multiple sets of vibration time-domain signals corresponding to different frequency bands; for each set of vibration time-domain signal, extract multiple time segments according to various preset durations. S2: Convert the vibration time-domain signal corresponding to each time segment into image data and input it into the pre-trained impact waveform recognition AI model. The AI ​​model will identify and output the position and number of impact waveforms in each time segment image. S3: Perform time-domain feature analysis on the vibration signals corresponding to each time segment in S1, extract candidate impact events that meet the preset physical constraints, and evaluate the periodicity parameters of the candidate impact events. Candidate impact events that reach the periodicity parameter threshold are judged as suspected fault impacts. S4: Cross-validate the identification results based on the AI ​​model with the suspected fault impacts. If the two meet the preset consistency conditions in terms of impact location and quantity, it is determined that there is a suspected fault impact in this time segment. S5: Based on the judgment results of multiple time segments in a single sampling, and the judgment results of multiple consecutive samplings at the same measuring point, a comprehensive diagnostic conclusion for early faults in the transmission chain is output.

[0030] It should be noted that a fault triggers an impact waveform in the signal, but the impact waveform in the signal is not necessarily caused by the fault. The impact caused by the fault is not only periodic, but the number of impacts per second also conforms to the physical dimensions and operating speed of the wind turbine generator set's transmission chain. Only such an impact waveform is a fault-caused impact. When a fault occurs, the impact waveform will persist. This periodicity will persist not only throughout the entire process of the same sampled signal, but also in different samples. In contrast, non-fault impacts caused by yaw, unstable winds, or changing operating conditions will not persist.

[0031] At the same time, the impact waveforms caused by each impact in a real fault are not all typical or equally regular. The AI ​​model may miss such impacts. When the AI ​​model identifies impact waveforms, there may also be cases where an impact waveform is identified repeatedly. As a result, the periodicity of the impact is judged by the number and position of the impact waveforms identified by the AI ​​model, which is worse than the periodicity of the actual impact. This also requires auxiliary correction from the time domain characteristics.

[0032] This invention utilizes time-domain features to verify whether the impact identified by the AI ​​model is caused by a real fault, thus excluding non-fault impacts.

[0033] In one embodiment, S1 includes: performing bandpass filtering on the original vibration time-domain signal at multiple center frequencies with overlapping steps in multiple different target frequency ranges to extract filtered signals covering different frequency bands.

[0034] In one embodiment, S1 further includes signal preprocessing to remove abnormal acquisition signals before multi-band filtering of the original vibration time-domain signal: The original vibration time-domain signal is input into the abnormal acquisition signal elimination AI model to identify abnormal acquisition signals of the original vibration time-domain signal and discard the abnormal acquisition signals.

[0035] It should be noted that the abnormal acquisition signal rejection AI model is independent of the impact waveform recognition AI model. In this embodiment, when training the abnormal acquisition signal rejection AI model, abnormal acquisition signals are selected as training samples to train a separate AI model. This model is only used to perform abnormal acquisition signal rejection in the signal preprocessing stage and cannot be used to perform step S2. When using this abnormal acquisition signal rejection AI model to determine whether the corresponding time-domain signal is an abnormal signal, if it is, step S2 does not perform inference and diagnosis on that time-domain signal.

[0036] like Figure 2 As shown in (1), (2), and (3), AI models are used to remove three common abnormal acquisition signals during signal preprocessing.

[0037] In one embodiment, the preset duration in S1 is configured as: a preset multiple of the theoretical impact period caused by a single fault source when the operating speed of the fan exceeds a specified percentage multiple of the rated speed.

[0038] In practical implementation, the setting of multiple filter frequency upper and lower limits should be balanced by considering the width of the frequency range [upper limit - lower limit]. It should not be too wide or too narrow. It should cover the frequency range excited by the impact to facilitate the display of the impact waveform. At the same time, it should also take into account the increased computational burden caused by an excessively narrow frequency range. For example, for a vibration signal with a sampling frequency of 25600Hz, the following settings can be used simultaneously: (60,2000, 0.8), (100,2000, 0.6), (200,3000, 0.8), (300,5000, 0.6), (500,NyquistFs, 0.8).

