A fan blade residual life prediction method based on working condition correction of voiceprint degradation features

Abstract

The application discloses a fan blade residual life prediction method based on working condition correction of voiceprint degradation features, comprising synchronously collecting voiceprint signals and working condition parameters, calculating blade passing frequency based on impeller rotating speed and extracting voiceprint features; generating normal voiceprint reference features under the current working condition through a working condition encoder and a voiceprint benchmark generator, calculating the residual of actual voiceprint and reference features to obtain working condition decoupled voiceprint residual features, and weighting with blade passing frequency and its multiple frequency as prior anchor points to obtain enhanced voiceprint degradation features; constructing health indexes and dynamic degradation trajectories based on the features; screening credible incremental health samples from operation data, performing incremental update of the encoder and the benchmark generator under the constraint of benchmark maintenance, and inhibiting benchmark drift; and fusing health state and working condition parameters into a Transform network to output residual life prediction values. The application can adapt to individual differences and long-term voiceprint distribution changes, and improve prediction accuracy.

Description

Technical Field

[0001] This invention relates to the fields of wind turbine condition monitoring, acoustic signal processing and remaining life prediction, and in particular to a method for predicting the remaining life of wind turbine blades based on the acoustic degradation characteristics of operating conditions correction. Background Technology

[0002] Wind turbine blades operate in environments with strong winds, rain and snow, temperature variations, and alternating loads for extended periods, making them prone to degradation issues such as leading-edge erosion, cracks, delamination, icing, and fatigue damage. Therefore, continuous monitoring and prediction of blade health and remaining lifespan are necessary.

[0003] Existing blade inspection methods mainly include manual inspection, UAV image inspection, vibration monitoring, acoustic emission detection, and life assessment based on SCADA data. However, they still have problems such as long inspection cycles, susceptibility to weather and observation conditions, difficulty in sensor deployment, and difficulty in continuously characterizing the blade life evolution process.

[0004] The acoustic signature of wind turbine blades has advantages such as non-contact acquisition and suitability for continuous monitoring, but its characteristics are easily affected by factors such as wind speed, impeller speed, output power, blade pitch angle, and environmental noise. If the original acoustic signature or the acoustic signature characteristics without operating condition correction are used directly for life prediction, normal operating condition fluctuations can easily be misjudged as blade degradation.

[0005] Existing voiceprint benchmark models are typically trained offline using historical healthy samples and remain fixed after deployment. This makes it difficult to adapt to changes in normal voiceprint distribution caused by individual differences in wind turbines, seasonal variations, changes in background noise, and sensor drift. If the model is updated directly using continuously running data, abnormal or degraded samples may be incorrectly included in healthy samples, causing drift in the healthy voiceprint benchmark.

[0006] Therefore, there is an urgent need for a method to predict the remaining life of wind turbine blades that can generate normal acoustic signature reference features by combining operating conditions and use trusted healthy samples to perform constrained self-learning updates on the acoustic signature benchmark, so as to improve the accuracy and stability of life prediction under complex operating conditions and long-term operation. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method for predicting the remaining life of wind turbine blades based on the degradation characteristics of acoustic signatures corrected for operating conditions. This method addresses the problems that the acoustic signatures of wind turbine blades are easily affected by fluctuations in operating conditions such as wind speed, impeller speed, output power, and pitch angle, as well as the difficulty of adapting fixed healthy acoustic signature benchmarks to individual differences in wind turbines and long-term changes in normal acoustic signature distribution. At the same time, by performing reliable screening and constrained self-learning updates on incremental healthy samples, the drift of the healthy acoustic signature benchmark caused by abnormal or degraded samples is reduced, thereby improving the accuracy and stability of blade degradation feature extraction and remaining life prediction.

[0008] The wind turbine blade remaining life prediction method based on operating condition-corrected acoustic signature degradation characteristics of the present invention includes the following steps:

[0009] Step S1: Synchronously acquire the raw acoustic signature signal x during the operation of the wind turbine blades. t (n) and operating condition parameter vector C t The operating condition parameter vector includes at least wind speed, impeller speed, output power, and blade pitch angle. The original acoustic signature signal is preprocessed, transformed in the time and frequency domains, and features are extracted to obtain time-frequency domain features. The blade passage frequency is calculated based on the impeller speed, and the periodic sweeping sound features corresponding to the blade passage frequency and its harmonics are extracted. The time-frequency domain features are concatenated with the periodic sweeping sound features to obtain the acoustic signature feature vector A. t ;

[0010] Step S2: Construct a working condition voiceprint reference generation network with self-learning and updating capabilities. The working condition voiceprint reference generation network includes a working condition encoder, a voiceprint reference generator, a voiceprint residual extraction module, a sensitive frequency band attention module, a health sample screening module, and a self-learning and updating module.

[0011] The operating condition parameter vector C t Input the operating condition encoder to obtain the operating condition status code Gt, and then input the operating condition status code Gt. t Input the acoustic signature reference generator to generate normal acoustic signature reference features for healthy blades under the current operating conditions. ;

[0012] The voiceprint feature vector A is calculated using the voiceprint residual extraction module. t Compared with normal voiceprint reference features The residuals between them are used to obtain the acoustic residual characteristics R after decoupling under the operating conditions. t ;

[0013] The sensitive frequency band attention module is based on the acoustic residual feature R t Operating condition parameter vector C t And a priori frequency band mask constructed centered on the blade passage frequency and its harmonics is used to generate the frequency band attention weight W. t And utilize the frequency band attention weight W t For the residual features R of the voiceprint t Weighting is performed to obtain the enhanced voiceprint degradation features. ;

[0014] The health sample screening module uses the voiceprint residual feature R t Enhanced voiceprint degradation features The system uses the stability of operating conditions and blade abnormality indicators to screen incremental healthy samples from continuous operating data. When the preset update conditions are met, the self-learning update module uses the incremental healthy samples to incrementally update the adaptive parameters of the operating condition encoder and / or acoustic signature reference generator. The system also uses the reference retention constraint to limit the deviation between the normal acoustic signature reference features generated for historical healthy operating conditions before and after the update. The updated network is used to generate normal acoustic signature reference features for subsequent time windows.

[0015] Step S3, based on the enhanced voiceprint degradation features Multiple degradation characteristics were extracted, and the leaf acoustic health index (HI) was obtained by normalization and weighted fusion. t The blade acoustic health index within a continuous time window is smoothed to obtain a smoothed health index. Based on the smoothed health index within the continuous time window, a dynamic degradation trajectory is constructed and the degradation rate V is calculated. t ;

[0016] Step S4, the smoothness health index and degradation rate V are... t Dynamic degradation trajectory and operating condition parameter vector C t The life prediction input features are fused together to obtain the life prediction input features; the life prediction input features corresponding to multiple consecutive time windows are combined into a life prediction time series input sequence and input into the Transformer life prediction network.

[0017] The Transformer lifetime prediction network introduces a condition modulation term generated from operating parameters at different time windows into its self-attention mechanism. The value vector is then weighted and aggregated according to the self-attention weights of the condition modulation to obtain the temporal degradation state code E. t And according to the time-degradation state encoding E t The predicted remaining life of the output fan blades.

[0018] Further, in step S1, the time-frequency domain features include Mel spectrum features or MFCC features and spectral structure features; the periodic sweeping sound features of the blades include the frequency band energy features near the blade's passing frequency and its harmonics, and the periodic acoustic perturbation features of the blades; the Mel spectrum features or MFCC features, spectral structure features, frequency band energy features, and the periodic acoustic perturbation features of the blades are concatenated to obtain the acoustic feature vector:

[0019]

[0020] Among them, M t Represents the mel spectrum or MFCC characteristics, Q t B represents the spectral structure characteristics. t P represents the frequency band energy characteristics of the blade passing through and its harmonics.t This indicates the periodic acoustic perturbation characteristics of the blade.

