A wind turbine soundprint feature dimension reduction and evolution analysis method and related device

Abstract

The application discloses a wind turbine voiceprint feature dimension reduction and evolution analysis method and a related device, comprising: collecting a multi-channel voiceprint signal of a wind turbine, recording the rotating speed and load working condition parameters of the wind turbine; performing improved MFCC time-frequency joint feature on the multi-channel voiceprint signal of the wind turbine to obtain a three-dimensional feature tensor; inputting the three-dimensional feature tensor into a 3D-CNN network to obtain a 128-dimensional deep feature vector; compensating the 128-dimensional deep feature vector through the rotating speed and load working condition parameters of the wind turbine; inputting the compensated 128-dimensional deep feature vector into an Attention-LSTM network to obtain a voiceprint feature evolution trajectory; and performing degradation stage quantization and critical point prediction based on the voiceprint feature evolution trajectory. The method and the related device can solve the technical problems of high voiceprint feature dimension, insufficient time-frequency coupling representation, difficult degradation evolution trend quantization and poor cross-working condition generalization performance in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis and intelligent operation and maintenance technology, and relates to a method and related device for dimensionality reduction and evolution analysis of acoustic features of wind turbine generators. Background Technology

[0002] As the core equipment for clean energy power generation, the operational reliability of wind turbines directly affects the economic benefits of wind farms and the stability of the power grid. During long-term operation, major components of wind turbines (including gearboxes, generators, and main shaft bearings) are subjected to the combined effects of alternating loads, environmental corrosion, and lubrication deterioration, gradually leading to progressive degradation such as fatigue cracks, wear, and pitting. While traditional vibration monitoring methods can capture mechanical fault characteristics, they have inherent limitations, including restrictions on sensor installation, rapid attenuation of high-frequency signals, and insensitivity to early, subtle faults.

[0003] Acoustic fingerprint monitoring technology, as a non-contact, broadband state sensing method, can capture the acoustic characteristics generated by the coupling of mechanical vibration, aerodynamic noise, and electromagnetic noise during the operation of wind turbines. Compared with vibration signals, acoustic fingerprint signals have advantages such as convenient acquisition, wide coverage, and sensitivity to early faults, making them particularly suitable for state monitoring of wind turbines under complex operating conditions.

[0004] However, existing voiceprint analysis technology has the following prominent problems in practical applications: 1. The curse of dimensionality in feature extraction. The frequency range of wind turbine acoustic signature signals is wide (usually covering 20Hz~20kHz). The static feature vectors extracted by traditional Mel-Frequency Cepstral Coefficients (MFCC) have a high dimension (usually 13~40 dimensions) and lack the ability to characterize the dynamic evolution information of time and frequency. It is difficult to fully represent the spatiotemporal evolution law of acoustic signature signals, resulting in redundant inputs and low computational efficiency for subsequent classification or prediction models.

[0005] 2. Insufficient spatiotemporal coupling representation of time-frequency features. Acoustic fingerprint signals are essentially time-varying, non-stationary signals, and their spectral characteristics evolve dynamically over time. Existing methods mostly use one-dimensional convolutional neural networks (1D-CNN) or two-dimensional convolutional neural networks (2D-CNN) to process time-domain or frequency-domain features respectively, failing to effectively integrate the three-dimensional coupling relationship of acoustic fingerprint signals in the time domain, frequency domain, and channel dimensions. This results in insufficient feature representation capabilities and an inaccurate characterization of the evolution trend of the gradual degradation process of wind turbine units.

[0006] 3. Difficulty in quantifying degradation evolution trends. The evolution of large components in wind turbines from normal state to failure is a gradual process, and the acoustic signature characteristics exhibit nonlinear and non-monotonic changes with the degree of degradation. Existing methods mostly focus on fault classification and identification, lacking the ability to continuously model and quantify the evolution trajectory of acoustic signature characteristics, and cannot accurately predict the critical points of degradation stage transitions, making it difficult to meet the engineering requirements of early detection and early warning.

[0007] 4. Poor generalization performance across operating conditions. Wind turbines operate under complex and varied conditions (wind speed 3~25m / s, load 30%~100% of rated power), resulting in significant differences in the fundamental frequency characteristics and harmonic distribution of acoustic signature signals under different conditions. Existing acoustic signature analysis methods do not fully consider the adaptive compensation mechanism for operating conditions, leading to a sharp decline in model performance and insufficient generalization ability when migrating across operating conditions.

