A dynamic predictive maintenance method for a main bearing of a fan considering different critical states

Abstract

The present application relates to the technical field of wind turbine intelligent operation and maintenance, and discloses a wind turbine main bearing dynamic predictive maintenance method considering different critical states, which solves the problems of multi-source data fusion difficulty, static conservative maintenance strategy and high cost in the prior art. The present application comprises: acquiring multi-source heterogeneous monitoring data; constructing a multi-source information fusion domain adaptive model, extracting space-time characteristics and outputting a remaining useful life prediction result; dividing the degradation process into low, medium and high critical states according to the prediction result, respectively matching binary state checking, quantitative trend checking and accurate root cause checking, and selecting corrective maintenance, preventive maintenance or no maintenance according to the checking result; and minimizing the full life cycle maintenance cost rate as the target, and using a particle swarm optimization algorithm to optimize the decision variable. The present application improves the accuracy and reliability of the remaining useful life prediction, significantly reduces the operation and maintenance cost, and experiments show that the maintenance cost rate is reduced by about 30%-50% compared with the traditional strategy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for wind turbine generator sets. Specifically, it relates to a dynamic predictive maintenance method for wind turbine main bearings that takes into account different critical states. Background Technology

[0002] In recent years, global investment in wind power technology and assets has been steadily increasing. Direct-drive wind turbines, with their advantages of simple structure, high power generation efficiency, and environmental friendliness, are playing an increasingly important role in the green energy process. The operation and maintenance phases of wind power equipment represent the most complex aspect of equipment lifecycle management, involving a wide range of content, complex information, and data. Operation and maintenance costs remain a significant obstacle to value enhancement in the wind power industry. Wind power is a capital-intensive industry, and cost control is a key focus for the energy sector. Maintenance costs for wind power equipment account for as much as 10% to 15% of total wind farm revenue. Efficient and scientific maintenance strategies are the core engine for ensuring safe asset operation, improving power generation efficiency, and maximizing asset value throughout its entire lifecycle.

[0003] Direct-drive wind turbine generators possess advantages such as simple structure, high power generation efficiency, and environmental friendliness, playing a vital role in the green energy development process. As the core load-bearing and transmission hub of a direct-drive wind turbine, the main bearing is subjected to harsh environments such as low-speed variable loads and high-temperature salt spray over extended periods. Its performance degradation process is complex, making it difficult to guarantee the long-term accuracy and reliability of lifespan prediction. Traditional maintenance strategies are mainly divided into reactive maintenance and preventative maintenance: reactive maintenance is costly and results in significant downtime losses; preventative maintenance, on the other hand, incurs high operation and maintenance costs due to unnecessary intervention in most healthy assets.

[0004] With the development of artificial intelligence technology, predictive maintenance has gradually become a research hotspot. Existing methods mostly use single-sensor data (such as vibration signals) to predict remaining service life, neglecting the complementarity and synergistic effects between multi-source heterogeneous data. Furthermore, single sensors are susceptible to interference in extreme environments, reducing data reliability. Some studies have attempted to fuse multi-sensor data, but the heterogeneity of multi-source data leads to differences in feature distribution and noise levels, and the data distribution offset problem between the source and target domains has not been effectively resolved. Regarding maintenance strategies, most existing predictive maintenance decisions are based on fixed service life thresholds, failing to adjust inspection and maintenance actions in stages according to the dynamic changes in the degradation process, resulting in limited room for cost optimization.

[0005] In view of this, the present invention is proposed. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows:

[0007] A dynamic predictive maintenance method for wind turbine main bearings considering different critical states. This method requires no model training, is highly interpretable, and can be used for online prediction. The method includes the following steps:

[0008] 1.1 Acquire multi-source heterogeneous monitoring data of the main bearing of the wind turbine, wherein the multi-source heterogeneous monitoring data includes SCADA data, CMS vibration data, ultrasonic sensor data and temperature sensor array data;

[0009] 1.2 Construct a multi-source information fusion domain adaptation model based on spatiotemporal graph convolutional neural network, extract and fuse spatiotemporal features of the multi-source heterogeneous monitoring data, and output the prediction result of the remaining service life of the main bearing;

[0010] 1.3 Based on the predicted remaining service life, the degradation process of the main bearing is divided into low-critical state, medium-critical state and high-critical state;

[0011] 1.4. For the low critical state, a binary state inspection method is used; for the medium critical state, a quantitative trend inspection method is used; for the high critical state, a precise root cause inspection method is used; and corrective maintenance, preventive maintenance, or maintenance-free operation is selected based on the inspection results.