[0021] Furthermore, in step S2, the operating condition acoustic signature reference generation network is trained using historical healthy samples during the pre-training phase. These historical healthy samples include the acoustic signature feature vector A of healthy blades. t and its corresponding operating condition parameter vector C t The training loss in the pre-training phase includes the voiceprint reference generation loss. and temporal smoothing loss Total loss during pre-training phase for:

[0022]

[0023] in,

[0024]

[0025]

[0026] Where N represents the number of healthy samples participating in the pre-training. The squared error is represented by t, which represents the time window number. λ represents the normal voiceprint reference feature generated by the voiceprint reference generator based on the operating condition parameters corresponding to the t-th time window, and λ represents the non-negative weighting coefficient of the temporal smoothing loss.

[0027] Furthermore, in step S2, the health sample screening module screens incremental health samples from continuously running data, including:

[0028] Calculate the confidence level of the healthy samples corresponding to the current time window. :

[0029]

[0030] in, This indicates that based on the aforementioned voiceprint residual feature R t The obtained screening acoustic residual intensity, This indicates that based on the enhanced voiceprint degradation characteristics The obtained comprehensive degradation response intensity, Z represents the stability of the operating conditions within the current time window. t The indicators represent abnormal leaf conditions, with α1, α2, α3, and α4 being non-negative weighting coefficients. This is a normalized mapping function that maps the calculation results to a pre-defined confidence interval.

[0031] The acoustic residual intensity and the comprehensive degradation response intensity used for screening are respectively:

[0032]

[0033]

[0034] Where, d R Represents the residual characteristics R of the voiceprint t The dimension, d E This indicates the enhanced voiceprint degradation characteristics. The dimension;

[0035] The stability of the operating conditions is determined based on the fluctuation of the standardized operating condition parameters within the current time window.

[0036]

[0037] Where, d C Indicates the number of operating parameters, s t,r ω represents the standard deviation of the parameter for the r-th standardized operating condition within the current time window. r Let be the non-negative weights corresponding to the parameters of the r-th type of working condition, and satisfy:

[0038]

[0039] The leaf abnormality index Z t Determined based on SCADA fault alarm information, historical anomaly records, and blade inspection results; when any information indicates a blade anomaly within the current time window, let Z... t =1, otherwise let Z =1. t =0.

[0040] When the confidence level of the healthy sample Not less than the preset health confidence threshold τ h And the leaf abnormality index Z t When =0, the voiceprint feature vector A corresponding to the current time window will be... t and operating condition parameter vector C t The samples were identified as incremental healthy samples and stored in the incremental healthy sample database.

[0041] Furthermore, in step S3, the degradation characterization quantities include the intensity of the acoustic residual, the degree of abnormal frequency band shift, the degree of periodic acoustic disturbance, and the degree of time-frequency structure distortion, which are calculated as follows:

[0042] Acoustic residual intensity ,

[0043] Abnormal frequency band shift ,

[0044] Periodic voiceprint perturbation level ,

[0045] Time-frequency structural distortion ;

[0046] Among them, Ω b For the sensitive frequency band range related to blade degradation, To represent the average baseline of the enhanced residual characteristics under healthy conditions at frequency f, f b The frequency of the blade passage, kf b K represents the kth harmonic of the frequency through which the blade passes. b Indicates the harmonic order involved in the calculation. The gradient of the enhanced voiceprint degradation features in the time or frequency direction.

[0047] Furthermore, the leaf acoustic health index mentioned in step S3 Calculated by the following formula:

[0048]

[0049] in, , , , These represent the normalized residual intensity of the voiceprint, the degree of anomalous frequency band shift, the degree of periodic voiceprint perturbation, and the degree of time-frequency structure distortion, respectively; α, β, γ, and δ are non-negative weighting coefficients and satisfy the following conditions: The leaf acoustic health index HI t The higher the value, the greater the degree of degradation of the leaf's acoustic signature;

[0050] The smooth health index The result was obtained through a moving average:

[0051]

[0052] Where L is the length of the sliding time window;

[0053] A dynamic degradation trajectory is constructed based on a smoothed health index over several consecutive time windows:

[0054]

[0055] in, This represents the dynamic degradation trajectory of the blade acoustic signature up to the t-th time window, where m is the number of historical windows;

[0056] The degradation rate is calculated based on the smoothed health index corresponding to adjacent time windows:

[0057]

[0058] Among them, V tΔt represents the voiceprint degradation rate corresponding to the t-th time window, and Δt represents the time interval between adjacent time windows.

[0059] Furthermore, step S3 also includes a degradation monotonic constraint to suppress the abnormal decrease in the smoothness health index of the same wind turbine blade during continuous degradation. The degradation monotonic constraint loss is:

[0060]

[0061] Where M represents the number of continuous time windows participating in the degenerate monotonic constraint. and These represent the smoothed health index corresponding to adjacent time windows.

[0062] Furthermore, the Transformer lifetime prediction network in step S4 also includes an uncertainty assessment module, which evaluates the uncertainty based on the temporal degradation state code E. t Output prediction variance σ t 2 ,

[0063]

[0064] Among them, F u (·) indicates the uncertainty assessment module;

[0065] Then based on the remaining life prediction value and the predicted standard deviation σ t The remaining lifetime prediction range is obtained. , where z represents the interval coefficient corresponding to the preset confidence level.

[0066] Furthermore, the training loss of the Transformer lifetime prediction network in step S4 includes the remaining lifetime prediction loss. Lifetime trend constraint loss and life risk level prediction loss The total loss is:

[0067]

[0068] Where η1 and η2 are weighting coefficients;

[0069] The remaining lifetime prediction loss L rul Gaussian negative log-likelihood loss is used:

[0070]

[0071] in, This represents the predicted remaining lifetime value corresponding to the t-th time window. N is the label for the actual remaining lifetime. R The total number of time windows participating in training, where t is the time window number;

[0072] The lifetime trend constraint loss is used to constrain the overall trend of the remaining lifetime prediction to decrease over time, thus suppressing unexpected increases.

[0073]

[0074] in A non-negative relaxation factor that allows for small fluctuations in the remaining lifetime prediction.

[0075] The accuracy of risk level classification is measured using cross-entropy loss to determine the prediction loss for life risk levels.

[0076]

[0077] Where C represents the number of categories of the preset life risk level. This indicates that the true life risk level corresponding to the t-th time window belongs to category c. This represents the probability of the c-th type of prediction output from the life risk level prediction branch.

[0078] Furthermore, the Transformer lifetime prediction network in step S4 also includes a lifetime risk level prediction branch, which outputs the predicted probability of each lifetime risk level and determines the category corresponding to the maximum predicted probability as the lifetime risk level corresponding to the t-th time window:

[0079]

[0080] Where C represents the number of categories of life risk level, This represents the predicted probability that the t-th time window belongs to the c-th life risk level. This represents the life risk level corresponding to the t-th time window; each life risk level is coded as 1, 2, ..., C in order of increasing risk level.

[0081] Based on the predicted remaining lifespan and life risk level Determine the recommended maintenance time window:

[0082]

[0083] in, This represents the suggested maintenance time window set, τ r τ represents the remaining lifespan warning threshold. y This indicates the risk level warning threshold.

[0084] The beneficial effects of this invention are:

[0085] 1. This invention generates normal acoustic signature reference features of healthy blades based on operating conditions such as wind speed, impeller speed, output power, and blade pitch angle. It also extracts the decoupled acoustic signature degradation features by using the residual between the actual acoustic signature and the reference features. This reduces the interference of normal operating condition fluctuations on acoustic signature analysis and improves the stability and accuracy of degradation feature extraction.