[0008] Therefore, there is an urgent need for a voiceprint analysis method that can efficiently reduce dimensionality, accurately characterize the time-frequency evolution features of voiceprints, quantify degradation trends, and have adaptive operating conditions, so as to break through the existing technical bottlenecks and provide reliable technical support for the intelligent operation and maintenance of large components of wind turbine units. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and related device for dimensionality reduction and evolution analysis of acoustic features of wind turbine generators. This method and related device can solve the technical problems of high dimensionality of acoustic features, insufficient time-frequency coupling characterization, difficulty in quantifying degradation and evolution trends, and poor generalization performance across operating conditions in the prior art.

[0010] To achieve the above objectives, this invention discloses a method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators, comprising: Collect multi-channel acoustic fingerprint signals from wind turbine generators and record the speed and load parameters of the wind turbine generators; Improved MFCC time-frequency joint features were applied to the multi-channel acoustic signature signal of the wind turbine to obtain a three-dimensional feature tensor. ; The three-dimensional feature tensor The input is fed into a 3D-CNN network to obtain a 128-dimensional deep feature vector; The 128-dimensional deep feature vector is compensated by the wind turbine's rotational speed and load parameters. The compensated 128-dimensional deep feature vector is input into the Attention-LSTM network to obtain the voiceprint feature evolution trajectory; Based on the evolution trajectory of the voiceprint features, the degradation stage is quantified and the critical point is predicted.

[0011] Further, multi-channel voiceprint signals of the wind turbine unit are collected through bone conduction sensors and ultrasonic sensor arrays deployed near the gearbox bearing seat, generator end cover and main shaft bearing seat.

[0012] Further, the Mel-Frequency Cepstral Coefficients (MFCC) of each frame of the voiceprint signal are calculated, and the first-order difference ΔMFCC and the second-order difference Δ²MFCC are extracted; The MFCC, ΔMFCC and Δ²MFCC are stacked along the channel dimension to generate a three-dimensional feature tensor , where T is the number of time frames, F is the Mel frequency dimension, C is the number of channels, and C=3.

[0013] Further, the 3D-CNN network structure includes: Conv3D(32,k=(3,3,3))→BatchNorm→ReLU→MaxPool3D(2,2,1)→Conv3D(64,k=(3,3,3))→BatchNorm→ReLU→MaxPool3D(2,2,1)→Conv3D(128,k=(3,3,3))→BatchNorm→ReLU→GlobalAveragePool3D→Flatten→Dense(256,ReLU)→Dropout(0.3)→Dense(128,ReLU) Further, the LSTM network adopts a 2-layer stacked structure, each layer contains 128 hidden units, and the input sequence length corresponds to 30 days of continuous monitoring data; the attention mechanism adopts multi-head self-attention.

[0014] Further, based on the evolution trajectory of the voiceprint features, the health index H(t)∈[0,1] is defined, and the model output is mapped to the health index interval through the Sigmoid function, where H(t)>0.85 is the normal period, 0.6<H(t)≤0.85 is the slow degradation period, 0.3<H(t)≤0.6 is the accelerated degradation period, and H(t)≤0.3 is the critical failure period; The change rate of the health index v(t)=dH(t) / dt is calculated, and when v(t) exceeds the preset threshold and lasts for more than 3 time steps, it is determined as a stage transition critical point, and early warning information is output.

[0015] Further, the method also includes: outputting and displaying the health index degradation curve, degradation stage division result and stage transition critical point prediction information of the large components of the wind turbine unit.

[0016] The present invention discloses a system for dimensionality reduction and evolution analysis of voiceprint features of a wind turbine unit, comprising: a collection module, configured to collect multi-channel voiceprint signals of the wind turbine unit, and record the rotation speed and load condition parameters of the wind turbine unit; The extraction module is used to perform improved MFCC time-frequency joint feature extraction on the multi-channel acoustic signature signal of the wind turbine generator to obtain a three-dimensional feature tensor. ; Input module, used to input the three-dimensional feature tensor The input is fed into a 3D-CNN network to obtain a 128-dimensional deep feature vector; The compensation module is used to compensate the 128-dimensional depth feature vector using the wind turbine's speed and load parameters. The inference module is used to input the compensated 128-dimensional deep feature vector into the Attention-LSTM network to obtain the evolution trajectory of the voiceprint features; The prediction module is used to quantify the degradation stage and predict the critical point based on the evolution trajectory of the voiceprint features.

