Method for predicting service life of main transmission chain part of wind driven generator under sparse data condition

Abstract

A method for predicting the service life of a main transmission chain part of a wind driven generator under a sparse data condition comprises the following steps: constructing a degradation evolution model in an index form to describe the state change of the part, and introducing working condition disturbance into the model to improve the characterization capability of actual operation uncertainty; a time scale translation mechanism is introduced, the relative position of observation data in the degradation process is calibrated, and matching between different observation points and a degradation model under the unified time scale is achieved; in the parameter estimation process, a combined loss function fusing fitting errors and degradation monotonicity constraints is designed, and the convergence stability and fitting reliability of the model under the sparse data condition are improved in combination with a parameter optimization strategy of the self-adaptive backtracking step length; on the basis of degradation model parameters obtained through optimization, the residual life of the main transmission chain component is calculated through degradation trend extrapolation and failure threshold judgment; according to the method, the part degradation rule is accurately described under the limited observation condition, and the stability and engineering applicability of the life prediction result are improved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine main drivetrain life prediction technology, specifically to a method for predicting the life of wind turbine main drivetrain components under sparse data conditions. Background Technology

[0002] With the ongoing transformation of the energy structure, wind turbines have gradually become an important pillar of the new energy power generation system, and their safe and stable operation is crucial for ensuring energy supply. However, wind turbines are typically deployed in complex operating environments such as the field or offshore for extended periods, making them susceptible to multiple adverse factors such as drastic temperature fluctuations, wind and sand erosion, and humidity changes. This results in significant sparsity and irregularity in the temporal distribution of data collected by sensors. Harsh environments can easily trigger sensor failures, and the combined effects of environmental disturbances lead to uncertainties in data acquisition, further exacerbating the problem of missing data. The time span, starting position, and specific location of such missing data have no fixed pattern, exhibiting significant randomness.