[0086] 2. This invention dynamically calculates the blade passage frequency based on the impeller rotation speed, and uses the blade passage frequency and its harmonics as prior frequency band anchor points to guide the sensitive frequency band attention module to focus on the frequency band related to blade degradation, thereby improving the physical interpretability and noise resistance of acoustic degradation features.

[0087] 3. This invention extracts the residual intensity of the acoustic signature, the degree of abnormal frequency band shift, the degree of periodic acoustic signature disturbance, and the degree of time-frequency structure distortion from the acoustic signature degradation characteristics, and constructs the blade acoustic signature health index, dynamic degradation trajectory, and degradation rate to achieve continuous characterization of the blade degradation process.

[0088] 4. This invention integrates the smoothing health index, degradation rate, dynamic degradation trajectory and operating condition parameters, and introduces a condition modulation mechanism into the life prediction network, enabling the model to adjust the degree of attention to each time window according to different load conditions, thereby improving the accuracy and reliability of remaining life prediction.

[0089] 5. This invention outputs the remaining life prediction value, the remaining life prediction range, the life risk level, and the recommended maintenance time window set, providing a decision-making basis for the formulation of blade inspection, maintenance, and replacement plans.

[0090] 6. This invention filters incremental healthy samples that meet preset confidence conditions from continuous operation data, and incrementally updates the operating condition encoder and / or acoustic signature generator through benchmark preservation constraints, so that the healthy acoustic signature benchmark can adapt to individual differences of wind turbines and long-term normal acoustic signature distribution changes, while reducing benchmark drift caused by abnormal or degraded samples, and improving the stability of remaining life prediction under long-term operating conditions. Attached Figure Description

[0091] Figure 1 A schematic diagram of the network structure for generating and self-learning the voiceprint reference under operating conditions.

[0092] Figure 2 A flowchart for constructing the leaf acoustic health index and generating dynamic degradation trajectory.

[0093] Figure 3 This is a prediction of blade remaining lifetime based on a Transformer lifetime prediction network.

[0094] Figure 4 This is a flowchart of the method for predicting the remaining life of wind turbine blades. Detailed Implementation

[0095] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0096] The wind turbine blade remaining life prediction method based on operating condition correction acoustic signature degradation characteristics in this embodiment includes the following steps:

[0097] Step S1: During wind turbine operation, acoustic sensors deployed outside the nacelle, at the top of the tower, or near the blade sweeping sound propagation path are used to collect the original acoustic signature signal x during wind turbine blade operation. t (n); simultaneously, relying on the wind turbine main control or SCADA system, the operating condition parameter vector C within the same time window is obtained. t The operating condition parameter vector includes at least the wind speed v. t Impeller speed ω t Output power p t and propeller pitch angle θ t .

[0098] 1) Extracting time-frequency domain features

[0099] The original voiceprint signal is preprocessed by framing, filtering, normalization, etc., and each preprocessed frame signal is transformed in the time-frequency domain (specifically, fast Fourier transform can be used). Based on this, time-frequency domain features are extracted, including Mel spectrum features, MFCC features, and spectrum structure features.

[0100] Specifically, the preprocessed voiceprint signal is subjected to frame-by-frame windowing and Fast Fourier Transform to obtain the power spectrum S corresponding to the l-th frame and the r-th frequency point. t (l,r). The power spectrum is input into the Mel filter bank and the logarithm is taken to obtain the Mel spectral characteristics. :

[0101]

[0102] Among them, H m (r) represents the weight of the m-th Mel filter, r represents the discrete frequency index of the spectrum (FFT frequency point number); ε is a minimum positive number to prevent log(0) from occurring in the logarithmic operation; the Mel spectrum features are subjected to discrete cosine transform to obtain MFCC features.

[0103] The spectral structure feature Q t It includes at least one of the following: spectral centroid, spectral bandwidth, spectral kurtosis, and spectral flatness.

[0104] 2) Calculate the blade passage frequency based on the impeller rotation speed, and extract the periodic sweeping sound features of the blades corresponding to the blade passage frequency and its harmonics; concatenate the time-frequency domain features with the periodic sweeping sound features of the blades to obtain the acoustic signature feature vector A. t .

[0105] The blade passes through frequency f b Based on the impeller speed ω t Calculation. In the actual operation of a wind turbine generator, the rotating blades periodically cut the airflow and aerodynamically interfere with surrounding stationary structures (such as the tower, nacelle, or guide vanes), thus generating a periodic component with significant energy in the acoustic signature. The blades transmit the signal at frequency f. b This refers to the number of times the blade tip sweeps past a stationary reference point per unit time, and its value is determined by the impeller speed ω. t The decision is made using the following formula:

[0106]

[0107] Among them, f b N represents the frequency of the blade passage. b Indicates the number of leaves.

[0108] The periodic sweeping sound characteristics of the blades include the frequency band energy characteristics of the blades passing through and their harmonics. , Leaf periodic acoustic perturbation characteristics P t Because the blades generate not only fundamental frequency sweeping noise during rotation, but also, due to blade vibration mode excitation, tip vortex shedding, and multiple reflections and interferences with stationary structures, noise at its harmonic frequency (2f) is also generated. b ,3f b Energy accumulates at locations such as f. During normal wind turbine operation and early stages of minor blade structural degradation (e.g., slight coating peeling, slight leading edge wear), the acoustic spectrum at f... b It exhibits stable energy peak characteristics near its first few harmonics. Therefore, this embodiment selects f b Using its harmonics as frequency band anchors for acoustic fingerprint feature extraction, it can effectively capture the acoustic fingerprint dominated by the periodic motion of the blades, eliminate broadband interference factors such as wind speed changes and turbulent background noise, thereby improving the robustness of subsequent degradation feature extraction.

[0109] With the blade passing frequency f b Centered on its harmonics, calculate the frequency band energy characteristics within a preset half-bandwidth Δf. :

[0110]

[0111] Where k represents the harmonic order; This represents the average power spectrum of each frame within the t-th time window at the r-th frequency point; based on the normalized autocorrelation coefficient of the acoustic signature signal at the blade passage period and its integer multiples of the delay position. The periodic acoustic perturbation characteristics P of the blade were obtained. t P t Solving for:

[0112] First, based on the blade's passing frequency f b Determine the sampling point delay corresponding to the blade's passage period:

[0113]

[0114] Among them, f s τ represents the sampling frequency of the voiceprint signal. b This represents the number of sampling points corresponding to the cycle of a single blade.

[0115] The normalized autocorrelation coefficient of the voiceprint signal at delay τ within the t-th time window is:

[0116]

[0117] Where, N t This represents the number of sampling points within a time window of t. This represents the average value of the voiceprint signal within that time window.

[0118] The periodic acoustic signature perturbation characteristics of the blade are composed of the normalized autocorrelation coefficients at the blade's passage period and its integer multiples of the delay position.

[0119]

[0120] Among them, K p This indicates the lag order for participation in periodic analysis.

[0121] 3) Concatenate the time-frequency domain features with the periodic sweeping sound features of the blades to obtain the acoustic signature feature vector A. t :

[0122]

[0123] Among them, M t Represents the mel spectrum or MFCC characteristics, Q t B represents the spectral structure characteristics. t P represents the frequency band energy characteristics of the blade passing through and its harmonics. t This indicates the periodic acoustic perturbation characteristics of the blade.

[0124] 4) Convert the voiceprint feature vector A for each time window t. t With synchronous operating condition parameter vector Pairing up to form a joint sample set:

[0125]

[0126] Where D represents the joint sample set and T represents the number of samples. This step provides basic data support for subsequent operating condition decoupling, voiceprint degradation feature extraction, and remaining lifetime prediction.

