Contrastive learning based prognostics and health management method

Abstract

The present application relates to the technical field of fault analysis, and more particularly to a fault prediction and health management method based on contrast learning, comprising: constructing a wind turbine health operation dataset to be input into a space-time contrast encoder to generate a wind turbine health benchmark feature library; comparing and analyzing the health deviation degree of the wind turbine according to the wind turbine operation data, wind turbine location meteorological data and wind turbine health benchmark feature library; analyzing the fault state of the wind turbine according to the health deviation degree analysis result of the wind turbine, wind turbine oil detection data and wind turbine blade inspection image data; importing the health deviation degree comparison analysis result and fault state analysis result of the wind turbine into a wind turbine dynamic scheduling strategy analysis model to generate a wind turbine dynamic scheduling strategy, and dynamically scheduling the operation of the wind turbine in the wind turbine generator set; the load of the high-risk wind turbine can be adaptively adjusted to reduce the fault risk, and the utilization rate of the low-risk wind turbine is improved to meet the total demand, thereby improving the equipment life and grid stability.

Description

Technical Field

[0001] This invention relates to the field of fault analysis technology, and in particular to a fault prediction and health management method based on comparative learning. Background Technology

[0002] With the rapid global energy structure shifting towards cleaner and lower-carbon energy, wind power, as a crucial pillar of renewable energy, has become a key pathway to ensuring energy security and achieving "dual-carbon" goals through large-scale deployment and efficient operation and maintenance. However, wind turbines operate in complex and ever-changing natural environments, constantly subjected to harsh conditions such as wind speed fluctuations, extreme temperatures and humidity, and salt spray corrosion, leading to frequent problems such as wear and tear on mechanical components, aging of electrical systems, and damage to blade structures. Traditional wind turbine operation and maintenance methods mainly rely on periodic inspections, threshold alarms, and reactive maintenance, but these methods have significant limitations: First, manual inspections are inefficient and costly, making it difficult to cover the numerous turbines in large wind farms; second, threshold alarms, based on static rules with a single parameter, cannot capture early fault characteristics under multi-parameter coupling, easily resulting in missed or false alarms; finally, reactive maintenance is a passive response to faults, often causing unplanned downtime, leading to power generation losses and increased maintenance costs. Existing wind turbine generator fault analysis and maintenance technologies typically construct intelligent fault analysis models by integrating multi-dimensional data, enabling a shift from "post-fault maintenance" to "predictive maintenance" in traditional methods.

[0003] However, existing technologies, when analyzing the dynamic coupling relationship between wind turbine operating status and complex environmental factors, do not consider adaptive modeling of wind turbine health baseline characteristics under multiple operating conditions. This results in insufficient discriminative power in feature extraction, making it difficult to accurately identify abnormal states under the influence of external conditions such as wind speed zoning and temperature changes. Furthermore, existing methods often use power generation efficiency as the sole optimization objective, neglecting the importance of equipment health to wind turbine operating conditions and load distribution, and failing to consider the dynamic balance between health deviation and load rate. This leads to the continuous accumulation of wind turbine failure risks under high load operation, making it difficult to achieve the optimal trade-off between wind turbine power generation revenue and equipment lifespan.

[0004] To address these issues, this application proposes a fault prediction and health management method based on contrastive learning. Summary of the Invention

[0005] The purpose of this invention is to provide a fault prediction and health management method based on contrastive learning. This method involves constructing a wind turbine health benchmark feature library; analyzing the health deviation and fault status of wind turbines using multidimensional data and the wind turbine health benchmark feature library; generating a dynamic scheduling strategy for wind turbines based on the analysis results; and dynamically scheduling the operation of wind turbines in the wind turbine generator set. This method can adaptively adjust the load of high-risk wind turbines to reduce the risk of failure, and improve the utilization rate of low-risk wind turbines to meet total demand, thereby improving equipment lifespan and grid stability.

[0006] This invention is implemented as follows:

[0007] In a first aspect, the present invention provides a fault prediction and health management method based on contrastive learning, comprising the following steps:

[0008] S1. Acquire wind turbine operation data, wind turbine location meteorological data, and historical health label data of the wind turbine generator set; at the same time, acquire wind turbine oil detection data and wind turbine blade inspection image data;

[0009] S2. Based on historical health tag data and wind turbine operation data, construct a wind turbine health operation dataset, input the wind turbine health operation dataset into the spatiotemporal comparison encoder, and generate a wind turbine health benchmark feature library.

[0010] S3. Import the wind turbine operation data, wind turbine location meteorological data, and wind turbine health benchmark feature library into the health deviation comparison analysis model, and compare and analyze the health deviation of the wind turbine.