[0003] Existing methods for predicting the lifespan and monitoring the condition of wind turbine main drivetrain components largely rely on traditional supervised modeling approaches, whose effectiveness depends on complete historical data. However, in practical applications, wind turbine main drivetrain components often lack complete lifecycle data, data sampling stability is insufficient, and the degradation stage of the current observation point within the entire lifecycle cannot be accurately determined. This results in low prediction accuracy and poor adaptability of traditional methods, making it difficult to meet actual engineering needs. Therefore, developing lifespan prediction technology for wind turbine main drivetrain components for sparse data scenarios has become a critical issue that urgently needs to be addressed in the current wind turbine operation and maintenance field. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention aims to provide a method for predicting the lifespan of wind turbine main drivetrain components under sparse data conditions. By employing signal preprocessing, multi-dimensional feature extraction, adaptive parameter optimization, and time scale calibration techniques, the method reduces reliance on complete historical data and stable sampling conditions, significantly improving the accuracy and reliability of lifespan prediction for wind turbine main drivetrain components under sparse data scenarios, and providing precise decision support for preventive maintenance of wind turbine equipment.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting the lifespan of main drivetrain components of a wind turbine under sparse data conditions includes the following steps: Step 1, Degradation function setting: Based on the nonlinear degradation characteristics of the main drivetrain components of a wind turbine (such as the main bearing and the generator), an exponential degradation model is used to describe its state evolution process. The expression of the exponential degradation model is as follows: in, The equipment is in a degraded state. It is relative to the time of the onset of degradation. These are the parameters corresponding to the degradation process. Uncertainties arising from operating condition disturbances such as speed fluctuations; degenerate function ,in, For the degenerate initial state coefficients, For degradation rate coefficient, This is the degraded baseline offset; The second step is time scale translation: To address the irregular temporal distribution of sparse data, a time shift is introduced to calibrate the time axis of the observed data. The formula is as follows: in, This is the time shift, representing the offset of the observed data throughout the entire degradation trend, used to correct time misalignments caused by missing data; This represents the degradation trend function of the calibrated sparse data, ensuring that different observation points match the degradation model at a uniform time scale. For the first The actual time corresponding to each observation point For the first Degradation state observations at each observation point For the first Uncertainties caused by operating condition disturbances such as constant speed fluctuations; The third step is to set the loss function: To address the constraints on fitting accuracy and monotonicity of degradation trends in sparse data, a combined loss function is designed, as follows: The expression for the loss function is: in, The Huber loss term is used to reduce the interference of outliers on the fitting results, while taking into account both the gradient stability of the mean squared error and the robustness of the absolute error. Its calculation method is as follows: in, The number of samples in the observation data. For the first The actual degradation state observations of each sample For the first Predicted degradation status of each sample The threshold parameter for Huber loss; This is a monotonicity loss term. Since the degradation process of wind turbine equipment has irreversible monotonicity (eigenvalues ​​monotonically increase / decrease over time), this loss term constrains the monotonicity of the prediction results, avoiding the unreasonable situation of "reverse fluctuations in degradation eigenvalues." Its calculation method is as follows: in, For the first The predicted degradation state of each sample, and this loss term is used to constrain the monotonicity of the degradation trend; The monotonicity constraint weights are used to balance the priority between fitting accuracy and monotonicity constraints. The weights of the monotonicity loss are dynamically updated during the iteration process, using the following formula: in, For the first The monotonicity constraint weights for the next iteration; For the first The monotonicity constraint weights for the next iteration; The learning rate used to update the weights; This represents the mean of the monotonicity loss; This represents the mean of Huber's loss; To prevent the minimum constant value of the denominator from being 0; Step 4: Adaptive backtracking step size strategy and convergence determination: 4.1) To improve the stability and efficiency of parameter