[0127] Step S2: Construct a working condition voiceprint reference generation network with self-learning and updating capabilities. The working condition voiceprint reference generation network includes a working condition encoder, a voiceprint reference generator, a voiceprint residual extraction module, a sensitive frequency band attention module, a health sample screening module, and a self-learning and updating module.

[0128] In this embodiment, the working condition encoder adopts a two-layer fully connected neural network.

[0129] First, the operating condition parameter vector C... t Standardization is performed to obtain a standardized operating condition parameter vector:

[0130]

[0131] Where, μ C and σ C ε represents the mean and standard deviation of each operating condition parameter in the historical health sample, respectively. C To prevent positive numbers with a denominator of zero.

[0132] Standardized operating condition parameter vector The first fully connected layer is input, and the output is a 32-dimensional hidden feature. After ReLU activation, it is input into the second fully connected layer to obtain a 16-dimensional working condition code.

[0133]

[0134] Among them, F c (·) indicates the operating condition encoder. The operating condition code is used to characterize the normal variation pattern of blade acoustic signature under different operating conditions.

[0135] The operating condition state is encoded as G. t Input the acoustic signature reference generator to generate normal acoustic signature reference features for healthy blades under the current operating conditions. :

[0136]

[0137] Among them, F g (·) indicates the voiceprint reference generator, which is used to generate normal voiceprint reference features under the current operating condition based on the operating condition encoding. This represents the normal voiceprint reference features generated under the current operating conditions. In this embodiment, the voiceprint reference generator uses a three-layer fully connected neural network, where the output dimensions of the two hidden layers are 64 and 128 respectively, and ReLU activation is used; the output layer uses linear activation, and the output is the same as the voiceprint feature vector A. t Normal voiceprint reference features with consistent dimensions and feature order The normal acoustic signature reference features are used to characterize the acoustic signature baseline state that a healthy blade should have under current operating conditions.

[0138] The voiceprint feature vector A is calculated using the voiceprint residual extraction module. t Compared with the normal voiceprint reference features The residuals are used to obtain the acoustic signature residual characteristics after decoupling under the operating conditions:

[0139]

[0140] Among them, R t This represents the acoustic signature residual characteristics within the t-th time window; these residual characteristics characterize the deviation of the actual blade acoustic signature from the healthy acoustic signature baseline under current operating conditions. Through residual calculation, the influence of normal operating condition variations such as wind speed, impeller speed, output power, and blade pitch angle on the acoustic signature characteristics can be reduced, allowing the acoustic signature residual characteristics to more effectively reflect the abnormal acoustic response caused by blade structural degradation.

[0141] The sensitive frequency band attention module is based on the voiceprint residual feature R t Operating condition parameter vector C t and the blade passing frequency f b The frequency band attention weight W is generated by using a predefined prior frequency band mask centered on its frequency harmonics and its multiples. t In this embodiment, the frequency band attention weight W t The generation process is as follows:

[0142] First, select the frequency f through which the blade passes. b and its 2nd harmonic 2f b and 3 times the frequency 3f b The prior frequency band anchor point set {f} is obtained. b ,2f b ,3f b}, and then predefine an a priori frequency band mask centered on the a priori frequency band anchor point. The a priori frequency band mask The length of the acoustic residual feature R t The number of frequency bands is the same, and a first preset value is assigned at the corresponding frequency band position of each anchor point, and a second preset value is assigned at other positions; or, the prior frequency band mask A Gaussian window or triangular window function centered at each anchor point is predefined, and the window width covers a predetermined frequency band range near the anchor point.

[0143] Subsequently, a subnetwork F is generated using trainable sensitive frequency band weights. w (·) Based on the aforementioned voiceprint residual feature R t and operating condition parameter vector C t Generate trainable frequency band weight components and combine these components with the prior frequency band mask. We perform weighted fusion to obtain the frequency band attention weights:

[0144]

[0145] In this embodiment, the trainable sensitive frequency band weight generation subnetwork F w (·) Specifically, a two-layer one-dimensional convolutional neural network is used to implement the voiceprint residual feature R. t The operating condition parameter vector C, after being fully connected and extended to the same frequency band length. t Channel concatenation is performed and used as input to the two-layer one-dimensional convolutional neural network. In the two-layer one-dimensional convolutional neural network: the first layer uses 32 convolutional kernels of length 5 with a stride of 1, followed by batch normalization and ReLU activation; the second layer uses a single convolutional kernel of length 3 with a stride of 1, and outputs trainable frequency band weights with the same frequency band length as the voiceprint residual features. Both convolutional layers use zero padding to maintain the frequency band dimension.

[0146] The prior frequency band mask weight coefficient μ is used to adjust the influence of the blade passage frequency and its harmonic prior on the frequency band attention weight. Its value ranges from 0.1 to 2, preferably 0.5; the normalized activation function Φ w (·) The Sigmoid function is used to map the output to the (0,1) interval.

[0147] Bandwidth attention weight W t With the aforementioned voiceprint residual feature R t Weighted analysis yields the enhanced voiceprint degradation features:

[0148]

[0149] in, This indicates the enhanced voiceprint degradation characteristics. This indicates element-wise multiplication. Through the above processing, the abnormal acoustic response near the blade's passage frequency and its harmonics can be enhanced, while reducing the interference of irrelevant environmental noise and non-critical frequency bands on degradation characteristics.

[0150] To enable the operating condition encoder and acoustic signature reference generator to establish a mapping relationship between operating conditions and normal acoustic signatures of healthy blades, and to adapt to changes in the distribution of normal acoustic signatures during long-term operation, this embodiment divides the operation process of the operating condition acoustic signature reference generation network into a pre-training stage, a real-time inference stage, and a self-learning update stage. The pre-training stage uses historical healthy samples for training, the real-time inference stage extracts acoustic signature degradation features, constructs health indices, and selects incremental healthy samples, and the self-learning update stage uses the selected incremental healthy samples to incrementally update the network parameters under the constraint of maintaining the reference.

[0151] (1) Pre-training phase

[0152] In this embodiment, the operating condition acoustic signature reference generation network is trained in the pre-training stage using a set of historical healthy samples selected from the historical healthy sample library. Specifically, the pre-training training dataset includes sample pairs consisting of acoustic signature feature vectors collected when the blades are in a healthy state and their corresponding operating condition parameter vectors.

[0153] Pre-training loss includes speaker benchmark generation loss. and temporal smoothing loss Total loss during pre-training phase for:

[0154]

[0155] Wherein, λ is the weighting coefficient; in this embodiment, the weighting coefficient λ of the temporal smoothing loss is determined based on the voiceprint reference generation error on the validation set and the fluctuation degree of the reference features in adjacent time windows, and its value range can be 0.01 to 1, preferably 0.1. When the generated normal voiceprint reference features fluctuate greatly between adjacent time windows, λ can be appropriately increased; when the voiceprint reference generation error is large, λ can be appropriately decreased.

[0156]

[0157]

[0158] Where N represents the number of healthy samples, denoted by , where t represents the squared error and t represents the time window number.

[0159] (2) Real-time reasoning stage

[0160] During the real-time inference phase, the operating condition voiceprint reference generation network receives the real-time acquired voiceprint feature vector A. t and operating condition parameter vector C t The signal passes through a working condition encoder, a voiceprint reference generator, a voiceprint residual extraction module, and a sensitive frequency band attention module in sequence to obtain the voiceprint residual feature R. tand enhanced voiceprint degradation features .