[0011] S4. Based on the health deviation analysis results of the wind turbine, the oil detection data of the wind turbine, and the inspection image data of the wind turbine blades, a wind turbine fault state analysis model is constructed to analyze the fault state of the wind turbine.

[0012] S5. Construct a dynamic scheduling strategy analysis model for wind turbines. Import the health deviation comparison analysis results and fault status analysis results of wind turbines into the dynamic scheduling strategy analysis model to generate a dynamic scheduling strategy for wind turbines and dynamically schedule the operation of wind turbines in the wind turbine generator set.

[0013] Preferably, in step S2, a wind turbine health operation dataset is constructed based on historical health tag data and wind turbine operation data, specifically including:

[0014] S21. Under the same monitoring time window, monitor the wind turbine operation data in the wind turbine generator set, and monitor the wind speed at the location of the wind turbine in real time to obtain the wind speed range at the location of the wind turbine.

[0015] S22. Divide the wind speed range at the location of the wind turbine into multiple wind speed sub-ranges at equal intervals; randomly select historical wind turbine operation data located in a certain wind speed sub-range from the historical health label data of the wind turbine as wind turbine anchor point samples;

[0016] S23. From the historical health label data of wind turbines, randomly select multiple historical wind turbine operation data corresponding to the same wind speed sub-interval as the wind turbine anchor point sample when there are no fault records and the wind turbine is operating normally as positive samples of wind turbine health.

[0017] S24. From the historical health label data of wind turbines, randomly select multiple historical wind turbine operation data corresponding to the same wind speed sub-interval as the wind turbine anchor point sample when there is a fault record or abnormal operation as wind turbine health negative samples.

[0018] S25. The selected wind turbine anchor point samples, positive wind turbine health samples, and negative wind turbine health samples are used as the wind turbine health operation dataset.

[0019] Preferably, in step S2, the wind turbine health operation dataset is input into the spatiotemporal comparison encoder to generate a wind turbine health benchmark feature library, specifically including:

[0020] S26. Import the wind turbine health operation dataset into the spatiotemporal contrast encoder model. With the training objective of maximizing the similarity of samples of the same class and minimizing the similarity of samples of different classes, train the spatiotemporal contrast encoder model until the model loss value converges; output the trained spatiotemporal contrast encoder model.

[0021] S27. Input the historical wind turbine operation data corresponding to the same wind turbine in all wind speed sub-intervals when there are no fault records and the wind turbine is operating normally into the trained spatiotemporal comparison encoder model, and output the wind turbine health feature vector in all wind speed sub-intervals.

[0022] S28. Based on the wind turbine health feature vectors in all wind speed sub-intervals, perform cluster analysis according to the wind speed sub-intervals to generate the wind turbine health benchmark feature vector corresponding to each wind speed sub-interval. Import the wind turbine health benchmark feature vectors corresponding to all wind speed sub-intervals into the database to obtain the wind turbine health benchmark feature library.

[0023] Preferably, step S3 involves a comparative analysis of the health deviation of the wind turbine, specifically including the following steps:

[0024] S31. Extract wind turbine operation data, wind turbine location meteorological data, and wind turbine health benchmark feature library;

[0025] S32. Construct a health deviation comparison analysis model. Import wind turbine operation data, wind turbine location meteorological data, and wind turbine health benchmark feature library into the health deviation comparison analysis model, compare and analyze the health deviation of the wind turbine, and obtain the health deviation analysis results of the wind turbine.

[0026] Preferably, the construction process of the health deviation comparison analysis model in step S32 includes the following specific steps:

[0027] S321. Extract the real-time wind speed from the meteorological data of the wind turbine location; based on the real-time wind speed, extract the wind turbine health benchmark feature vector corresponding to the wind speed sub-interval from the wind turbine health benchmark feature library;

[0028] S322. Obtain the wind turbine operation data at the current moment, input the wind turbine operation data into the spatiotemporal comparison encoder model, and obtain the feature vector of the wind turbine operation data at the current moment;

[0029] S323. Based on the feature vector of the wind turbine operation data at the current moment and the wind turbine health benchmark feature vector of the corresponding wind speed sub-interval, calculate the health deviation of the wind turbine at the current moment.

[0030] The formula for calculating health deviation is:

[0031]

[0032] In the formula, Dh is the health deviation of the wind turbine at the current moment, fb is the wind turbine health benchmark feature vector of the wind speed sub-interval where the real-time wind speed is located, and fc is the feature vector of the wind turbine operation data at the current moment.