optimization, an adaptive backtracking step size with a safety reduction coefficient is used to update the model parameters. The specific process is as follows: The formula for parameter update is: The iteration parameters of the current model are: The loss function is The loss function has respect to the parameters gradient , These are the candidate iteration parameters after the parameter update. This is the step size for the current iteration, used to control the magnitude of parameter updates; in, For safety reduction factor, This represents the sum of squared gradients, ensuring that the loss function decreases significantly enough when the gradient is large. This is a scaling factor for the step size, used to reduce the step size when the candidate step size does not meet the conditions, ensuring a decrease in the loss function. If the above conditions are not met, then the scaling factor will apply. Decreasing step size: Repeat this process until the step size satisfies the loss function descent condition, thereby avoiding the problem of "the loss function rising due to an excessively large step size" during parameter updates; 4.2) To ensure the stability and effectiveness of the parameter optimization results, a dual convergence criterion is used to determine whether the iteration terminates: First is the relative step size criterion, used to determine whether the parameter update magnitude is small enough. The formula is: in, The parameters are those for two consecutive iterations. Step size threshold, It is the 2-norm, used to measure the length of the parameter vector; To prevent the minimum constant value of the denominator from being 0; The loss variation criterion was also used to determine whether the loss function tends to be stationary. The formula is as follows: in, The loss value is the sum of the losses between two consecutive iterations. For the loss threshold, To prevent the minimum constant of the denominator being 0; The fifth step, after calculating the loss function and adjusting the step size, involves iteratively updating the model parameters using gradient descent combined with an adaptive step size. The specific mechanism is as follows: The formula for parameter update is: in, The optimization parameters are for the k-th iteration. For the first The adaptive backtracking step size ensures that the total loss function satisfies the descent condition after the update through the aforementioned backtracking mechanism; The sixth step involves calculating the remaining lifespan of the wind turbine equipment based on the converged optimal parameters through degradation model extrapolation and failure threshold determination. The specific process is as follows: 6.1) Degradation Model Reconstruction: Using the optimal parameters obtained from parameter optimization, reconstruct the complete equipment degradation model: in, For the degenerate initial state coefficients, For degradation rate coefficient, This is the degraded reference offset. This is the time shift amount. It contains key information such as the amplitude, rate, baseline, and time offset of the degradation curve. Using these parameters, a complete degradation model can be constructed. 6.2) Failure time determination: Based on the failure definition of the main drive chain components of the wind turbine, determine the characteristic value failure threshold. When the degradation characteristic value reaches the threshold, the component is determined to have failed. By extrapolating from the degradation model, the time to first reach the failure threshold, i.e., the failure time, is calculated. Its expression is: The "infimum" represents the minimum time required to satisfy the condition. The time corresponding to the last observation data. The failure threshold is the characteristic value of component degradation. In numerical implementation, it can be done in intervals. The degenerate trajectory corresponding to the discrete-time grid is calculated above. It detects the time point when the threshold is first exceeded, where H is the maximum prediction duration, used to limit the search range of the failure time and avoid unlimited extrapolation.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method for predicting the lifespan of wind turbine main drivetrain components under sparse data conditions. It effectively addresses the drawback of traditional technologies' reliance on complete, continuous data. By adapting time-scale shifting and an exponential degradation model to the sparse and irregular data of wind turbine equipment, it accurately characterizes the degradation trend. Combining a combined loss function and an adaptive step-size strategy reduces interference from anomalous data while constraining the irreversibility of the degradation trend. Furthermore, it reduces computational load through parameter inheritance, adapting to online real-time monitoring requirements. Even without complete full lifecycle data, this method can control the remaining lifespan prediction error within ±25%, accurately determining equipment failure time and improving the lifespan prediction technology system for wind turbine equipment in sparse data scenarios. Attached Figure Description