[0017] This invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the wind turbine acoustic feature dimensionality reduction and evolution analysis method.

[0018] This invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the wind turbine acoustic signature feature dimensionality reduction and evolution analysis method.

[0019] The present invention has the following beneficial effects: The wind turbine acoustic signature feature dimensionality reduction and evolution analysis method and related device described in this invention, in specific operation, constructs an improved MFCC acoustic signature feature extraction system, integrates the improved MFCC and dynamic difference coefficients to generate a time-frequency joint feature tensor; uses a 3D-CNN network to deeply characterize and reduce the three-dimensional spatiotemporal coupling relationship of acoustic signature features, extracting high-discriminative low-dimensional feature vectors; combines an attention-based long short-term memory network to continuously model the evolution trajectory of acoustic signature features, quantifying the critical points of degradation stage transitions; and introduces condition-adaptive compensation to achieve model generalization optimization under cross-condition conditions, ultimately outputting the health status assessment results and degradation trend prediction curves of major components of the wind turbine, providing precise technical support for the full life cycle health management of wind turbines. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A schematic diagram illustrating the improvement of time-frequency joint feature extraction for MFCC; Figure 3 Diagram of the 3D-CNN deep feature dimensionality reduction network structure; Figure 4 This is a schematic diagram of the adaptive compensation module for operating conditions. Figure 5 Network structure diagram for modeling the evolution trajectory of Attention-LSTM. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0024] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0026] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0027] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0029] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0030] The method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators described in this invention includes the following steps: 1) Synchronous acquisition and preprocessing of multi-channel acoustic signature signals; The acoustic signature signals of key components of the wind turbine (gearbox, generator, main shaft bearing) are collected synchronously, and the speed and load parameters of the wind turbine are recorded.

[0031] The acoustic signature signal is acquired by a bone conduction sensor and an ultrasonic sensor array deployed near the gearbox bearing housing, generator end cover, and main shaft bearing housing. The sampling frequency is not less than 51.2 kHz, and the frequency response range is 20 Hz to 20 kHz. The rotational speed is acquired by a photoelectric encoder, and the load is acquired by a torque sensor. The sampling frequency is synchronized with the acoustic signature signal, and the time synchronization error does not exceed ±5 μs.

[0032] Preprocessing includes: adaptive noise cancellation of the acoustic signal based on improved spectral subtraction to eliminate wind noise and environmental interference; pre-emphasis filtering (coefficient α=0.97) to compensate for high-frequency attenuation; and short-term signal stabilization through Hamming window framing (frame length 25ms, frame shift 10ms).

[0033] 2) Improved MFCC time-frequency joint feature extraction; An improved MFCC was constructed, which dynamically adjusts the center frequency and bandwidth of the filter according to the spectral energy distribution of the acoustic signature signal of the wind turbine, thereby enhancing the characteristic resolution of the fault-sensitive frequency band.

[0034] The Mel frequency cepstral coefficients (MFCC) of each frame of the voiceprint signal are calculated, and the first-order difference (ΔMFCC) and second-order difference (Δ²MFCC) dynamic coefficients are extracted to form a time-frequency joint feature tensor.

[0035] Stack MFCC, ΔMFCC, and Δ²MFCC along the channel dimension to generate a three-dimensional feature tensor. Where T is the number of time frames, F is the Mel frequency dimension (F=40), and C is the number of channels, C=3, to realize the joint representation of the time-frequency dynamic information of the voiceprint signal.

[0036] 3) 3D-CNN deep feature dimensionality reduction and spatiotemporal coupling representation; By inputting the three-dimensional feature tensor X into a 3D-CNN network, the coupling relationship in the time domain, frequency domain, and channel dimension is captured simultaneously through three-dimensional convolutional kernels, achieving efficient dimensionality reduction and deep representation of voiceprint features.

[0037] The 3D-CNN network architecture includes: Conv3D(32,k=(3,3,3))→BatchNorm→ReLU→MaxPool3D(2,2,1)→Conv3D(64,k=(3,3,3))→BatchNorm→ReLU→MaxPool3D(2,2,1)→Conv3D(128,k=(3,3,3))→BatchNorm→ReLU→GlobalAveragePool3D→Flatten→Dense(256,ReLU)→Dropout(0.3)→Dense(128,ReLU), finally outputting a 128-dimensional deep feature vector.