[0033] Preferably, step S4 involves analyzing the fault status of the wind turbine, including the following specific steps:

[0034] S41. Obtain the health deviation analysis results of the wind turbine, and at the same time obtain the wind turbine oil detection data and wind turbine blade inspection image data;

[0035] S42. Import the health deviation analysis results of the wind turbine, the oil detection data of the wind turbine, and the inspection image data of the wind turbine blades into the wind turbine fault state analysis model to analyze the fault state of the wind turbine and obtain the fault state analysis results of the wind turbine.

[0036] Preferably, the construction process of the wind turbine fault state analysis model in step S42 includes the following specific steps:

[0037] S421. Based on the fan oil detection data, calculate the abnormal index of wear particle distribution of the fan;

[0038] S422. Calculate the surface anomaly index of the wind turbine blades based on the wind turbine blade inspection image data;

[0039] S423. The health deviation of the fan, the abnormal index of wear particle distribution, and the abnormal index of blade surface are imported into the fan fault state calculation formula to calculate the fault state of the fan.

[0040] The formula for calculating the fault status of a wind turbine is as follows:

[0041]

[0042] In the formula, Bp represents the fault state of the fan at the current moment; when q=1, P1 represents the health deviation of the fan at the current moment; when q=2, P2 represents the abnormal distribution index of wear particles in the fan; and when q=3, P3 represents the abnormal surface index of the fan blades. This is the chain multiplication operator, where q represents any term from 1 to 3.

[0043] Preferably, step S5 involves constructing a dynamic scheduling strategy analysis model for wind turbines, specifically including:

[0044] S51. Obtain the health deviation comparison analysis results and fault status analysis results of the wind turbine obtained from the analysis;

[0045] S52. Import the health deviation comparison analysis results and fault status analysis results of the wind turbines into the wind turbine dynamic scheduling strategy analysis model to generate a wind turbine dynamic scheduling strategy and dynamically schedule the operation of wind turbines in the wind turbine generator set.

[0046] Preferably, the construction process of the wind turbine dynamic scheduling strategy analysis model in step S52 specifically includes:

[0047] S521. Import the health deviation comparison analysis results and fault status analysis results of the wind turbine into the calculation formula of the wind turbine dynamic scheduling decision index, and calculate the dynamic scheduling decision index of the wind turbine at the current moment.

[0048] The formula for calculating the dynamic scheduling decision index of wind turbines is:

[0049]

[0050] In the formula, S is the dynamic scheduling decision index of the wind turbine at the current moment, Lc is the load rate of the wind turbine at the current moment, and Lmax is the rated maximum load rate of the wind turbine.

[0051] S522. Obtain the dynamic scheduling decision index of all wind turbines in the wind turbine generator set at the current moment and substitute it into the wind turbine dynamic scheduling optimization objective function;

[0052] S523. Calculate the gradient of the wind turbine dynamic scheduling optimization objective function using the gradient descent method, iteratively adjust the load rate of the wind turbines along the negative gradient direction, and output the optimal load rate of all wind turbines that makes the wind turbine dynamic scheduling optimization objective function converge after iterative adjustment. Then, redistribute the load of all wind turbines in the wind turbine generator set according to the optimal load rate of all wind turbines.

[0053] In a second aspect, the present invention provides an electronic device, comprising: a processor and a memory, wherein the memory stores a computer program that can be called by the processor, and the processor executes a fault prediction and health management method based on contrastive learning by calling the computer program stored in the memory.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] This invention constructs a wind turbine health operation dataset, which is input into a spatiotemporal comparison encoder to generate a wind turbine health benchmark feature library. Based on wind turbine operation data, wind turbine location meteorological data, and the wind turbine health benchmark feature library, the health deviation of the wind turbines is compared and analyzed. The fault status of the wind turbines is analyzed based on the health deviation analysis results, wind turbine oil level detection data, and wind turbine blade inspection image data. The results of the wind turbine health deviation comparison analysis and fault status analysis are imported into a wind turbine dynamic scheduling strategy analysis model to generate a dynamic scheduling strategy for dynamically scheduling the operation of wind turbines in the wind turbine generator set. This strategy can adaptively adjust the load of high-risk wind turbines to reduce fault risk and improve the utilization rate of low-risk wind turbines to meet total demand, thereby improving equipment lifespan and grid stability. Attached Figure Description

[0056] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0057] Figure 1 This is a schematic diagram of the overall process of the fault prediction and health management method based on contrastive learning of the present invention;

[0058] Figure 2 This is an analysis flowchart of step S3 in the fault prediction and health management method based on contrastive learning of the present invention;

[0059] Figure 3 This is an analysis flowchart of step S4 in the fault prediction and health management method based on contrastive learning of the present invention;

[0060] Figure 4 This is an analysis flowchart of step S5 of the fault prediction and health management method based on contrastive learning of the present invention. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0062] Example 1

[0063] like Figure 1 As shown, this embodiment provides a fault prediction and health management method based on contrastive learning, which specifically includes the following steps:

[0064] S1. Acquire wind turbine operation data, wind turbine location meteorological data, and historical health label data of the wind turbine generator set; at the same time, acquire wind turbine oil detection data and wind turbine blade inspection image data;

[0065] S2. Based on historical health tag data and wind turbine operation data, construct a wind turbine health operation dataset, input the wind turbine health operation dataset into the spatiotemporal comparison encoder, and generate a wind turbine health benchmark feature library.