[0007] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0008] Figure 2 This is a diagram illustrating the predicted degradation process of a wind turbine generator drive end in a wind farm, according to an embodiment of the present invention. Detailed Implementation

[0009] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0010] Reference Figure 1 A method for predicting the lifespan of main drivetrain components of a wind turbine under sparse data conditions includes the following steps: Step 1, Degradation function setting: Based on the nonlinear degradation characteristics of the main drivetrain components of a wind turbine (such as the main bearing and the generator), an exponential degradation model is used to describe its state evolution process. The expression of the exponential degradation model is as follows: in, The equipment is in a degraded state. It is relative to the time of the onset of degradation. These are the parameters corresponding to the degradation process. Uncertainties arising from operating condition disturbances such as speed fluctuations; degenerate function ,in, For the degenerate initial state coefficients, For degradation rate coefficient, This is the degraded baseline offset; The second step is time scale translation: To address the irregular temporal distribution of sparse data, a time shift is introduced to calibrate the time axis of the observed data. The formula is as follows: in, This is the time shift, representing the offset of the observed data throughout the entire degradation trend, used to correct time misalignments caused by missing data; This represents the degradation trend function of the calibrated sparse data, ensuring that different observation points match the degradation model at a uniform time scale. For the first The actual time corresponding to each observation point For the first Degradation state observations at each observation point For the first Uncertainties caused by operating condition disturbances such as constant speed fluctuations; The third step is to set the loss function: To address the constraints on fitting accuracy and monotonicity of degradation trends in sparse data, a combined loss function is designed, as follows: The expression for the loss function is: in, The Huber loss term is used to reduce the interference of outliers on the fitting results, while taking into account both the gradient stability of the mean squared error and the robustness of the absolute error. Its calculation method is as follows: in, The number of samples in the observation data. For the first The actual degradation state observations of each sample For the first Predicted degradation status of each sample The threshold parameter for Huber loss; This is a monotonicity loss term. Since the degradation process of wind turbine equipment has irreversible monotonicity (eigenvalues ​​monotonically increase / decrease over time), this loss term constrains the monotonicity of the prediction results, avoiding the unreasonable situation of "reverse fluctuations in degradation eigenvalues." Its calculation method is as follows: in, For the first The predicted degradation state of each sample, and this loss term is used to constrain the monotonicity of the degradation trend; The monotonicity constraint weights are used to balance the priority between fitting accuracy and monotonicity constraints. The weights of the monotonicity loss are dynamically updated during the iteration process, using the following formula: in, For the first The monotonicity constraint weights for the next iteration; For the first The monotonicity constraint weights for the next iteration; The learning rate used to update the weights; This represents the mean of the monotonicity loss; This represents the mean of Huber's loss; To prevent the minimum constant value of the denominator from being 0; Step 4: Adaptive backtracking step size strategy and convergence determination: 4.1) To improve the stability and efficiency of parameter optimization, an adaptive backtracking step size with a safety reduction coefficient is used to update the model parameters. The specific process is as follows: The formula for parameter update is: The iteration parameters of the current model are: The loss function is The loss function has respect to the parameters gradient , These are the candidate iteration parameters after the parameter update. This is the step size for the current iteration, used to control the magnitude of parameter updates; in, For safety reduction factor, This represents the sum of squared gradients, ensuring that the loss function decreases significantly enough when the gradient is large. This is a scaling factor for the step size, used to reduce the step size when the candidate step size does not meet the conditions, ensuring a decrease in the loss function. If the above conditions are not met, then the scaling factor will apply. Decreasing step size: Repeat this process until the step size satisfies the loss function descent condition, thereby avoiding the problem of "the loss function rising due to an excessively large step size" during parameter updates; 4.2) To ensure the stability and effectiveness of the parameter optimization results, a dual convergence criterion is used to determine whether the iteration terminates: First is the relative step size criterion, used to determine whether the parameter update magnitude is small enough. The formula is: in, The parameters are those for two consecutive iterations. Step size threshold, It is the 2-norm, used to measure the length of the parameter vector; To prevent the minimum constant value of the denominator from being 0; The loss variation criterion was also used to determine whether the loss function tends to be stationary. The formula is as follows: in, The loss value is the sum of the losses between two consecutive iterations. For the loss threshold, To prevent the minimum constant of the denominator being 0; The fifth step, after calculating the loss function and adjusting the step size, involves iteratively updating the model parameters using gradient descent combined with an adaptive step size. The specific mechanism is as follows: The formula for parameter update is: in, The optimization parameters are for the k-th iteration. For the first The adaptive backtracking step size ensures that the total loss function satisfies the descent condition after the update through the aforementioned backtracking mechanism; The sixth step involves calculating the remaining lifespan of the wind turbine equipment based on the converged optimal parameters through degradation model extrapolation and failure threshold determination. The specific process is as follows: 6.1) Degradation Model Reconstruction: Using the optimal parameters obtained from parameter optimization, reconstruct the complete equipment degradation model: in, For the degenerate initial state coefficients, For degradation rate coefficient, This is the degraded reference offset. This is the time shift amount. It contains key information such as the amplitude, rate, baseline, and time offset of the degradation curve. Using these parameters, a complete degradation model can be constructed. 6.2) Failure time determination: Based on the failure definition of the main drive chain components of the wind turbine, determine the characteristic value failure threshold. When the degradation characteristic value reaches the threshold, the component is determined to have failed. By extrapolating from the degradation model, the time to first reach the failure threshold, i.e., the failure time, is calculated. Its expression is: The "infimum" represents the minimum time required to satisfy the condition. The time corresponding to the last observation data. The failure threshold is the characteristic value of component degradation. In numerical implementation, it can be done in intervals. The degenerate trajectory corresponding to the discrete-time grid is calculated above. It detects the time point when the threshold is first exceeded, where H is the maximum prediction duration, used to limit the search range of the failure time and avoid unlimited extrapolation; Through the above steps, the method in this embodiment can accurately fit the degradation process of wind turbine equipment under sparse data conditions and reliably predict its remaining lifespan. The stability and accuracy of the prediction results are significantly improved. Figure 2 As shown, the left subplot is a comparison of the prediction results for the entire lifespan. The solid line in the figure represents the actual remaining lifespan of the main drivetrain components of the wind turbine, while the dashed line represents the predicted remaining lifespan obtained using this method. The light-colored areas correspond to error ranges of ±20% and ±50%, respectively. It can be seen that the prediction curve is always within the ±20% error range, showing a high degree of fit with the actual lifespan. The right subplot is a comparison of the prediction error for the remaining lifespan. The curve in the figure represents the error between the predicted value and the actual value. The light-colored areas correspond to error ranges of ±25% and ±50%. It can be seen that within the current time (160~200 days), the prediction error is stably within the ±25% range, verifying the prediction accuracy of this method.