[0161] The enhanced voiceprint degradation feature On one hand, it is passed to step S3 for constructing the blade acoustic signature health index and dynamic degradation trajectory; on the other hand, it is combined with the acoustic signature residual feature R. t The stability of the operating conditions and the abnormal state of the blades are input into the healthy sample screening module to determine whether the sample corresponding to the current time window can be used as an incremental healthy sample.

[0162] Incremental health sample screening:

[0163] The health sample screening module is based on the voiceprint residual feature R t Enhanced voiceprint degradation features The stability of operating conditions and indicators of abnormal blade status are used to screen incremental healthy samples from continuous operating data, including:

[0164] Calculate the confidence level of the healthy samples corresponding to the current time window. :

[0165]

[0166] in, This indicates that based on the aforementioned voiceprint residual feature R t The obtained acoustic residual intensity, This indicates the enhanced voiceprint degradation features. The obtained degradation response intensity, Z represents the stability of the operating conditions within the current time window. t This represents the blade abnormality status index determined based on blade fault alarm information and historical anomaly records. α1, α2, α3, and α4 are non-negative weighting coefficients. To map the calculation results to a pre-set confidence value range, a normalization mapping function is used; in this embodiment, the Sigmoid function is used.

[0167] The screening uses residual acoustic signature intensity for:

[0168]

[0169] Where, d R Represents the residual characteristics R of the voiceprint t Dimensions.

[0170] The degradation response intensity for:

[0171]

[0172] Where, d E This indicates the enhanced voiceprint degradation characteristics. The dimension. The stated and Used for incremental healthy sample screening, but not as a degradation characterization metric for constructing the leaf acoustic health index in step S3.

[0173] The wind speed, impeller speed, output power, and blade pitch angle within the current time window are standardized, and the standard deviation s of each standardized operating condition parameter is calculated. t,r The stability of the operating conditions is as follows:

[0174]

[0175] Where, d C ω represents the number of operating parameters. r Let be the non-negative weights corresponding to the parameters of the r-th type of working condition, and satisfy:

[0176]

[0177] The leaf abnormality index Z t Determined based on SCADA fault alarm information, historical anomaly records, and blade inspection results; when any information indicates a blade anomaly within the current time window, let Z... t =1, otherwise let Z =1. t =0.

[0178] The weighting coefficients α1, α2, α3, and α4 satisfy the following:

[0179]

[0180] Each weight can be determined based on the false acceptance rate and false rejection rate of healthy samples on the validation set. In one embodiment, α1, α2, α3, and α4 are set to 0.30, 0.30, 0.25, and 0.15, respectively; in other embodiments, they can be adjusted based on the false acceptance rate and false rejection rate of healthy samples on the validation set.

[0181] When the confidence level of the healthy sample Not less than the preset health confidence threshold And the leaf abnormality index Z t When =0, the voiceprint feature vector A corresponding to the current time window will be... t and operating condition parameter vector C t The samples were identified as incremental healthy samples and stored in the incremental healthy sample database.

[0182] (3) Self-learning and updating stage

[0183] When the number of samples in the incremental health sample library reaches the preset update quantity or the preset update cycle, the self-learning update module selects an incremental health sample set from the incremental health sample library:

[0184]

[0185] And select a set of historical health samples from the historical health sample database:

[0186]

[0187] Where, N u This indicates the number of incremental healthy samples participating in this incremental update. and N represents the voiceprint feature vector and the operating condition parameter vector corresponding to the i-th incremental health sample, respectively; h This indicates the number of historical healthy samples that participated in maintaining the baseline constraint. These represent the operating condition parameter vectors corresponding to the j-th historical healthy sample.

[0188] The self-learning update module incrementally updates the adaptive parameters of the operating condition encoder and / or voiceprint reference generator based on the incremental health sample set and the historical health operating condition sample set.

[0189] In this embodiment, the adaptive parameters include the weight parameters and bias parameters of the second fully connected layer of the working condition encoder and the last hidden layer and output layer of the voiceprint reference generator; the parameters of the first fully connected layer of the working condition encoder and the first hidden layer of the voiceprint reference generator are kept frozen during the incremental update process to reduce the overall model drift caused by a small number of incremental samples.

[0190] The incremental update uses the Adam optimization algorithm, with an initial learning rate set to 1×10⁻⁶. −4 Batch size is set to 16–64, with incremental updates for 5–20 training rounds; when the self-learning loss decreases by less than 10 for three consecutive rounds. −4 The incremental update will stop when the preset number of training rounds is reached. After the incremental update is completed, the updated network will only be used for the generation of normal voiceprint reference features in subsequent time windows.

[0191] The incremental update uses a self-learning loss function. for:

[0192]

[0193] in, This represents the incremental voiceprint reference generation loss. Indicates the baseline holding loss. The non-negative weighting coefficient represents the baseline retention loss.

[0194] The incremental acoustic signature generation loss for:

[0195]

[0196] The benchmark retention loss for:

[0197]

[0198] in, and These represent the operating condition encoder and the voiceprint reference generator during the incremental update process, respectively. and These represent the encoder and voiceprint reference generator under operating conditions before the incremental update, respectively.

[0199] The incremental voiceprint baseline generation loss is used to adapt the updated network to the newly screened healthy voiceprint-work condition samples; the baseline preservation loss is used to limit the deviation between the normal voiceprint reference features generated before and after the incremental update for the same historical healthy work condition, so as to suppress catastrophic forgetting and the drift of the healthy voiceprint baseline to an abnormal or degenerate state.

[0200] The incrementally updated operating condition encoder and acoustic signature reference generator are used for generating normal acoustic signature reference features for subsequent time windows. The acoustic signature residual feature R corresponding to the current time window... t The enhanced voiceprint degradation features are still calculated by the network before the incremental update, so as to avoid the voiceprint residuals of the current samples being artificially reduced after participating in the network update.

[0201] Through the aforementioned pre-training, real-time inference, healthy sample screening, and self-learning update processes, the operating condition acoustic signature reference generation network can generate normal acoustic signature reference features for healthy blades based on the current operating conditions. It also utilizes reliable incremental healthy samples to adapt to individual differences in wind turbines, seasonal variations, background noise changes, and sensor drift. Simultaneously, it reduces the risk of healthy acoustic signature drift through signature maintenance constraints, providing stable acoustic signature degradation features for the construction of the blade acoustic signature health index and the generation of dynamic degradation trajectories in step S3. .

[0202] Step S3, based on the enhanced voiceprint degradation features obtained in step S2 Four degradation characteristics were extracted: residual intensity of the acoustic signature, degree of anomalous frequency band shift, degree of periodic acoustic signature perturbation, and degree of time-frequency structure distortion. Their calculations are as follows:

[0203] Acoustic residual intensity ,

[0204] Abnormal frequency band shift ,

[0205] Periodic voiceprint perturbation level ,

[0206] Time-frequency structural distortion ;

[0207] Among them, Ω b For the sensitive frequency band range related to blade degradation, The average residual reference at the corresponding frequency position under healthy conditions, where f is the frequency and kf is the reference value. b K represents the kth harmonic of the frequency through which the blade passes. b Indicates the harmonic order involved in the calculation. This represents the gradient of voiceprint degradation features in the time or frequency direction.

[0208] The four degradation characteristics obtained were normalized and then weighted and fused to construct the leaf acoustic health index HI. t :

[0209]

[0210] in, , , , These represent the normalized residual intensity of the voiceprint, the degree of anomalous frequency band shift, the degree of periodic voiceprint perturbation, and the degree of time-frequency structure distortion, respectively; α, β, γ, and δ are weighting coefficients and satisfy the following conditions: The weighting coefficients can be adjusted according to the specific degradation type: for aerodynamic acoustic anomalies such as leading edge erosion and icing, the weights of anomalous frequency band shift and periodic disturbances are increased; for structural acoustic anomalies such as cracks and fatigue, the weights of residual intensity and time-frequency distortion are increased, thereby taking into account the acoustic differences of different degradation mechanisms.