[0066] S3. Import the wind turbine operation data, wind turbine location meteorological data, and wind turbine health benchmark feature library into the health deviation comparison analysis model, and compare and analyze the health deviation of the wind turbine.

[0067] S4. Based on the health deviation analysis results of the wind turbine, the oil detection data of the wind turbine, and the inspection image data of the wind turbine blades, a wind turbine fault state analysis model is constructed to analyze the fault state of the wind turbine.

[0068] S5. Construct a dynamic scheduling strategy analysis model for wind turbines. Import the health deviation comparison analysis results and fault status analysis results of wind turbines into the dynamic scheduling strategy analysis model to generate a dynamic scheduling strategy for wind turbines and dynamically schedule the operation of wind turbines in the wind turbine generator set.

[0069] In this embodiment, step S2 involves constructing a wind turbine health operation dataset based on historical health tag data and wind turbine operation data, specifically including:

[0070] S21. Under the same monitoring time window, monitor the wind turbine operation data in the wind turbine generator set, and monitor the wind speed at the location of the wind turbine in real time to obtain the wind speed range at the location of the wind turbine.

[0071] S22. Divide the wind speed range at the location of the wind turbine into multiple wind speed sub-ranges at equal intervals; randomly select historical wind turbine operation data located in a certain wind speed sub-range from the historical health label data of the wind turbine as wind turbine anchor point samples;

[0072] S23. From the historical health label data of wind turbines, randomly select multiple historical wind turbine operation data corresponding to the same wind speed sub-interval as the wind turbine anchor point sample when there are no fault records and the wind turbine is operating normally as positive samples of wind turbine health.

[0073] S24. From the historical health label data of wind turbines, randomly select multiple historical wind turbine operation data corresponding to the same wind speed sub-interval as the wind turbine anchor point sample when there is a fault record or abnormal operation as wind turbine health negative samples.

[0074] S25. The selected wind turbine anchor point samples, positive wind turbine health samples, and negative wind turbine health samples are used as the wind turbine health operation dataset.

[0075] In this embodiment, step S2 involves inputting the wind turbine health operation dataset into a spatiotemporal comparison encoder to generate a wind turbine health benchmark feature library, specifically including:

[0076] S26. Import the wind turbine health operation dataset into the spatiotemporal contrast encoder model. With the training objective of maximizing the similarity of samples of the same class and minimizing the similarity of samples of different classes, train the spatiotemporal contrast encoder model until the model loss value converges; output the trained spatiotemporal contrast encoder model.

[0077] S27. Input the historical wind turbine operation data corresponding to the same wind turbine in all wind speed sub-intervals when there are no fault records and the wind turbine is operating normally into the trained spatiotemporal comparison encoder model, and output the wind turbine health feature vector in all wind speed sub-intervals.

[0078] For example, the training objective of the spatiotemporal contrastive encoder model provided in this embodiment is based on contrastive learning. By constructing positive samples (similar healthy data) and negative samples (abnormal data), the model narrows the distance between similar positive samples and widens the distance between dissimilar negative samples in the feature space. Specifically, the spatiotemporal contrastive encoder model provided in this embodiment maximizes the feature similarity of similar samples of multiple historical wind turbine operating data within the same wind speed range, thereby enhancing the model's ability to recognize patterns in healthy states; and maximizes the difference of dissimilar samples of wind turbine operating data in fault states, thereby improving the model's sensitivity to abnormal features. Furthermore, wind turbine operating data has spatiotemporal correlation. This embodiment uses a spatiotemporal contrastive encoder to capture the multidimensional features of wind turbine operating states more comprehensively by fusing temporal and spatial information.

[0079] S28. Based on the wind turbine health feature vectors in all wind speed sub-intervals, perform cluster analysis according to the wind speed sub-intervals to generate the wind turbine health benchmark feature vector corresponding to each wind speed sub-interval. Import the wind turbine health benchmark feature vectors corresponding to all wind speed sub-intervals into the database to obtain the wind turbine health benchmark feature library.