[0039] Let's take (60, 2000, 0.8) as an example to illustrate its meaning: The filter bandwidth is 60Hz to filter out impulse waveforms with frequencies below 60Hz. This bandwidth is applied within the 0-2000Hz range. The lowcut increments by 60 * 0.8Hz, meaning a 20% overlap in frequency range with each increment. Impulse waveforms with frequencies below 60Hz generally do not exceed 2000Hz; higher frequencies are not selected to reduce computational burden. The 20% overlap is to prevent omissions or to more completely display impulse waveforms at frequency crossover points.

[0040] In practical implementation, the various time segment length settings should cover the upper limit of the number of impact waveforms caused by transmission chain faults when the wind turbine is operating above 90-95% of its rated speed. The time segment length should be long enough to ensure that, even considering potential omissions by the AI ​​model, the segment still contains a sufficient number of impacts from the same fault source when the wind turbine is stably operating at high speeds (>95% of rated speed). This supports reliable subsequent assessment of the impact periodicity; a 20% margin is generally sufficient. The segment should also contain at least five impact waveforms. For example, for a 6-second vibration signal, the time segment length can be set to six different values: 1s, 2s, 3s, 4s, 5s, and 6s.

[0041] Each time segment of the signal is captured, it overlaps with the previous one by 15%. This overlap rate is set to ensure that no impact across time segments is missed.

[0042] Ideally, each wind turbine and each measuring point should have at least 10 sampled data points analyzed by AI per day.

[0043] In one embodiment, S2 further includes the following steps: converting the time-domain vibration signal of the time segment into a grayscale image, with signals of the same filtering parameter and the same length of time segment being grouped together.

[0044] In this embodiment, the time-domain signal acquired by the CMS system is converted into grayscale graphic files such as JPG, PNG, and BMP, and the impact waveform part in the graphic file is marked by a rectangular box or a segmented polygonal box for classification and recognition.

[0045] The graphic file can be saved as a JPG, PNG, BMP or other format and input into the AI ​​model for impact waveform recognition and counting. It can also be directly input into the AI ​​model as a memory graphic file for impact waveform recognition and counting.

[0046] The selection of the width and height pixel count should be conducive to the display of the impact waveform characteristics. For a given time slice signal length, pixels that are too wide or too narrow will distort the impact waveform, and the same applies to the height. The selection of width and height pixels should also be considered in conjunction with computing power configuration factors. In this embodiment, (320~640)*(160~320) is chosen as a better value during model training and inference, which is conducive to the display of the impact waveform without putting too much computing power pressure on inference.

[0047] In one embodiment, the impact waveform recognition AI model in S2 is a deep learning model based on image target localization and recognition. The positive samples of vibration signal images used for pre-training are impact waveform regions with impact waveform labels. The impact waveform regions include: rising edge waveform regions that meet steepness requirements and waveform regions with oscillation decay characteristics after the rising edge.

[0048] In practice, the impact waveform of the actual fault vibration signal that occurred in the actual operating unit's drive chain is selected as the sample data for model training.

[0049] The key to selecting training samples is to identify waveforms with steep rising edges and subsequent oscillation decay in the labeled signal as impulse waveforms. At the same time, identify waveforms that are spike-modulated but have non-steep rising edges and / or no oscillation decay characteristics. This auxiliary method can identify impulse waveforms more accurately. Another key to selecting training samples is requiring the fan speed to be above 90-95% of its rated speed, excluding waveforms modulated during fan startup that are not actual impacts. The 90-95% parameter is particularly important here, as it eliminates the modulation effect of varying operating conditions on the vibration signal, ensuring that the identified disturbances are fault impacts rather than disturbances caused by changes in operating conditions.