[0038] By fusing time-frequency channel information through three-dimensional convolution operations, the feature dimensionality reduction efficiency is improved by more than 40% compared with traditional 1D-CNN or 2D-CNN, while preserving key information on the spatiotemporal evolution of the voiceprint signal.

[0039] 4) Adaptive compensation for operating conditions and generalized optimization across operating conditions; Construct a working condition adaptive compensation module, and perform working condition adaptive compensation on the voiceprint features according to the rotational speed and load working condition parameters to obtain compensated voiceprint features, so as to eliminate the interference of working condition fluctuations on the feature distribution.

[0040] The working condition compensation formula is: , where g(ω,T) is the compensation function based on working condition parameters (rotational speed ω, load T), and Gaussian Process Regression (GPR) is used to model the working condition-feature mapping relationship.

[0041] An Adversarial Domain Adaptation mechanism for working conditions is introduced in the training phase, and a Gradient Reversal Layer (GRL) is used to constrain the feature extractor to learn working condition-invariant features, so as to improve the cross-working condition generalization ability of the model.

[0042] 5) Voiceprint feature evolution trajectory modeling based on Attention-LSTM; Input the 128-dimensional deep feature vector sequence into the Attention-LSTM network, and adaptively weight the features of key time steps through the attention mechanism to capture the evolution law of voiceprint features along with the degradation process.

[0043] The LSTM network adopts a 2-layer stacked structure, each layer contains 128 hidden units, and the input sequence length corresponds to 30 days of continuous monitoring data; the attention mechanism adopts multi-head self-attention (8 heads, hidden dimension 128), calculates the importance weight of features at each time step, and performs weighted summation to obtain the evolution feature vector.

[0044] Establish the voiceprint feature evolution trajectory model: , where H(t) is the health indicator at time t, is the 128-dimensional feature vector at the i-th time step, is the Attention-LSTM mapping function.

[0045] 6) Degradation stage quantification and critical point prediction; Based on the voiceprint feature evolution trajectory, the health indicator H(t)∈[0,1] is defined, and the model output is mapped to the health indicator interval through the Sigmoid function.

[0046] Divide the degradation stages: H(t)>0.85 is the normal period, 0.6<H(t)≤0.85 is the slow degradation period, 0.3<H(t)≤0.6 is the accelerated degradation period, and H(t)≤0.3 is the critical failure period.

[0047] A critical point prediction method based on evolution rate analysis is adopted: the rate of change of health indicators v(t) = dH(t) / dt is calculated, and when v(t) exceeds a preset threshold... If the condition persists for more than 3 time steps, it is determined to be a critical point for stage transition, and an early warning message is output.

[0048] 7) Result output and dynamic updates; Output the health index degradation curves of major components of the wind turbine, the degradation stage division results, and the prediction information of the stage transition critical point.

[0049] Incremental learning is performed using the Elastic Weight Consolidation (EWC) algorithm. When the accumulated new data reaches 200 sets and the prediction error exceeds 5%, the model is updated to ensure long-term prediction stability.

[0050] The displayed content is dynamically updated, and historical data review and trend comparison analysis are supported.

[0051] Example 1 1) Acquisition of voiceprint signals; A Knowles SPH0645LM4H-B bone conduction sensor (sensitivity -22dBFS, frequency response 20Hz~20kHz) and an ultrasonic sensor array (center frequency 40kHz, bandwidth ±5kHz) were selected and deployed near the gearbox bearing housing (P1-P4), generator end cover (P5-P6), and main shaft bearing housing (P7-P8), respectively. The sampling frequency was 51.2kHz, which meets the requirements of the Nyquist sampling theorem.

[0052] Operating condition parameter acquisition: Speed ​​measurement uses a photoelectric encoder (resolution 2048 pulses / revolution); load measurement uses a strain gauge torque sensor (range 0~8000Nm, accuracy ±0.05%); the synchronization mechanism adopts the IEEE 1588-2008 PTPv2 precise time protocol, and through a combination of hardware triggering and software calibration, it ensures that the sampling time error between the acoustic signal and the operating condition parameters does not exceed ±5μs.

[0053] 2) Voiceprint signal preprocessing; Adaptive noise cancellation: An improved spectral subtraction method is used to perform frequency domain filtering based on noise power spectrum estimation, which improves the signal-to-noise ratio by no less than 15dB.