[0080] For example, this embodiment performs cluster analysis on the wind turbine health feature vector based on the physical characteristics and actual needs of the wind turbine operation data. Different wind speed ranges correspond to different wind turbine operating modes, such as standby mode at low wind speeds and full-load operation mode at high wind speeds. Therefore, clustering the health features according to wind speed sub-ranges can more accurately reflect the wind turbine health benchmark under specific operating conditions. By using a clustering algorithm to extract the most representative features from multiple health feature vectors within the same wind speed range, the influence of noisy data is eliminated, ensuring that the benchmark features can both cover typical health states and adapt to the diversity of data distribution.

[0081] In this embodiment, as Figure 2 As shown, step S3 involves a comparative analysis of the health deviation of the wind turbine, specifically including the following steps:

[0082] S31. Extract wind turbine operation data, wind turbine location meteorological data, and wind turbine health benchmark feature library;

[0083] S32. Construct a health deviation comparison analysis model. Import wind turbine operation data, wind turbine location meteorological data, and wind turbine health benchmark feature library into the health deviation comparison analysis model, compare and analyze the health deviation of the wind turbine, and obtain the health deviation analysis results of the wind turbine.

[0084] In this embodiment, the construction process of the health deviation comparison analysis model in step S32 includes the following specific steps:

[0085] S321. Extract the real-time wind speed from the meteorological data of the wind turbine location; based on the real-time wind speed, extract the wind turbine health benchmark feature vector corresponding to the wind speed sub-interval from the wind turbine health benchmark feature library;

[0086] S322. Obtain the wind turbine operation data at the current moment, input the wind turbine operation data into the spatiotemporal comparison encoder model, and obtain the feature vector of the wind turbine operation data at the current moment;

[0087] S323. Based on the feature vector of the wind turbine operation data at the current moment and the wind turbine health benchmark feature vector of the corresponding wind speed sub-interval, calculate the health deviation of the wind turbine at the current moment.

[0088] The formula for calculating health deviation is:

[0089]

[0090] In the formula, Dh is the health deviation of the wind turbine at the current moment, fb is the wind turbine health baseline feature vector of the wind speed sub-interval where the real-time wind speed is located, and fc is the feature vector of the wind turbine operation data at the current moment.

[0091] For example, in this embodiment, the calculation of health deviation can quantify the degree of difference between the current operating feature vector and the health baseline feature vector. This embodiment directly reflects the degree of abnormality in the current operating state of the wind turbine through health deviation; it not only provides a quantitative indicator for subsequent fault analysis but also supports the optimization process of dynamic scheduling strategies. For example, wind turbines with high health deviation can be preferentially reduced in load to decrease the probability of failure. Furthermore, this embodiment can also distinguish between transient disturbances and persistent anomalies by continuously monitoring the changing trend of health deviation, thereby improving the accuracy of fault diagnosis.

[0092] In this embodiment, as Figure 3 As shown, step S4 analyzes the fault status of the wind turbine, including the following specific steps:

[0093] S41. Obtain the health deviation analysis results of the wind turbine, and at the same time obtain the wind turbine oil detection data and wind turbine blade inspection image data;

[0094] S42. Import the health deviation analysis results of the wind turbine, the oil detection data of the wind turbine, and the inspection image data of the wind turbine blades into the wind turbine fault state analysis model to analyze the fault state of the wind turbine and obtain the fault state analysis results of the wind turbine.

[0095] In this embodiment, the construction process of the wind turbine fault state analysis model in step S42 includes the following specific steps:

[0096] S421. Based on the fan oil detection data, calculate the abnormal index of wear particle distribution of the fan;

[0097] The formula for calculating the wear particle distribution anomaly index is:

[0098]

[0099] In the formula, Lw is the abnormal distribution index of wear particles in the fan, D is the classification interval of wear particle size in the fan oil in the fan oil detection data, Nd is the number of wear particles with size located in the d classification interval detected in the fan oil detection data, Nt is the total number of wear particles detected in the fan oil detection data, Ud is the ratio of the total number of wear particles to the number of classification intervals, and Sd is the standard deviation of the number of wear particles in all classification intervals in the fan oil detection data.

[0100] For example, the calculation formula for the wear particle distribution anomaly index provided in this embodiment is based on the statistical distribution characteristics of oil detection data. The abrasive particle size distribution in the wind turbine oil follows specific statistical laws, while abnormal wear will cause the abrasive particle size distribution to deviate from theoretical expectations. Specifically, the calculation formula for the wear particle distribution anomaly index provided in this embodiment is used to quantify the degree of wear anomaly in mechanical components. For example, early wear of the wind turbine gearbox will lead to an abnormal increase in the proportion of small-sized abrasive particles, while bearing damage may trigger a sudden increase in large-sized abrasive particles. By observing the changes in the wear particle distribution anomaly index, potential faulty components can be located, and the severity of the fault can be judged in conjunction with the health deviation; thus, the limitations of a single fault state judgment indicator are overcome, providing a multi-dimensional analysis standard for the analysis of wind turbine fault states.