[0012] 1.5. With minimizing the life-cycle maintenance cost rate as the optimization objective, the particle swarm optimization algorithm is used to optimize the decision variables in the maintenance strategy to obtain the optimal dynamic predictive maintenance strategy.

[0013] In a preferred embodiment of the present invention, the multi-source information fusion domain adaptation model in step 1.2 includes multiple domain-specific feature extractors, multiple domain-specific discriminators, and multiple domain-specific regressors; the domain-specific feature extractor adopts a "space-time-space" sandwich structure, which includes a first temporal convolutional layer, a graph convolutional layer, and a second temporal convolutional layer in sequence, and introduces residual connections.

[0014] As a preferred embodiment of the present invention, the spatiotemporal feature extraction and fusion further includes: adaptive weighted fusion of input features based on an attention mechanism, wherein the attention weights are calculated through a multilayer perceptron and normalized using the SoftMax function.

[0015] In a preferred embodiment of the present invention, the multi-source information fusion domain adaptation model further includes an intra-group alignment strategy and an inter-group alignment strategy; the intra-group alignment strategy uses the multi-core maximum mean difference to measure the distance between the feature distributions of the source domain and the target domain under the same operating condition; the inter-group alignment strategy measures the distance between the feature distributions of the source domain under different operating conditions.

[0016] In a preferred embodiment of the present invention, the domain discriminator consists of two fully connected layers, and its adversarial loss is implemented using a gradient inversion layer to enable the feature extractor to learn domain-invariant features.

[0017] In a preferred embodiment of the present invention, the classification of low-critical state, medium-critical state and high-critical state in step 1.3 is as follows: when the remaining useful life prediction result is higher than the binary state check threshold, it is determined to be a low-critical state; when the remaining useful life prediction result is lower than the binary state check threshold but higher than the precise root cause check threshold, it is determined to be a medium-critical state; when the remaining useful life prediction result is lower than the precise root cause check threshold, it is determined to be a high-critical state.

[0018] In a preferred embodiment of the present invention, step 1.4 further includes:

[0019] 1.7.1 The binary state inspection method monitors whether the key performance parameters of the main bearing exceed the abnormal state trigger threshold and outputs whether the system is operating normally.

[0020] 1.7.2 Quantitative trend inspection method: By periodically and repeatedly measuring and recording the specific values ​​of key state parameters, a performance degradation model is established to quantify the degradation rate;

[0021] 1.7.3 The precise root cause detection method increases the detection frequency and performs synchronous analysis and spatiotemporal feature extraction of multi-source information such as vibration, temperature, sound signals and current.

[0022] In a preferred embodiment of the present invention, the maintenance operation selection rule in step 1.4 is as follows:

[0023] 1.8.1 When the inspection results indicate that the main bearing is in a faulty state, corrective maintenance shall be performed;

[0024] 1.8.2 When the inspection results show that the main bearing is in normal operating condition but the predicted remaining service life is lower than the preventive maintenance threshold, preventive maintenance shall be carried out; when the predicted remaining service life is lower than the preventive replacement threshold, preventive replacement shall be carried out.

[0025] 1.8.3 Otherwise, no maintenance operation will be performed.

[0026] In a preferred embodiment of the present invention, the decision variables in step 1.5 include the inspection interval, the binary state inspection threshold, the precise root cause inspection threshold, the preventive maintenance threshold, and the preventive replacement threshold; the maintenance cost rate consists of the inspection cost, the maintenance cost, and the cost of power generation loss due to downtime.

[0027] In a preferred embodiment of the present invention, the wind turbine is a direct-drive wind turbine generator set, and the main bearing is the main bearing of the direct-drive wind turbine.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] 1. The present invention proposes a multi-source information fusion domain adaptation method based on spatiotemporal graph convolutional neural network, which realizes the spatiotemporal feature extraction and fusion of multi-source monitoring data, effectively solves the domain distribution offset problem, and significantly improves the accuracy and robustness of remaining service life prediction.