[0054] Preemphasis Filtering: Transfer Function α=0.97, to compensate for the energy attenuation of the high-frequency band of the voiceprint signal.

[0055] Framed windowing: Frame length 25ms (1280 sampling points), frame shift 10ms (512 sampling points), Hamming window function w(n)=0.54-0.46cos(2πn / (N-1)), N=1280.

[0056] 3) Improved MFCC time-frequency joint feature extraction Improved MFCC Construction: Standard Mel filter banks have high resolution in the low-frequency band but low resolution in the high-frequency band. For the fault-sensitive frequency bands of wind turbine acoustic signature signals (gear meshing frequency typically 2~10kHz, bearing fault characteristic frequency 5~15kHz), an adaptive adjustment strategy based on spectral energy entropy is adopted: calculating the energy entropy of each frequency band. ,when Time reduction filter bandwidth This enhances the characteristic resolution of the frequency band; conversely, it widens the bandwidth, where β=0.3 is an adjustment coefficient. This is the energy threshold.

[0057] MFCC calculation: Perform FFT transformation on each frame of signal, take the logarithm after improving MFCC filtering, and then perform Discrete Cosine Transform (DCT) to retain the first 40 coefficients, thus obtaining a 40-dimensional MFCC feature vector.

[0058] Dynamic difference coefficients: Calculate the first-order difference ΔMFCC(t)=MFCC(t)-MFCC(t-1) and the second-order difference Δ²MFCC(t)=ΔMFCC(t)-ΔMFCC(t-1), each with 40 dimensions.

[0059] 3D Feature Tensor Generation: Stack MFCC, ΔMFCC, and Δ²MFCC along the channel dimension to form... Where T=1000 is the number of time frames (corresponding to 10 seconds of voiceprint signal).

[0060] 4) Dimensionality reduction of deep features in 3D-CNN; Network parameter configuration: Input layer:

[0061] Conv3D-1: 32 convolutional kernels, kernel size (3,3,3), stride (1,1,1), same padding, output (1000×40×3×32). BatchNorm-1 + ReLU + MaxPool3D(2,2,1): Output (500×20×1×32) Conv3D-2: 64 convolutional kernels, kernel size (3,3,3), output (500×20×1×64) BatchNorm-2 + ReLU + MaxPool3D(2,2,1): Output (250×10×1×64) Conv3D-3: 128 convolutional kernels, kernel size (3,3,3), output (250×10×1×128) BatchNorm-3 + ReLU + GlobalAveragePool3D: Output 128 dimensions Flatten→Dense(256, ReLU)→Dropout(0.3)→Dense(128, ReLU) The final output is a 128-dimensional deep feature vector, which is 99.89% lower than the original 120,000-dimensional (1000×40×3) input.

[0062] 5) Attention-LSTM evolution trajectory modeling; Network parameter configuration: Input layer: 128-dimensional feature vector sequence, sequence length L=4320 (corresponding to 30 days, 144 samples per day, sampled every 10 minutes). LSTM-1: 128 hidden units, returns a sequence LSTM-2: 128 hidden units, returns a sequence MultiHeadAttention: 8 heads, 128 hidden dimensions, output weighted sequence GlobalAveragePooling1D: Outputs 128-dimensional evolutionary features Dense(64, ReLU)→Dense(1, Sigmoid): Outputs health indicators

[0063] Training configuration: The dataset consists of 20 2.0MW permanent magnet synchronous wind turbines from a wind farm, collected from March 2022 to March 2024 (24 months), totaling 20,736,000 samples; the optimizer is Adam, with an initial learning rate of 1e-4, and cosine annealing scheduling; the loss function is... ,in, , α=0.05 is the smoothing weight coefficient, which is determined through validation set tuning.

[0064] 6) Quantification of degradation stages and prediction of critical points; Health Indicator Mapping: =Sigmoid(z), where z is the model output logits.

[0065] The stage division threshold is adaptively determined using K-means clustering (k=4), with a typical value being: normal period >0.85, slow degradation period 0.6< ≤0.85, accelerated degradation period 0.3< ≤0.6, critical failure period ≤0.3.

[0066] Critical point prediction: Calculate the evolution rate v(t) = (H(t) - H(t - Δt)) / Δt, where Δt = 1 day. When v(t) < -0.02 / day and persists for more than 3 days, it is determined to be a critical point of transition from the slow degradation phase to the accelerated degradation phase, and an early warning message is output.