[0101] S422. Calculate the surface anomaly index of the wind turbine blades based on the wind turbine blade inspection image data;

[0102] The formula for calculating the surface anomaly index of wind turbine blades is:

[0103]

[0104] In the formula, At is the surface anomaly index of the wind turbine blade, n is the number of equal-area image blocks obtained by segmenting the blade inspection image in the wind turbine blade inspection image data, Ei is the texture entropy value of the i-th equal-area image block in the wind turbine blade inspection image data, Emax is the maximum value of the texture entropy values ​​of all equal-area image blocks in the wind turbine blade inspection image data, Emin is the minimum value of the texture entropy values ​​of all equal-area image blocks in the wind turbine blade inspection image data, and i is any term from 1 to n;

[0105] For example, in this embodiment, the calculation of the wind turbine blade surface anomaly index is based on image texture analysis technology. Damage caused by cracks or corrosion on the wind turbine blade surface, or contamination caused by ice and snow adhesion, can lead to significant changes in the local texture entropy value in the blade inspection image. By dividing the blade image into equal-area blocks, calculating the texture entropy of each block, and normalizing it, the influence of lighting and shooting angle can be eliminated. Specifically, this embodiment can effectively identify local damage or contamination on the wind turbine blade surface. For example, the texture entropy value of cracked areas in the blade inspection image is usually higher than that of smooth surfaces, while ice and snow adhesion can cause uneven entropy distribution. Combining the blade surface anomaly index provided in this embodiment with the wind turbine's health deviation and wear particle distribution anomaly index to form a multimodal fault diagnosis framework can significantly improve the comprehensiveness of fault detection.

[0106] S423. The health deviation of the fan, the abnormal index of wear particle distribution, and the abnormal index of blade surface are imported into the fan fault state calculation formula to calculate the fault state of the fan.

[0107] The formula for calculating the fault status of a wind turbine is as follows:

[0108]

[0109] In the formula, Bp represents the fault state of the fan at the current moment; when q=1, P1 represents the health deviation of the fan at the current moment; when q=2, P2 represents the abnormal distribution index of wear particles in the fan; and when q=3, P3 represents the abnormal surface index of the fan blades. This is the chain multiplication operator, where q represents any term from 1 to 3.

[0110] For example, this embodiment comprehensively evaluates the overall wind turbine fault status through multiplication operations. By fusing multi-source data, the false alarm rate of a single indicator is significantly reduced. For instance, instantaneous fluctuations in health deviation may be caused by environmental disturbances, but when they are accompanied by an abnormal increase in the wear particle distribution index and blade surface abnormalities, the wind turbine's fault status will clearly increase. This embodiment enables high-precision fault diagnosis of wind turbines under complex operating conditions and provides an optimization basis for the dynamic scheduling of wind turbine operation.

[0111] In this embodiment, as Figure 4 As shown, step S5 involves constructing a dynamic scheduling strategy analysis model for wind turbines, specifically including:

[0112] S51. Obtain the health deviation comparison analysis results and fault status analysis results of the wind turbine obtained from the analysis;

[0113] S52. Import the health deviation comparison analysis results and fault status analysis results of the wind turbines into the wind turbine dynamic scheduling strategy analysis model to generate a wind turbine dynamic scheduling strategy and dynamically schedule the operation of wind turbines in the wind turbine generator set.

[0114] In this embodiment, the construction process of the wind turbine dynamic scheduling strategy analysis model in step S52 specifically includes:

[0115] S521. Import the health deviation comparison analysis results and fault status analysis results of the wind turbine into the calculation formula of the wind turbine dynamic scheduling decision index, and calculate the dynamic scheduling decision index of the wind turbine at the current moment.

[0116] The formula for calculating the dynamic scheduling decision index of wind turbines is:

[0117]

[0118] In the formula, S is the dynamic scheduling decision index of the wind turbine at the current moment, Lc is the load rate of the wind turbine at the current moment, and Lmax is the rated maximum load rate of the wind turbine.