[0050] In this embodiment, the impact waveform recognition AI model can also be incrementally learned or retrained using confirmed fault data in subsequent operations to continuously optimize model performance.

[0051] In one embodiment, the shock waveform recognition AI model employs any deep learning model capable of locating and recognizing targets in an image, including but not limited to the YOLO series, R-CNN series, or SSD model, and is trained through transfer learning to recognize shock waveforms in vibration signals.

[0052] In one embodiment, the step of performing time-domain feature analysis on the vibration signal corresponding to each time segment in S3 includes: S31: Vertically project the time segment image to extract the contour boundary of the signal and obtain a time domain contour image; S32: From the temporal contour image, candidate peaks are selected based on the preset theoretical minimum time interval and the first amplitude ratio threshold. The theoretical minimum time interval is determined based on the physical dimensions of the transmission chain components and the real-time speed of the fan. S33: Calculate the gradient of the time domain contour image corresponding to the time segment, locate the gradient extremum points, and associate and match the gradient extremum points with the positions of candidate peaks to filter out gradient peaks that simultaneously meet the time interval requirements and gradient requirements. S34: Verify in turn whether the waveform of the gradient spike has an oscillating decay pattern, and filter out gradient spikes whose signal amplitude is lower than the second amplitude ratio threshold to obtain the final filtered gradient spikes. S35: Calculate the coefficient of variation (CV) of the time interval between adjacent peaks based on the final screening gradient peak. If the coefficient of variation (CV) is less than or equal to the preset CV threshold, the vibration signal corresponding to the final screening gradient peak is determined to be a suspected fault impact.

[0053] In one embodiment, before S32, the method further includes: filtering out preliminary peaks from the temporal contour image based on a preset minimum pixel interval and amplitude ratio threshold; determining whether the preliminary peaks exceed a physical upper limit threshold, wherein the physical upper limit threshold is determined based on the physical dimensions of the transmission chain components and the real-time rotational speed of the fan; if the number of preliminary peaks does not exceed the physical upper limit threshold, then proceeding to S32.

[0054] In one embodiment, S35 further includes: determining the period of the final selected gradient peak by calculating the time interval between adjacent peaks based on the final selected gradient peak, searching for candidate peaks located in the period to update them as the final selected gradient peaks, and recalculating the coefficient of variation (CV) based on the set of the final selected gradient peaks.

[0055] The specific execution process of S3 is given below: S310: Vertically project the grayscale image of each time segment signal to obtain the waveform outline boundary.

[0056] S320: Based on waveform contour boundaries, remove detail noise and highlight key features.

[0057] S330: Performs signal smoothing processing on the denoised grayscale image, preserving impact features while removing high-frequency noise.

[0058] S340: Based on the minimum distance of 5 pixels and different height ratio thresholds, identify the number and location of spikes. The height ratio in this step is the percentage of a single spike height in the time-domain signal waveform relative to the total number of spike heights. The height ratio threshold can be set to [0.0, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], where 0.0 corresponds to all spikes, 0.4 corresponds to spikes with a height ratio exceeding 40%, and 0.5 corresponds to spikes with a height ratio exceeding 50%. Their numbers significantly exceed the number of impacts that could be caused by a fault in the unit's drivetrain within that time segment, thus eliminating the possibility of fault identification.

[0059] S350: Based on the physical dimensions of the unit's drive train and the minimum distance set by the rated operating speed of the unit, as well as the threshold values ​​for different height ratios [0.0, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], find the number and location of the spikes.

[0060] S360: Find the gradient peak based on the minimum distance and 90th percentile of height set in S350, and the significance calculated based on np.std(gradient)*2.

[0061] S370: Find the peaks near the gradient peak, i.e., gradient spikes.

[0062] S380: Determine the impact damping oscillation characteristics of gradient peaks and exclude peaks that are gradient peaks but do not meet the impact damping oscillation characteristics.

[0063] S390: Filters out gradient spikes with a height ratio less than 0.5.