[0030] 2. This invention proposes a three-stage dynamic predictive maintenance strategy, dividing the main bearing degradation process into three critical states: low, medium, and high. Differentiated inspection and maintenance actions are matched for each stage. This dynamically balances the failure risk, lifespan utilization, and maintenance cost of the wind turbine main bearing. This method surpasses the static strategies of existing models, comprehensively optimizing inspection, maintenance, and operation.

[0031] 3. This invention takes maintenance cost rate as the optimization objective and uses particle swarm optimization algorithm to globally optimize decision variables such as inspection interval and inspection threshold. Experimental results show that compared with traditional corrective maintenance strategy, this invention can reduce maintenance cost rate by about 50.19%; compared with conventional predictive maintenance strategy, it can reduce it by about 29.83%.

[0032] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0033] In the attached diagram:

[0034] Figure 1 This is a schematic diagram of the main bearing of the direct-drive fan of the present invention;

[0035] Figure 2 This is the overall diagram of the dynamic predictive maintenance framework of the present invention;

[0036] Figure 3 This is a structural framework diagram of the multi-source information fusion domain adaptation model of the present invention;

[0037] Figure 4 This is a schematic diagram of the main bearing dynamic predictive maintenance plan of the present invention;

[0038] Figure 5 This is a diagram of the dynamic predictive maintenance logic framework of the present invention;

[0039] Figure 6 This is a schematic diagram of the wind turbine drivetrain remote condition monitoring system (private cloud platform) of the present invention;

[0040] Figure 7This is a graph showing the trend of maintenance cost rate and inspection interval in this invention;

[0041] Figure 8 This is a trend graph showing the changes in the binary state check threshold, the precise root cause check threshold, and the maintenance cost rate of the present invention.

[0042] Figure 9 This is a graph showing the changing trends of preventive maintenance, replacement thresholds, and maintenance cost rates according to the present invention.

[0043] Figure 10 This is a comparison chart of the maintenance cost rate of the ablation experiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention.

[0045] like Figure 1 - Figure 10 As shown, a dynamic predictive maintenance method for wind turbine main bearings considering different critical states is proposed. This method requires no model training, is highly interpretable, and can be used for online prediction. The method includes the following steps:

[0046] 1.1 Acquire multi-source heterogeneous monitoring data of the main bearing of the wind turbine. The multi-source heterogeneous monitoring data includes SCADA data, CMS vibration data, ultrasonic sensor data, and temperature sensor array data. The ultrasonic sensor is used to capture high-frequency stress waves (20kHz–1MHz) generated by micro-deformation, crack initiation, early pitting, lubrication failure, etc. inside the main bearing. The temperature sensor array is arranged at key positions on the outer circumference of the main bearing (lower half of the load area) to enhance and deepen the existing temperature measurement data of SCADA.

[0047] 1.2 Construct a multi-source information fusion domain adaptation model based on spatiotemporal graph convolutional neural network, extract and fuse spatiotemporal features of the multi-source heterogeneous monitoring data, and output the prediction result of the remaining service life of the main bearing;

[0048] 1.3 Based on the predicted remaining service life, the degradation process of the main bearing is divided into low-critical state, medium-critical state and high-critical state;

[0049] 1.4. For the low critical state, a binary state inspection method is used; for the medium critical state, a quantitative trend inspection method is used; for the high critical state, a precise root cause inspection method is used; and corrective maintenance, preventive maintenance, or maintenance-free operation is selected based on the inspection results.

[0050] 1.5. With minimizing the life-cycle maintenance cost rate as the optimization objective, the particle swarm optimization algorithm is used to optimize the decision variables in the maintenance strategy to obtain the optimal dynamic predictive maintenance strategy.

[0051] In step 1.2, the specific implementation steps are as follows:

[0052] Import the collected multi-source monitoring data into the Python software. The source domain contains labeled data under three different operating conditions (different wind speed ranges and different ambient temperatures), and the target domain is the unlabeled data of the wind turbine to be predicted. All data are standardized and synchronized to ensure data quality and consistency.