[0119] For example, this embodiment uses a wind turbine dynamic scheduling decision index to balance power generation efficiency and equipment health. This represents the ratio of the current load rate to the rated maximum load rate, used to reflect power generation efficiency. This is used to amplify the coupled effects of health deviation and fault state; when The larger the value, the greater the dynamic scheduling decision index of the wind turbine, indicating that the wind turbine needs to reduce its load to lower the risk of failure at the current moment. Specifically, in this embodiment, among two wind turbines with the same load rate, A larger wind turbine can be assigned a lower load, thus delaying the development of its failure, optimizing the short-term power generation efficiency of the wind turbine, and reducing long-term operation and maintenance costs by extending equipment life.

[0120] S522. Obtain the dynamic scheduling decision index of all wind turbines in the wind turbine generator set at the current moment and substitute it into the wind turbine dynamic scheduling optimization objective function;

[0121] The objective function for dynamic scheduling optimization of wind turbines is:

[0122]

[0123] In the formula, Sr is the dynamic scheduling decision index of the r-th wind turbine at the current time, Lcr is the load rate of the r-th wind turbine at the current time, LT is the total load demand of the wind turbine generator set, R is the number of wind turbines in the wind turbine generator set, and r is any one of 1 to R. The penalty coefficient is set to 0.1 by default in this embodiment.

[0124] For example, in this embodiment, the wind turbine dynamic scheduling optimization objective function takes minimizing the overall scheduling cost and meeting the total load demand as the optimization objective. Minimizing the sum of the products of the scheduling decision indices of all wind turbines and their load rates means allocating load under the premise of optimal health status; As a penalty, it is used to constrain the deviation between the total load factor and the total load demand.

[0125] S523. Calculate the gradient of the wind turbine dynamic scheduling optimization objective function using the gradient descent method, iteratively adjust the load rate of the wind turbines along the negative gradient direction, and output the optimal load rate of all wind turbines that makes the wind turbine dynamic scheduling optimization objective function converge after iterative adjustment. Then, redistribute the load of all wind turbines in the wind turbine generator set according to the optimal load rate of all wind turbines.

[0126] For example, this embodiment iteratively optimizes the load rate of each wind turbine using the gradient descent method, achieving a dynamic balance between power generation tasks and equipment health under actual operating conditions. For instance, when total load demand is high, this embodiment can appropriately increase the load of wind turbines with low dynamic scheduling decision indices while avoiding overuse of wind turbines with high dynamic scheduling decision indices. Specifically, this embodiment not only ensures stable power supply to the grid but also significantly improves the overall reliability of wind turbine generator sets through refined scheduling.

[0127] Example 2

[0128] An electronic device according to an embodiment of the present invention includes a processor and a memory, wherein the memory stores a computer program that can be called by the processor, and the processor executes a fault prediction and health management method based on contrastive learning by calling the computer program stored in the memory. It should be noted that all computer programs for the fault prediction and health management method based on contrastive learning are implemented using the C programming language.

[0129] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0130] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A fault prediction and health management method based on contrastive learning, characterized in that, Includes the following steps: S1. Acquire wind turbine operation data, wind turbine location meteorological data, and historical health label data of the wind turbine generator set; at the same time, acquire wind turbine oil detection data and wind turbine blade inspection image data; S2. Under equal monitoring time windows, monitor the wind turbine operation data of the wind turbine generator set and monitor the wind speed at the location of the wind turbine in real time to obtain the wind speed range at the location of the wind turbine; divide the wind speed range at the location of the wind turbine into multiple wind speed sub-ranges at equal distances; randomly select historical wind turbine operation data within a certain wind speed sub-range from the historical health label data of the wind turbine as wind turbine anchor point samples; and select positive and negative samples of wind turbine health in the corresponding sub-ranges to jointly construct the wind turbine health operation dataset; import the wind turbine health operation dataset into the spatiotemporal contrast encoder model to maximize the similarity of samples of the same class and minimize the similarity of samples of different classes as the training. The objective is to train a spatiotemporal contrast encoder model until the model loss value converges; output the trained spatiotemporal contrast encoder model; input historical wind turbine operation data corresponding to the same wind turbine within all wind speed sub-intervals when there are no fault records and the operation is normal into the trained spatiotemporal contrast encoder model, and output wind turbine health feature vectors within all wind speed sub-intervals; based on the wind turbine health feature vectors within all wind speed sub-intervals, perform cluster analysis by wind speed sub-interval to generate wind turbine health benchmark feature vectors corresponding to each wind speed sub-interval; import the wind turbine health benchmark feature vectors corresponding to all wind speed sub-intervals into a database to obtain a wind turbine health benchmark feature library. S3. Extract the real-time wind speed from the meteorological data of the wind turbine location; Based on real-time wind speed, wind turbine health benchmark feature vectors for corresponding wind speed sub-intervals are extracted from the wind turbine health benchmark feature library; wind turbine operation data at the current moment is obtained and input into the spatiotemporal comparison encoder model to obtain the feature vector of wind turbine operation data at the current moment; based on the feature vector of wind turbine operation data at the current moment and the wind turbine health benchmark feature vector for the corresponding wind speed sub-interval, the health deviation of the wind turbine at the current moment is calculated. S4. Based on the health deviation analysis results of the wind turbine, the oil detection data of the wind turbine, and the inspection image data of the wind turbine blades, a wind turbine fault state analysis model is constructed to analyze the fault state of the wind turbine. S5. Construct a dynamic scheduling strategy analysis model for wind turbines. Import the health deviation comparison analysis results and fault status analysis results of wind turbines into the dynamic scheduling strategy analysis model to generate a dynamic scheduling strategy for wind turbines and dynamically schedule the operation of wind turbines in the wind turbine generator set.