[0064] S3100: The number of impacts that are significantly less than the number that could be caused by a fault in the unit's drivetrain within that time segment are also ruled out as possible faults.

[0065] S3110: Evaluate the periodicity of the gradient spikes obtained in S390, denoted by CV. Multiple CV thresholds can be set: 0.05, 0.1, 0.15, 0.2, 0.25, 0.3. Here, the spikes found in S350 with different height ratio thresholds [0.0, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] can be used to supplement the evaluation of the periodicity of the gradient spikes obtained in S390. For example, if there are 10 periodic impacts per second, and the height ratio of the 3rd impact only reaches 0.4, it is not counted in the gradient spikes, but should be considered as a gradient spike, thus ensuring excellent periodicity of the waveform.

[0066] S3120: If the above characteristics are met and the CV reaches the set requirements, it is judged as a suspected fault impact. The CV balance point obtained by historical data verification is 0.1~0.15.

[0067] In one embodiment, when step S4 is specifically executed, if the location of the impact identified by the AI ​​model matches the location of the suspected fault impact determined according to the above time-domain features, and the error between the number of impacts and the number of suspected fault impacts determined according to the above time-domain features is less than 2, then the impact in the signal of that time segment is determined to be a fault impact.

[0068] In one embodiment, S5 includes the following steps: If the proportion of time segments of the same length in the vibration time domain signal corresponding to the same frequency band that are identified as containing suspected fault impacts exceeds the first set proportion threshold, the vibration time domain signal is marked as a suspected fault; if the number of vibration time domain signals sampled in each continuous period at the same measuring point that are marked as suspected faults exceeds the second set proportion threshold, it is determined to be a real fault and a maintenance alarm is triggered.

[0069] In practice, if 85% of the time segments of the same length extracted by the same filter parameter in a single sampled signal are suspected fault impacts, then the sampled signal is judged as a suspected fault. For each wind turbine and each measuring point, the AI ​​inference should be no less than 10 data points per day (this parameter can be set), and more than 80% of them should be verified as suspected faults as described above. It is recommended that the maintenance team pay close attention to this.

[0070] If more than 80% of the sampled signals from the same measuring point over several consecutive days (e.g., 3-5 days) are suspected faults according to AI inference, then it is determined to be a real fault, and it is recommended that wind turbine maintenance personnel go to the wind turbine site for inspection.

[0071] like Figures 3-4 The image shows an early stage of gear failure, in which... Figure 3 This is a time-domain waveform diagram of an early gear failure. Figure 4 yes Figure 3 The time-domain waveform diagram shown corresponds to a spectrum diagram. It can be seen from the diagram that in the early stages of the fault, no characteristic frequency (sideband) has yet appeared in the spectrum. At this time, the fault power is still very small and cannot be represented as a characteristic frequency in frequency domain analysis. Furthermore, its vibration RMS has not reached the warning threshold set by GB / T 35854, making it impossible to analyze using existing characteristic frequency analysis methods or GB / T 35854. However, this early gear fault can be easily and automatically identified using the method of this invention, such as... Figures 5-7 As shown, where, Figure 5 yes Figure 3 The time-domain waveform shown is the early fault impulse waveform after filtering. Figure 6It is automatically identified using the AI ​​model of the method in the embodiments of the present invention. Figure 3 The early fault impact waveform shown in the time-domain waveform diagram is... Figure 7 This is an early fault impulse waveform verified using the time-domain characteristics of the method in this embodiment of the invention, used in conjunction with... Figure 6 The early fault impulse waveforms shown are cross-validated.

[0072] like Figures 8-9 The diagram shows the third stage of a gear failure (specifically, a broken tooth failure). Figure 8 This is the time-domain waveform diagram of the third stage gear failure. Figure 9 yes Figure 8 The time-domain waveform diagram shown corresponds to a spectrum diagram. At this point, the fault power is already very high, and obvious characteristic frequencies (sidebands) appear in the spectrum. Analysts can use existing frequency characteristics to analyze the fault at this stage. However, the method of this invention can identify the fault more quickly and automatically. The specific identification results are as follows: Figure 10 As shown in the figure, the identified [items] are [the following]. Figure 8 The fault impact waveform of the segment signal shown in the time-domain waveform diagram provides a fast and feasible time-domain diagnostic scheme for gear fault diagnosis in the third stage.