Claims

1. A method for predicting the lifespan of main drivetrain components of a wind turbine under sparse data conditions, characterized in that, Includes the following steps: Step 1, Degradation function setting: Based on the nonlinear degradation characteristics of the main drivetrain components of a wind turbine (such as the main bearing and the generator), an exponential degradation model is used to describe its state evolution process. The expression of the exponential degradation model is as follows: in, The equipment is in a degraded state. It is relative to the time of the onset of degradation. These are the parameters corresponding to the degradation process. Uncertainties arising from operating condition disturbances such as speed fluctuations; degenerate function ,in, For the degenerate initial state coefficients, For degradation rate coefficient, This is the degraded baseline offset; The second step is time scale translation: To address the irregular temporal distribution of sparse data, a time shift is introduced to calibrate the time axis of the observed data. The formula is as follows: in, This is the time shift, representing the offset of the observed data throughout the entire degradation trend, used to correct time misalignments caused by missing data; This represents the degradation trend function of the calibrated sparse data, ensuring that different observation points match the degradation model at a uniform time scale. For the first The actual time corresponding to each observation point For the first Degradation state observations at each observation point For the first Uncertainties caused by constant speed fluctuations in operating conditions; The third step is to set the loss function: To address the constraints on fitting accuracy and monotonicity of degradation trends in sparse data, a combined loss function is designed, as follows: The expression for the loss function is: in, The Huber loss term is used to reduce the interference of outliers on the fitting results, while taking into account both the gradient stability of the mean squared error and the robustness of the absolute error. Its calculation method is as follows: in, The number of samples in the observation data. For the first The actual degradation state observations of each sample For the first Predicted degradation status of each sample The threshold parameter for Huber loss; This is a monotonicity loss term. Since the degradation process of wind turbine equipment has irreversible monotonicity (eigenvalues ​​monotonically increase / decrease over time), this loss term constrains the monotonicity of the prediction results, avoiding the unreasonable situation of "reverse fluctuations in degradation eigenvalues." Its calculation method is as follows: in, For the first The predicted degradation state of each sample, and this loss term is used to constrain the monotonicity of the degradation trend; The monotonicity constraint weights are used to balance the priority between fitting accuracy and monotonicity constraints. The weights of the monotonicity loss are dynamically updated during the iteration process, using the following formula: in, For the first The monotonicity constraint weights for each iteration; For the first The monotonicity constraint weights for each iteration; The learning rate used to update the weights; This represents the mean of the monotonicity loss; This represents the mean of Huber's loss; To prevent the minimum constant value of the denominator from being 0; Step 4: Adaptive backtracking step size strategy and convergence determination: 4.1) To improve the stability and efficiency of parameter optimization, an adaptive backtracking step size with a safety reduction coefficient is used to update the model parameters. The specific process is as follows: The formula for parameter update is: The iteration parameters of the current model are: The loss function is The loss function has respect to the parameters gradient , These are the candidate iteration parameters after the parameter update. This is the step size for the current iteration, used to control the magnitude of parameter updates; in, For safety reduction factor, This represents the sum of squared gradients, ensuring that the loss function decreases significantly enough when the gradient is large. This is a scaling factor for the step size, used to reduce the step size when the candidate step size does not meet the conditions, ensuring a decrease in the loss function. If the above conditions are not met, then the scaling factor will apply. Decrease step size: Repeat this process until the step size satisfies the loss function descent condition, thereby avoiding the problem of "the loss function rising due to an excessively large step size" during parameter updates; 4.2) To ensure the stability and effectiveness of the parameter optimization results, a dual convergence criterion is used to determine whether the iteration terminates: First is the relative step size criterion, used to determine whether the parameter update magnitude is small enough. The formula is: in, The parameters are those for two consecutive iterations. Step size threshold, It is the 2-norm, used to measure the length of the parameter vector; To prevent the minimum constant value of the denominator from being 0; The loss variation criterion was also used to determine whether the loss function tends to be stationary. The formula is as follows: in, The loss value is the sum of the losses between two consecutive iterations. For the loss threshold, To prevent the minimum constant of the denominator being 0; The fifth step, after calculating the loss function and adjusting the step size, involves iteratively updating the model parameters using gradient descent combined with an adaptive step size. The specific mechanism is as follows: The formula for parameter update is: in, The optimization parameters are for the k-th iteration. For the first The adaptive backtracking step size ensures that the total loss function satisfies the descent condition after the update through the aforementioned backtracking mechanism; The sixth step involves calculating the remaining lifespan of the wind turbine equipment based on the converged optimal parameters through degradation model extrapolation and failure threshold determination. The specific process is as follows: 6.1) Degradation Model Reconstruction: Using the optimal parameters obtained from parameter optimization, reconstruct the complete equipment degradation model: in, For the degenerate initial state coefficients, For degradation rate coefficient, This is the degraded reference offset. This is the time shift amount. It contains key information such as the amplitude, rate, baseline, and time offset of the degradation curve. Using these parameters, a complete degradation model can be constructed. 6.2) Failure time determination: Based on the failure definition of the main drive chain components of the wind turbine, determine the characteristic value failure threshold. When the degradation characteristic value reaches the threshold, the component is determined to have failed. By extrapolating from the degradation model, the time to first reach the failure threshold, i.e., the failure time, is calculated. Its expression is: The "infimum" represents the minimum time required to satisfy the condition. The time corresponding to the last observation data. The failure threshold is the characteristic value of component degradation.

2. The method for predicting the lifespan of wind turbine main drivetrain components under sparse data conditions according to claim 1, characterized in that, In numerical implementation, within the interval The degenerate trajectory corresponding to the discrete-time grid is calculated above. It detects the time point when the threshold is first exceeded, where H is the maximum prediction duration, used to limit the search range of failure time.

Images