[0211] To reduce the impact of short-term wind noise, rain noise, sudden environmental noise, and acquisition errors on the health index, the blade acoustic health index within a continuous time window is smoothed to obtain a smoothed health index:

[0212]

[0213] in, Let L represent the smoothed health index corresponding to the t-th time window, where L is the length of the sliding time window. The smoothed health index can more stably reflect the blade degradation trend and avoid misjudgment caused by accidental acoustic disturbances within a single time window.

[0214] Construct a dynamic degradation trajectory based on a smoothed health index within a continuous time window. :

[0215]

[0216] in, This represents the dynamic degradation trajectory of the blade acoustic signature up to the t-th time window, where m is the number of historical windows. This degradation trajectory describes the change in blade health status over time, providing time-series input for subsequent remaining lifetime prediction.

[0217] To further characterize the blade degradation rate, the degradation rate was calculated based on the smoothness health index of adjacent time windows:

[0218]

[0219] Among them, V t Let V represent the voiceprint degradation rate corresponding to the t-th time window, and Δt represent the time interval between adjacent time windows. t When the sound signature continues to increase, it indicates that the degree of blade acoustic signature degradation is accelerating, which may correspond to the expansion of blade defects, deterioration of aerodynamic shape, or aggravation of structural fatigue damage.

[0220] As a further improvement to this step, considering that the leaf degradation process is usually continuous and cumulative, the health index should, in general, gradually increase or remain at a high-risk level as the degree of degradation increases. Therefore, this step also includes applying a degradation monotonic constraint to the smoothed health index, which is expressed as:

[0221]

[0222] Where M represents the number of time windows involved in the constraint. A degradation monotonic constraint is introduced into the process of health index change. This constraint does not require the health index to rise strictly at every moment, but is used to reduce unreasonable sharp declines in the long-term degradation trend, so that the constructed degradation trajectory is more consistent with the law of leaf life degradation.

[0223] Finally, this step can represent the health status of the leaf acoustic signature as follows:

[0224]

[0225] Among them, S t This represents the leaf acoustic signature health status corresponding to the t-th time window. V represents the smoothed health index. t Indicates the rate of voiceprint degradation. This indicates the dynamic degradation trajectory of the blade's acoustic signature.

[0226] Through the above processing, this step decouples the voiceprint degradation characteristics after the operating conditions. This process transforms the acoustic signature into a continuously trackable blade health index and dynamic degradation trajectory. It not only reflects the presence of abnormal acoustic responses in the blades but also characterizes the changing trend of blade health status over operating time and accumulated load. Compared to traditional acoustic signature fault classification methods, this step further transforms the acoustic signature signal from a basis for instantaneous state identification into a basis for characterizing the lifespan degradation process, providing a continuous, stable, and physically meaningful input for subsequent prediction of the remaining lifespan of wind turbine blades.

[0227] Step S4: In order to simultaneously consider the current degradation state of the blade and the influence of external operating loads, the smoothing health index is... Degradation rate V t Dynamic degradation trajectory With the operating condition parameter vector C t The features are then fused to obtain the lifetime prediction input features.

[0228]

[0229] Among them, U t S represents the lifetime prediction input feature corresponding to the t-th time window. t C represents the health status of leaf acoustic signature. t This indicates the parameters for synchronous operation.

[0230] Input the lifetime prediction feature U t The time series input sequence for lifetime prediction is as follows:

[0231]

[0232] in, This represents the lifetime prediction time series input sequence up to the t-th time window, where m represents the length of the historical window involved in the remaining lifetime prediction. This lifetime prediction time series input sequence can simultaneously reflect the changing trend of blade acoustic signature health status, the change in degradation rate, and the impact of operating loads on blade lifetime degradation.

[0233] Input lifetime prediction time series The input is fed into a Transformer lifetime prediction network, which includes a self-attention mechanism and introduces a condition modulation term ρ(C) into the self-attention mechanism. i C j This is used to adjust the attention score based on the similarity of operating conditions between different time windows.

[0234] The Transformer lifetime prediction network inputs a lifetime prediction time series. These are mapped to query vector, key vector, and value vector, respectively:

[0235]

[0236] Among them, F Q (·), F K (·), F V (·) represent the mapping functions for the query vector, key vector, and value vector, respectively. In this embodiment, the mapping functions are implemented using independent linear transformations, and for the lifetime prediction input feature U corresponding to the i-th time window... i ,have:

[0237]

[0238]

[0239]

[0240] Among them, U i W represents the lifetime prediction input feature corresponding to the i-th time window. Q W K W V and b Q b K b V These represent the linear mapping parameters corresponding to the query vector, key vector, and value vector, respectively.

[0241] In this embodiment, the feature dimension d of the query vector, key vector, and value vector are all set to 64; the feature dimension can also be selected in the range of 32 to 128 according to the size of the lifetime prediction input features.

[0242] First, the operating condition parameter vectors for the i-th and j-th time windows are standardized to obtain... and The operating condition modulation term is defined as follows:

[0243]

[0244] Among them, W c γ is a trainable condition mapping matrix. c It is a non-negative modulation coefficient.

[0245] The operating condition modulation term is superimposed on the scaled dot product of the query vector and the key vector to obtain the operating condition modulation attention score between the i-th time window and the j-th time window:

[0246]

[0247] Where, d kThe dimension of the Key vector in the attention mechanism is used to scale the dot product result, avoiding excessively large values ​​that could lead to gradient instability in high-dimensional cases. The condition-modulated attention score is normalized using the Softmax function to obtain the condition-modulated self-attention weights.

[0248]

[0249] Among them, A ij represents the attention weight of the i-th time window to the j-th time window, and m represents the number of time windows in the lifetime prediction time series input sequence.

[0250] When the operating conditions of two time windows are closer, the absolute value of the operating condition modulation term is smaller, and the suppression of the original attention score is weaker; when the operating conditions of two time windows are more different, the absolute value of the negative value of the operating condition modulation term is larger, thereby reducing the attention correlation between different operating condition time windows.

[0251] The value vector is weighted and aggregated based on the self-attention weights of the operating condition modulation to obtain the temporal degradation state code E. i :

[0252]

[0253] Among them, E i V represents the temporal degradation state code corresponding to the i-th time window. j Let E represent the value vector corresponding to the j-th time window. When the i-th time window is the t-th time window to be predicted, let i = t, and the corresponding temporal degradation state code is denoted as E. t This encoding simultaneously contains information on the evolution of blade acoustic signature health and the impact of operating load conditions on the degradation process.

[0254] The Transformer lifetime prediction network also includes a lifetime prediction head and an uncertainty assessment module. The lifetime prediction head F... p (·) and uncertainty assessment module F u (·) A shared hidden layer and two independent output branches are used. The temporal degradation state encoding E t Input a fully connected hidden layer with 64 neurons and apply ReLU activation to obtain hidden features:

[0255]

[0256] Where h t W represents the hidden feature vector of the t-th time window. h Let b be the hidden layer weight matrix. hThe hidden layer is biased; the lifetime prediction branch uses a linear output layer to obtain the predicted remaining lifetime of the wind turbine blades corresponding to the t-th time window:

[0257]

[0258] in, W represents the predicted remaining life of the wind turbine blades corresponding to the t-th time window. p Let b be the output layer weight matrix. p For the output layer bias term, and:

[0259]

[0260] The uncertainty assessment branch outputs the prediction variance:

[0261]

[0262] in, W represents the variance of the remaining lifetime prediction corresponding to the t-th time window. The Softplus activation function is used to ensure that the prediction variance is positive. u Let b be the weight matrix of the uncertain branch. u For the uncertainty branch bias term, ε u To prevent positive numbers with zero prediction variance, each weight matrix is ​​obtained through backpropagation training, and:

[0263]

[0264] Therefore, based on the remaining useful life prediction and the predicted standard deviation σ t The remaining lifetime prediction range is obtained as follows:

[0265]

[0266] Here, z represents the confidence coefficient. In this way, the Transformer lifetime prediction network can not only output a single remaining lifetime prediction value, but also provide the confidence range of the prediction results, making it easier for maintenance personnel to formulate inspection and maintenance plans based on prediction uncertainties.