[0053] Multiple domain-specific feature extractors are constructed, each corresponding to a source domain condition. A "space-time-space" sandwich structure is adopted, which includes a first temporal convolutional layer, a graph convolutional layer, and a second temporal convolutional layer, and residual connections are introduced. The graph convolutional network model is shown in Equation (1):

[0054] (1)

[0055] In the formula, For the first The feature matrix of the layer, For the first The learnable weight matrix of the layer, To add self-connected adjacency matrices, This is the corresponding degree matrix. This is a non-linear activation function. Through aggregation and transformation operations, new node representations are generated based on the features of the node itself and its local neighborhood. To further capture the spatial dependencies and temporal dynamics of multi-source data simultaneously, a 1D CNN temporal convolution that operates only along the time dimension is incorporated. For the input features... , Represents the number of points in space. Indicates the depth level. The feature dimensions are represented (from different data sources). The temporal convolution operation is shown in Equation 2.

[0056] (2)

[0057] In the formula, The convolution kernel weights, This represents the size of the convolution kernel in the time dimension. Indicates the fill size. This is the bias term. The kernel size is set to... This ensures that convolution occurs only along the time dimension.

[0058] An adaptive weighted fusion of input features is performed based on an attention mechanism. First, a set of attention weights associated with each feature vector is calculated using the SoftMax function. The feature attention network is implemented through a multilayer perceptron. The original features are then weighted and aggregated to obtain the fused features.

[0059] Intra-group alignment and inter-group alignment strategies are employed to eliminate data distribution offset between the source and target domains. Intra-group alignment uses multi-core maximum mean difference to measure the distance between the feature distributions of the source and target domains under the same operating conditions, while inter-group alignment measures the distance between the feature distributions of the source domain under different operating conditions. The alignment loss is shown in equations (3) and (4):

[0060] (3)

[0061] (4)

[0062] Multiple domain-specific discriminators are constructed, each consisting of two fully connected layers. Adversarial training is achieved using gradient inversion layers, enabling the feature extractor to learn domain-invariant features. The adversarial loss is shown in Equation (5).

[0063] (5)

[0064] The extracted high-quality spatiotemporal features are input into a domain-specific regressor, which outputs the remaining lifetime prediction results. The regression loss uses mean squared error, as shown in Equation (6);

[0065] (6)

[0066] In step 1.3, based on the remaining service life prediction results, the degradation process of the main bearing is divided into low-critical, medium-critical, and high-critical states. The classification criteria are as follows: when the remaining service life prediction result is higher than the binary state check threshold, it is determined to be a low-critical state; when the remaining service life prediction result is lower than the binary state check threshold but higher than the precise root cause check threshold, it is determined to be a medium-critical state; and when the remaining service life prediction result is lower than the precise root cause check threshold, it is determined to be a high-critical state.

[0067] In step 1.4, differentiated inspection methods are matched for different critical states:

[0068] In the implementation case, the specific implementation steps are as follows:

[0069] Low-critical state: A binary state inspection method is adopted, which monitors whether the key performance parameters of the main bearing exceed the abnormal state trigger threshold and outputs whether the system is operating normally, thus minimizing inspection costs. The output of the binary state inspection is shown in equation (7):

[0070] (7)

[0071] In the formula This is the output of the binary state check method. This is a key performance degradation monitoring signal for the main bearing. This is the threshold for triggering abnormal states.

[0072] Intermediate critical state: A quantitative trend check method is used to periodically and repeatedly measure and record the specific values ​​of key state parameters, establish a performance degradation model, and quantify the degradation rate. For example, track the change curve of the effective value of vibration velocity over time to observe whether it rises steadily, or establish a correlation model between bearing temperature and ambient temperature, and fan power / speed.

[0073] High-critical state: Employing precise root cause analysis methods and increasing the detection frequency, synchronously analyzing and extracting spatiotemporal features from multiple sources such as vibration, temperature, acoustic signals, and current, and obtaining a large amount of physical information to further improve the accuracy of the prognostic model.