[0073] like Figures 11-12 As shown, this represents the fourth stage of gear failure development, in which... Figure 11 This is the time-domain waveform diagram of the fourth stage gear failure. Figure 12 yes Figure 11 The time-domain waveform diagram shown corresponds to the spectrum diagram. At this point, the fault is already very serious, but the characteristic frequencies have completely disappeared in the spectrum. At this stage, identification can only be done by relying on GB / T 35854.

[0074] like Figure 13 The image shows a non-fault-related yaw impact. Using the method of this invention, such non-fault-related impacts can be automatically eliminated. The method of this invention eliminates waveforms that are modulated pulses but not impacts, i.e., waveforms that do not have the characteristics of a steep rising edge and oscillating decay at the falling edge. Specific identification results are shown below. Figure 14 As shown.

[0075] like Figure 15 As shown in (1) and (2) in the figure, this is an example of using time-domain features to help correct the true periodicity of the impact waveform that AI failed to identify.

[0076] In Figure (1), the fifth impact waveform was not identified, resulting in poor periodicity of the impact waveform identified by AI. However, it can be clearly determined from the time domain peak that there is a peak at this position. As shown in Figure (2), time domain correction was performed and it was added. After that, the actual impact waveform had good periodicity.

[0077] like Figure 16 As shown in (1), (2), and (3) in the figure, this is Example 2, which uses time-domain features to help correct the true periodicity of the impact waveform repeatedly identified by AI.

[0078] In Figure (1), the fourth impact waveform was repeatedly identified, resulting in poor periodicity of the impact waveform identified by AI. However, it can be clearly determined from the time domain peaks that there is only one peak at this position. As shown in Figure (2), time domain correction was performed to merge the two peaks. The actual impact waveform has good periodicity, as shown in Figure (3).

[0079] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0080] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. 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 the invention. Therefore, the invention 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.

Claims

1. An AI automatic real-time analysis and diagnosis method for early faults of a wind turbine drive train, characterized in that, Includes the following steps: S1: Collect the original vibration time-domain signal of the wind turbine generator drive chain; perform multi-band filtering on the original vibration time-domain signal to obtain multiple sets of vibration time-domain signals corresponding to different frequency bands; for each set of vibration time-domain signals, extract multiple time segments according to multiple preset durations. S2: Convert the vibration time-domain signal corresponding to each time segment into image data and input it into the pre-trained shock waveform recognition AI model. The shock waveform recognition AI model will identify and output the position and number of shock waveforms in each time segment image. S3: Perform time-domain feature analysis on the vibration signals corresponding to each time segment in S1, extract candidate impact events that meet preset physical constraints, and evaluate the periodicity parameters of the candidate impact events. Candidate impact events that reach the periodicity parameter threshold are judged as suspected fault impacts. The steps of performing time-domain feature analysis on the vibration signals corresponding to each time segment include: S31: Vertically project the time segment image to extract the contour boundary of the signal and obtain a time domain contour image; S32: From the time-domain contour image, candidate peaks are selected based on the preset theoretical minimum time interval and the first amplitude ratio threshold. The theoretical minimum time interval is determined based on the physical dimensions of the transmission chain components and the real-time speed of the fan. S33: Calculate the gradient of the temporal contour image corresponding to the time segment, locate the gradient extremum point, and associate and match the gradient extremum point with the position of the candidate peak to filter out the gradient peak that simultaneously meets the time interval requirement and the gradient requirement. S34: Verify in sequence whether the waveform of the gradient peak has an oscillation decay pattern, and filter out gradient peaks with signal amplitude lower than the second amplitude ratio threshold to obtain the final filtered gradient peaks. S35: Calculate the coefficient of variation (CV) of the time interval between adjacent peaks based on the final screening gradient peak. If the coefficient of variation (CV) is less than or equal to a preset CV threshold, then the vibration signal corresponding to the final screening gradient peak is determined to be a suspected fault impact. S4: The recognition results based on the impact waveform recognition AI model are cross-validated with the suspected fault impact. If the two meet the preset consistency conditions in terms of impact location and quantity, it is determined that there is a suspected fault impact in this time segment. S5: Based on the judgment results of multiple time segments in a single sampling, and the judgment results of multiple consecutive samplings at the same measuring point, a comprehensive diagnostic conclusion for early faults in the transmission chain is output.