[0267] In a further improvement, the lifetime prediction head in step S4 also outputs a lifetime risk level, and based on the remaining lifetime prediction value... The recommended maintenance time window is determined based on the aforementioned lifespan risk level.

[0268]

[0269] in, Indicates the suggested maintenance time window, τ r τ represents the remaining lifespan warning threshold.y Indicates the risk level warning threshold. This indicates the life risk level output by the life prediction head. By predicting the probability of each risk category The category corresponding to the maximum value is obtained. When the predicted remaining lifespan is lower than the set threshold, or the lifespan risk level reaches a high-risk state, the system outputs maintenance warning information to assist the wind farm in formulating blade inspection, maintenance, or shutdown plans.

[0270] Considering the cumulative and irreversible nature of blade life degradation, the Transformer life prediction network in this embodiment introduces a life trend constraint during training, ensuring that the prediction results generally conform to the law that the remaining life gradually decreases with operating time and load accumulation. The training loss of the Transformer life prediction network includes the remaining life prediction loss. Lifetime trend constraint loss and life risk level prediction loss The total loss is:

[0271]

[0272] Wherein, η1 and η2 are weighting coefficients, and in this embodiment, η1=0.1 and η2=0.05.

[0273] In this embodiment, the remaining lifetime prediction loss L rul Gaussian negative log-likelihood loss is used:

[0274]

[0275] in, This represents the predicted remaining lifetime value corresponding to the t-th time window. N is the label for the actual remaining lifetime. R The total number of time windows for training is represented by t, where t is the time window number.

[0276] In this embodiment, the lifetime trend constraint loss Used to constrain the overall trend of the remaining useful life forecast to decrease with operating time, and to suppress unexpected increases:

[0277]

[0278] in A non-negative relaxation factor that allows for small fluctuations in the remaining lifetime prediction.

[0279] In this embodiment, the life risk level predicts the loss. Cross-entropy loss is used to measure the accuracy of risk level classification:

[0280]

[0281] Where C represents the number of categories of the preset life risk level. This indicates that the true life risk level corresponding to the t-th time window belongs to category c. This represents the probability of the c-th type of prediction output from the life risk level prediction branch.

[0282] In this embodiment, time-series samples with remaining life labels and life risk level labels are divided into training, validation, and test sets according to wind turbine blade objects or operating time periods, so that overlapping time windows of the same continuous operating segment are not simultaneously assigned to different datasets. The training set is used to optimize the Transformer life prediction network, and the validation set is used to determine the model parameters and the timing for stopping training.

[0283] The network is trained using the Adam optimization algorithm, with an initial learning rate set to 1×10⁻⁶. -4 The batch size is set to 32, and the maximum number of training epochs is set to 100. Training stops when the total loss of the validation set does not decrease for 10 consecutive epochs, and the model parameters corresponding to the lowest total loss of the validation set are saved. A Dropout ratio of 0.1 can be used during training to reduce the risk of overfitting.

[0284] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining life of wind turbine blades based on the degradation characteristics of acoustic signatures under operating conditions, characterized in that: Includes the following steps: Step S1: Synchronously acquire the raw acoustic signature signal x during the operation of the wind turbine blades. t (n) and operating condition parameter vector C t The operating condition parameter vector includes at least wind speed, impeller speed, output power, and blade pitch angle; the original acoustic signature signal is preprocessed, transformed in the time and frequency domains, and features are extracted to obtain time and frequency domain features; The blade passage frequency is calculated based on the impeller rotation speed, and the periodic sweeping sound features corresponding to the blade passage frequency and its harmonics are extracted. The time-frequency domain features are concatenated with the periodic sweeping sound features of the blades to obtain the acoustic signature feature vector A. t ; Step S2: Construct a working condition voiceprint reference generation network with self-learning and updating capabilities. The working condition voiceprint reference generation network includes a working condition encoder, a voiceprint reference generator, a voiceprint residual extraction module, a sensitive frequency band attention module, a health sample screening module, and a self-learning and updating module. The operating condition parameter vector C t Input the operating condition encoder to obtain the operating condition status code Gt, and then input the operating condition status code Gt. t Input the acoustic signature reference generator to generate normal acoustic signature reference features for healthy blades under the current operating conditions. ; The voiceprint feature vector A is calculated using the voiceprint residual extraction module. t Compared with normal voiceprint reference features The residuals between them are used to obtain the acoustic residual characteristics R after decoupling under the operating conditions. t ; The sensitive frequency band attention module is based on the acoustic residual feature R t Operating condition parameter vector C t And a priori frequency band mask constructed centered on the blade passage frequency and its harmonics is used to generate the frequency band attention weight W. t And utilize the frequency band attention weight W t For the residual features R of the voiceprint t Weighting is performed to obtain the enhanced voiceprint degradation features. ; The health sample screening module uses the voiceprint residual feature R t Enhanced voiceprint degradation features The system uses the stability of operating conditions and blade abnormality indicators to screen incremental healthy samples from continuous operating data. When the preset update conditions are met, the self-learning update module uses the incremental healthy samples to incrementally update the adaptive parameters of the operating condition encoder and / or acoustic signature reference generator. The system also uses the reference retention constraint to limit the deviation between the normal acoustic signature reference features generated for historical healthy operating conditions before and after the update. The updated network is used to generate normal acoustic signature reference features for subsequent time windows. Step S3, based on the enhanced voiceprint degradation features Multiple degradation characteristics were extracted, and the leaf acoustic health index (HI) was obtained by normalization and weighted fusion. t The blade acoustic health index within a continuous time window is smoothed to obtain a smoothed health index. Based on the smoothed health index within the continuous time window, a dynamic degradation trajectory is constructed and the degradation rate V is calculated. t ; Step S4, the smoothness health index and degradation rate V are... t Dynamic degradation trajectory and operating condition parameter vector C t The features are then fused to obtain the lifetime prediction input features. The lifetime prediction input features corresponding to multiple consecutive time windows are combined into a lifetime prediction time series input sequence and then input into the Transformer lifetime prediction network. The Transformer lifetime prediction network introduces a condition modulation term generated from operating parameters at different time windows into its self-attention mechanism. The value vector is then weighted and aggregated according to the self-attention weights of the condition modulation to obtain the temporal degradation state code E. t And according to the time-degradation state encoding E t The predicted remaining life of the output fan blades.

2. The method for predicting the remaining life of wind turbine blades based on the acoustic degradation characteristics of operating conditions as described in claim 1, characterized in that: In step S1, the time-frequency domain features include Mel spectrum features or MFCC features and spectral structure features; the periodic sweeping sound features of the blades include the frequency band energy features near the blade's passing frequency and its harmonics, and the periodic acoustic perturbation features of the blades; the Mel spectrum features or MFCC features, spectral structure features, frequency band energy features, and the periodic acoustic perturbation features of the blades are concatenated to obtain the acoustic feature vector: Among them, M t Represents the mel spectrum or MFCC characteristics, Q t B represents the spectral structure characteristics. t P represents the frequency band energy characteristics of the blade's passing frequency and its harmonics. t This indicates the periodic acoustic perturbation characteristics of the blade.