2. The fault prediction and health management method based on contrastive learning according to claim 1, characterized in that, The formula for calculating the health deviation is: ； In the formula, Dh is the health deviation of the wind turbine at the current moment, fb is the wind turbine health benchmark feature vector of the wind speed sub-interval where the real-time wind speed is located, and fc is the feature vector of the wind turbine operation data at the current moment.

3. The fault prediction and health management method based on contrastive learning according to claim 2, characterized in that, The analysis of the fault status of the wind turbine in step S4 includes the following specific steps: S41. Obtain the health deviation analysis results of the wind turbine, and at the same time obtain the wind turbine oil detection data and wind turbine blade inspection image data; S42. Import the health deviation analysis results of the wind turbine, the oil detection data of the wind turbine, and the inspection image data of the wind turbine blades into the wind turbine fault state analysis model to analyze the fault state of the wind turbine and obtain the fault state analysis results of the wind turbine.

4. The fault prediction and health management method based on contrastive learning according to claim 3, characterized in that, The construction process of the wind turbine fault state analysis model in step S42 includes the following specific steps: S421. Based on the fan oil detection data, calculate the abnormal index of wear particle distribution of the fan; S422. Calculate the surface anomaly index of the wind turbine blades based on the wind turbine blade inspection image data; S423. The health deviation of the fan, the abnormal index of wear particle distribution, and the abnormal index of blade surface are imported into the fan fault state calculation formula to calculate the fault state of the fan. The formula for calculating the fault status of a wind turbine is as follows: In the formula, Bp represents the fault state of the fan at the current moment; when q=1, P1 represents the health deviation of the fan at the current moment; when q=2, P2 represents the abnormal index of wear particle distribution of the fan; and when q=3, P3 represents the abnormal index of the fan blade surface. This is the chain multiplication operator, where q represents any term from 1 to 3.

5. The fault prediction and health management method based on contrastive learning according to claim 4, characterized in that, The step S5, which involves constructing a dynamic scheduling strategy analysis model for wind turbines, specifically includes: S51. Obtain the health deviation comparison analysis results and fault status analysis results of the wind turbine obtained from the analysis; S52. Import the health deviation comparison analysis results and fault status analysis results of the wind turbines into the wind turbine dynamic scheduling strategy analysis model to generate a wind turbine dynamic scheduling strategy and dynamically schedule the operation of wind turbines in the wind turbine generator set.

6. The fault prediction and health management method based on contrastive learning according to claim 5, characterized in that, The construction process of the wind turbine dynamic scheduling strategy analysis model in step S52 specifically includes: S521. Import the health deviation comparison analysis results and fault status analysis results of the wind turbine into the calculation formula of the wind turbine dynamic scheduling decision index, and calculate the dynamic scheduling decision index of the wind turbine at the current moment. The formula for calculating the dynamic scheduling decision index of wind turbines is: ； In the formula, S is the dynamic scheduling decision index of the wind turbine at the current moment, Lc is the load rate of the wind turbine at the current moment, and Lmax is the rated maximum load rate of the wind turbine. S522. Obtain the dynamic scheduling decision index of all wind turbines in the wind turbine generator set at the current moment and substitute it into the wind turbine dynamic scheduling optimization objective function; S523. Calculate the gradient of the wind turbine dynamic scheduling optimization objective function using the gradient descent method, iteratively adjust the load rate of the wind turbines along the negative gradient direction, and output the optimal load rate of all wind turbines that makes the wind turbine dynamic scheduling optimization objective function converge after iterative adjustment. Then, redistribute the load of all wind turbines in the wind turbine generator set according to the optimal load rate of all wind turbines.

7. An electronic device, comprising: A processor and a memory, wherein the memory stores a computer program that can be called by the processor; characterized in that the processor executes the fault prediction and health management method based on contrastive learning as described in any one of claims 1-6 by calling the computer program stored in the memory.

Images