2. The AI automatic real-time analysis and diagnosis method for early fault of a wind turbine drivetrain according to claim 1, characterized in that, S1 includes: performing bandpass filtering on the original vibration time-domain signal in multiple different target frequency ranges with multiple center frequencies having overlapping steps, so as to extract the filtered signal covering different frequency bands.

3. The AI automatic real-time analysis diagnostic method for early fault of a wind turbine drivetrain according to claim 1, characterized in that, S1 further includes signal preprocessing, which involves removing abnormal acquisition signals before multi-band filtering of the original vibration time-domain signal: The original vibration time-domain signal is input into the abnormal acquisition signal elimination AI model to identify abnormal acquisition signals of the original vibration time-domain signal and discard the abnormal acquisition signals.

4. The AI automatic real-time analysis diagnostic method for early fault of a wind turbine drivetrain according to claim 1, characterized in that, S2 further includes the following steps: converting the time-domain vibration signal of the time segment into a grayscale image, and grouping signals of the same filtering parameter and the same length of time segment together.

5. The AI automatic real-time analysis diagnostic method for early fault of the wind turbine drivetrain according to claim 1, characterized in that, The impact waveform recognition AI model in S2 is a deep learning model based on image target localization and recognition. The positive samples of vibration signal images used for pre-training are impact waveform regions with impact waveform labels. The impact waveform regions include: rising edge waveform regions that meet steepness requirements and waveform regions with oscillation decay characteristics after the rising edge.

6. The AI automatic real-time analysis diagnostic method for early fault of a wind turbine drivetrain chain according to claim 1, characterized in that, The shock waveform recognition AI model can be any deep learning model used to locate and recognize targets in images.

7. The AI automatic real-time analysis diagnostic method for early fault of a wind turbine drivetrain chain according to claim 1, characterized in that, Before step S32, the method further includes: filtering out preliminary peaks from the temporal contour image based on a preset minimum pixel interval and amplitude ratio threshold; determining whether the preliminary peaks exceed a physical upper limit threshold, wherein the physical upper limit threshold is determined based on the physical dimensions of the transmission chain components and the real-time rotational speed of the fan; if the number of preliminary peaks does not exceed the physical upper limit threshold, then proceeding to step S32.

8. The AI automatic real-time analysis diagnostic method for early fault of a wind turbine drivetrain chain according to claim 1, characterized in that, S35 further includes: calculating the time interval between adjacent peaks based on the final screening gradient peak to determine the period of the final screening gradient peak, searching for the candidate peak located in the period to update it as the final screening gradient peak, and recalculating the coefficient of variation CV based on the set of the final screening gradient peaks.

9. The AI automatic real-time analysis diagnostic method for early fault of a wind turbine drivetrain chain according to claim 1, characterized in that, S5 includes the following steps: If the proportion of time segments of the same length under the vibration time domain signal corresponding to the same frequency band that are determined to contain suspected fault impacts exceeds the first set proportion threshold, then the vibration time domain signal is marked as a suspected fault; if the number of vibration time domain signals sampled in each continuous period at the same measuring point that are marked as suspected faults exceeds the second set proportion threshold, then it is determined to be a real fault and a maintenance alarm is triggered.

Images