3. The method for predicting the remaining life of wind turbine blades based on the acoustic degradation characteristics of operating conditions as described in claim 1, characterized in that: In step S2, the operating condition acoustic signature reference generation network is trained using historical healthy samples during the pre-training phase. These historical healthy samples include the acoustic signature feature vector A of healthy blades. t and its corresponding operating condition parameter vector C t The training loss in the pre-training phase includes the voiceprint reference generation loss. and temporal smoothing loss Total loss during pre-training phase for: in, Where N represents the number of healthy samples participating in the pre-training. The squared error is represented by t, which represents the time window number. λ represents the normal voiceprint reference feature generated by the voiceprint reference generator based on the operating condition parameters corresponding to the t-th time window, and λ represents the non-negative weighting coefficient of the temporal smoothing loss.

4. The method for predicting the remaining life of wind turbine blades based on the acoustic degradation characteristics of operating conditions as described in claim 1, characterized in that: In step S2, the health sample screening module screens incremental health samples from continuously running data, including: Calculate the confidence level of the healthy samples corresponding to the current time window. : in, This indicates that based on the aforementioned voiceprint residual feature R t The obtained screening acoustic residual intensity, This indicates that based on the enhanced voiceprint degradation characteristics The obtained comprehensive degradation response intensity, Z represents the stability of the operating conditions within the current time window. t The indicators represent abnormal leaf conditions, with α1, α2, α3, and α4 being non-negative weighting coefficients. A normalized mapping function to map the calculation results to a pre-defined confidence interval; The acoustic residual intensity and the comprehensive degradation response intensity used for screening are respectively: Where, d R Represents the residual characteristics R of the voiceprint t The dimension, d E This indicates the enhanced voiceprint degradation features. The dimension; The stability of the operating conditions is determined based on the fluctuation of the standardized operating condition parameters within the current time window: Where, d C Indicates the number of operating parameters, s t,r ω represents the standard deviation of the parameter for the r-th standardized operating condition within the current time window. r Let be the non-negative weights corresponding to the parameters of the r-th type of working condition, and satisfy: The leaf abnormality index Z t Determined based on SCADA fault alarm information, historical anomaly records, and blade inspection results; when any information indicates a blade anomaly within the current time window, let Z... t =1, otherwise let Z =1. t =0; When the confidence level of the healthy sample Not less than the preset health confidence threshold τ h And the leaf abnormality index Z t When =0, the voiceprint feature vector A corresponding to the current time window will be... t and operating condition parameter vector C t The samples were identified as incremental healthy samples and stored in the incremental healthy sample database.

5. The method for predicting the remaining life of wind turbine blades based on the degradation characteristics of acoustic signatures under operating conditions, as described in claim 1, is characterized in that: In step S3, the degradation characterization quantities include the intensity of the acoustic residual, the degree of abnormal frequency band shift, the degree of periodic acoustic perturbation, and the degree of time-frequency structure distortion, which are calculated as follows: Acoustic residual intensity , Abnormal frequency band shift , Periodic voiceprint perturbation level , Time-frequency structural distortion ; Among them, Ω b For the sensitive frequency band range related to blade degradation, To represent the average baseline of the enhanced residual characteristics under healthy conditions at frequency f, f b The frequency of the blade passage, kf b K represents the kth harmonic of the frequency through which the blade passes. b Indicates the harmonic order involved in the calculation. The gradient of the enhanced voiceprint degradation features in the time or frequency direction.

6. The method for predicting the remaining life of wind turbine blades based on the acoustic signature degradation characteristics under operating conditions, as described in claim 5, is characterized in that: The leaf acoustic health index mentioned in step S3 Calculated by the following formula: in, , , , These represent the normalized residual intensity of the voiceprint, the degree of anomalous frequency band shift, the degree of periodic voiceprint perturbation, and the degree of time-frequency structure distortion, respectively; α, β, γ, and δ are non-negative weighting coefficients and satisfy the following conditions: The leaf acoustic health index HI t The higher the value, the greater the degree of degradation of the leaf's acoustic signature; The smooth health index The result was obtained through a moving average: Where L is the length of the sliding time window; A dynamic degradation trajectory is constructed based on a smoothed health index over several consecutive time windows: in, This represents the dynamic degradation trajectory of the blade acoustic signature up to the t-th time window, where m is the number of historical windows. The degradation rate is calculated based on the smoothed health index corresponding to adjacent time windows: Among them, V t Δt represents the voiceprint degradation rate corresponding to the t-th time window, and Δt represents the time interval between adjacent time windows.

7. The method for predicting the remaining life of wind turbine blades based on the acoustic degradation characteristics of operating conditions as described in claim 6, characterized in that: Step S3 also includes a degradation monotonic constraint to suppress the abnormal decrease in the smoothness health index of the same wind turbine blade during continuous degradation. The degradation monotonic constraint loss is: Where M represents the number of continuous time windows participating in the degenerate monotonic constraint. and These represent the smoothed health index corresponding to adjacent time windows.

8. The method for predicting the remaining life of wind turbine blades based on the acoustic degradation characteristics of operating conditions as described in claim 1, characterized in that: The Transformer lifetime prediction network in step S4 further includes an uncertainty assessment module, which evaluates the uncertainty based on the temporal degradation state code E. t Output prediction variance σ t 2 , Among them, F u (·) indicates the uncertainty assessment module; Then based on the remaining life prediction value and the predicted standard deviation σ t The remaining lifetime prediction range is obtained. , where z represents the interval coefficient corresponding to the preset confidence level.

9. The method for predicting the remaining life of wind turbine blades based on the acoustic degradation characteristics of operating conditions as described in claim 8, characterized in that: The training loss of the Transformer lifetime prediction network in step S4 includes the remaining lifetime prediction loss. Lifetime trend constraint loss and life risk level prediction loss The total loss is: Where η1 and η2 are weighting coefficients; The remaining lifetime prediction loss L rul Gaussian negative log-likelihood loss is used: in, This represents the predicted remaining lifetime value corresponding to the t-th time window. N is the label for the actual remaining lifetime. R The total number of time windows participating in training, where t is the time window number; The lifetime trend constraint loss is used to constrain the overall trend of the remaining lifetime prediction to decrease over time, thus suppressing unexpected increases. in A non-negative relaxation factor that allows for small fluctuations in the remaining lifetime forecast; The accuracy of risk level classification is measured using cross-entropy loss to determine the prediction loss for life risk levels. Where C represents the number of categories of the preset life risk level. This indicates that the true life risk level corresponding to the t-th time window belongs to category c. This represents the probability of the c-th type of prediction output from the life risk level prediction branch.

10. The method for predicting the remaining life of wind turbine blades based on the acoustic signature degradation characteristics under operating conditions, as described in claim 1, is characterized in that: The Transformer lifetime prediction network in step S4 further includes a lifetime risk level prediction branch. This branch outputs the predicted probability of each lifetime risk level and determines the category corresponding to the highest predicted probability as the lifetime risk level for the t-th time window. Where C represents the number of categories of life risk level, This represents the predicted probability that the t-th time window belongs to the c-th life risk level. This represents the life risk level corresponding to the t-th time window; each life risk level is coded as 1, 2, ..., C in order of increasing risk level. Based on the predicted remaining lifespan and life risk level Determine the recommended maintenance time window: in, This represents the suggested maintenance time window set, τ r τ represents the remaining lifespan warning threshold. y This indicates the risk level warning threshold.

Images