[0074] Based on the inspection results, select one of the following maintenance operations:

[0075] Corrective maintenance: This is performed when inspection results indicate that the main bearing is in a faulty state, and incurs corrective maintenance costs. ;

[0076] Preventive maintenance: This is used when inspection results indicate that the main bearing is in normal operating condition but the predicted remaining service life is lower than the preventive maintenance threshold; if it is lower than the preventive replacement threshold, preventive replacement is performed, incurring preventive maintenance costs. ;

[0077] Maintenance-free: This is used when the inspection results indicate that the main bearing is in normal operating condition and the predicted remaining service life is higher than the preventive maintenance threshold, with zero maintenance cost.

[0078] In step 1.5, the optimization objective is to minimize the life-cycle maintenance cost rate, and the particle swarm optimization algorithm is used to globally optimize the decision variables.

[0079] In the implementation case, the specific implementation steps are as follows:

[0080] The formula for calculating the maintenance cost rate is shown in equation (8):

[0081] (8)

[0082] in, The total cumulative maintenance cost includes inspection costs, maintenance costs, and costs of lost power generation due to downtime. The update cycle is specified. Decision variables include the inspection interval. Binary state check threshold Precise root cause detection threshold Preventive maintenance threshold and preventative replacement threshold The objective function of the particle swarm optimization algorithm is shown in equation (9):

[0083] (9)

[0084] The constraint is the range of values ​​for each decision variable.

[0085] The main bearing of a 1.5MW direct-drive wind turbine generator set in a wind farm in Inner Mongolia Autonomous Region, China, was used as the experimental object for verification.

[0086] In the implementation case, the specific implementation steps are as follows:

[0087] Maintenance cost data for the main bearing of a direct-drive wind turbine was calculated based on NREL reports and literature data. This includes inspection costs, maintenance costs, and loss rates.

[0088] The optimal decision variable values ​​were obtained using a particle swarm optimization algorithm (30 particles, maximum 100 iterations): inspection interval 13 months, binary state inspection threshold 0.65, precise root cause inspection threshold 0.25, preventive maintenance threshold 0.45, and preventive replacement threshold 0.15. At this point, the minimum maintenance cost rate was $8.612 / MWh.

[0089] Compared with traditional maintenance strategies: the maintenance cost rate for corrective maintenance strategy (S1) is $17.288 / MWh, for preventive maintenance strategy (S2) it is $12.273 / MWh, and for conventional predictive maintenance strategies (S3 and S4) it is $10.536 and $10.361 / MWh, respectively. The strategy proposed in this invention reduces the cost by approximately 50.19% compared to S1 and by approximately 29.83% compared to S4.

[0090] Ablation experiments verified the contribution of the multi-source information fusion module and the three-stage hierarchical strategy to cost reduction. Using only vibration signal threshold judgment (M1) resulted in the highest maintenance cost rate; not using the MIFDA model (M2) resulted in a higher cost rate than the present invention; and the cost rates of removing low, medium, and high critical states (M3, M4, and M5) were all higher than the complete strategy of the present invention (M6).

[0091] This invention proposes a dynamic predictive maintenance method for wind turbine main bearings, which greatly improves the accuracy of remaining service life prediction and significantly reduces operation and maintenance costs.

[0092] This invention utilizes multi-source heterogeneous monitoring data to construct a MIFDA model based on a spatiotemporal graph convolutional neural network for predicting remaining service life. Based on the prediction results, the degradation process is divided into three critical states: low, medium, and high. Binary state checks, quantitative trend checks, and precise root cause checks are then performed on each state. With the goal of minimizing the maintenance cost rate, a particle swarm optimization algorithm is used to optimize the decision variables, thereby achieving dynamic predictive maintenance.

[0093] It is understood that the present invention has been described through some embodiments, which are known to those skilled in the art. Various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A dynamic predictive maintenance method for wind turbine main bearings considering different critical states, characterized in that the method requires no model training, has strong interpretability, and can be used for online prediction, and the method includes the following steps: 1.1 Acquire multi-source heterogeneous monitoring data of the main bearing of the wind turbine, wherein the multi-source heterogeneous monitoring data includes SCADA data, CMS vibration data, ultrasonic sensor data and temperature sensor array data; 1.2 Construct a multi-source information fusion domain adaptation model based on spatiotemporal graph convolutional neural network, extract and fuse spatiotemporal features of the multi-source heterogeneous monitoring data, and output the prediction result of the remaining service life of the main bearing; 1.3 Based on the predicted remaining service life, the degradation process of the main bearing is divided into low-critical state, medium-critical state and high-critical state; 1.

4. For the low critical state, a binary state inspection method is used; for the medium critical state, a quantitative trend inspection method is used; for the high critical state, a precise root cause inspection method is used; and corrective maintenance, preventive maintenance, or maintenance-free operation is selected based on the inspection results. 1.

5. With minimizing the life-cycle maintenance cost rate as the optimization objective, the particle swarm optimization algorithm is used to optimize the decision variables in the maintenance strategy to obtain the optimal dynamic predictive maintenance strategy.

2. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 1, characterized in that, The multi-source information fusion domain adaptation model in step 1.2 includes multiple domain-specific feature extractors, multiple domain-specific discriminators, and multiple domain-specific regressors. The domain-specific feature extractor adopts a "space-time-space" sandwich structure, which includes a first temporal convolutional layer, a graph convolutional layer, and a second temporal convolutional layer in sequence, and introduces residual connections.

3. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 2, characterized in that, The spatiotemporal feature extraction and fusion further includes: adaptive weighted fusion of input features based on an attention mechanism, wherein the attention weights are calculated through a multilayer perceptron and normalized using the SoftMax function.

4. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 2, characterized in that, The multi-source information fusion domain adaptation model also includes an intra-group alignment strategy and an inter-group alignment strategy; the intra-group alignment strategy uses the multi-core maximum mean difference to measure the distance between the feature distributions of the source domain and the target domain under the same working condition; the inter-group alignment strategy measures the distance between the feature distributions of the source domain under different working conditions.

5. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 2, characterized in that, The domain discriminator consists of two fully connected layers, and its adversarial loss is implemented using a gradient inversion layer, which enables the feature extractor to learn domain-invariant features.

6. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 1, characterized in that, In step 1.3, the classification of low-critical state, medium-critical state, and high-critical state is as follows: when the remaining useful life prediction result is higher than the binary state check threshold, it is determined to be a low-critical state; when the remaining useful life prediction result is lower than the binary state check threshold but higher than the precise root cause check threshold, it is determined to be a medium-critical state; when the remaining useful life prediction result is lower than the precise root cause check threshold, it is determined to be a high-critical state.

7. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 1, characterized in that, Step 1.4 further includes: 1.7.1 The binary state inspection method monitors whether the key performance parameters of the main bearing exceed the abnormal state trigger threshold and outputs whether the system is operating normally. 1.7.2 Quantitative trend inspection method: By periodically and repeatedly measuring and recording the specific values ​​of key state parameters, a performance degradation model is established to quantify the degradation rate; 1.7.3 The precise root cause detection method increases the detection frequency and performs synchronous analysis and spatiotemporal feature extraction of multi-source information such as vibration, temperature, sound signals and current.

8. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 1, characterized in that, The maintenance operation selection rules in step 1.4 are as follows: 1.8.1 When the inspection results indicate that the main bearing is in a faulty state, corrective maintenance shall be performed; 1.8.2 When the inspection results show that the main bearing is in normal operating condition but the predicted remaining service life is lower than the preventive maintenance threshold, preventive maintenance shall be carried out; when the predicted remaining service life is lower than the preventive replacement threshold, preventive replacement shall be carried out. 1.8.3 Otherwise, no maintenance operation will be performed.

9. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 1, characterized in that, The decision variables in step 1.5 include the inspection interval, binary state inspection threshold, precise root cause inspection threshold, preventive maintenance threshold, and preventive replacement threshold; the maintenance cost rate consists of inspection cost, maintenance cost, and power generation loss cost during downtime.

10. The dynamic predictive maintenance method for wind turbine main bearings considering different critical states according to claim 1, characterized in that, The wind turbine is a direct-drive wind turbine generator set, and the main bearing is the main bearing of the direct-drive wind turbine.

Images