[0067] 7) Adaptive compensation based on operating conditions; Gaussian process regression modeling: The training data includes acoustic features and operating parameters under different operating conditions (wind speed 4~25m / s, load 20%~100%). The model is established as g(ω,T)~GP(0, k((ω,T),(ω',T'))), and the kernel function k is the radial basis function (RBF).

[0068] Adversarial Domain Adaptation: The feature extractor and the working condition discriminator are trained adversarially through a gradient inversion layer (GRL). The discriminator learns to distinguish different working conditions, while the feature extractor learns to generate working condition-invariant features. Let the feature extractor be F, the working condition discriminator be D, and the health indicator predictor be P. The overall optimization objective is:

[0069] in, This represents the mean squared error loss in the prediction of health indicators. λ represents the cross-entropy loss of the condition discriminator, and λ=0.1 is the adversarial weight coefficient. For input sample features, d represents the health indicator label, and d represents the working condition label. This is achieved through a gradient inversion layer (GRL). Adversarial optimization, the mathematical expression of the gradient inversion layer (GRL) is:

[0070]

[0071] The first formula is the forward propagation expression, and the second formula is the backward propagation expression. It is an identity matrix.

[0072] 8) System verification and performance evaluation; Test environment: 15 identical 2.0MW permanent magnet synchronous wind turbines in a wind farm, test period from April 2024 to October 2024 (6 months).

[0073] Evaluation indicators: Feature dimensionality reduction efficiency: 3D-CNN achieves a dimensionality reduction rate of 99.89%, which is 75% lower than traditional PCA dimensionality reduction (requiring 512 dimensions to retain 95% variance).

[0074] Feature representation capability: t-SNE visualization shows that 3D-CNN improves the clustering of similar samples by 35% and increases the inter-class distance by 28% when extracting features.

[0075] Accuracy of identification in degradation stages: 99.2% in normal stage, 96.8% in slow degradation stage, 94.5% in accelerated degradation stage, and 98.1% in critical failure stage.

[0076] Critical point prediction accuracy: The prediction stage transition is predicted 15 to 30 days in advance, with a mean absolute error (MAE) of 3.2 days.

[0077] Cross-condition generalization capability: Under unobserved conditions (wind speed > 20 m / s), the relative error of health index prediction decreased from 12.3% (uncompensated) to 4.1% (compensated).

[0078] Real-time performance: Single sample processing time is less than 150ms, meeting the requirements of online monitoring.

[0079] It should be noted that the present invention has the following characteristics: 1. Simultaneous improvement in feature dimensionality reduction efficiency and representation accuracy: Traditional MFCC static features have high dimensionality and lack dynamic information. This invention constructs a three-dimensional time-frequency joint feature tensor of MFCC-ΔMFCC-Δ²MFCC, which outputs a 128-dimensional feature vector after 3D-CNN deep dimensionality reduction, with a dimensionality reduction rate of 99.89%. The three-dimensional convolution simultaneously captures the three-dimensional coupling relationship of temporal evolution, frequency distribution, and channel complementarity, increasing the clustering degree of similar samples by 35% and the inter-class distance by 28%, breaking through the limitation of traditional 1D-CNN / 2D-CNN that can only handle single-dimensional relationships.

[0080] 2. Continuous Quantification and Precise Early Warning of Degradation Evolution Trends: Existing methods focus on fault classification and lack the ability to continuously model the degradation process. This invention establishes a voiceprint feature evolution trajectory model using Attention-LSTM. The attention mechanism adaptively focuses on key time steps; based on the evolution rate v(t)=dH(t) / dt, it realizes the quantitative division of the four stages of degradation, predicts the critical point 15~30 days in advance (mean absolute error MAE=3.2 days), and the accuracy of degradation stage identification is greater than 96.8%, meeting the engineering requirements of early detection and early warning.

[0081] 3. Significantly Enhanced Cross-Condition Generalization Performance: The variable operating conditions of wind turbines lead to feature distribution drift, causing a sharp decline in the cross-condition transfer performance of existing models. This invention establishes a condition compensation function g(ω,T) based on Gaussian process regression to achieve feature normalization, and combines an adversarial domain adaptive mechanism to learn condition-invariant features; the relative prediction error under unseen operating conditions is reduced from 12.3% to 4.1%, ensuring stable prediction across the entire operating condition range.

[0082] 4. The system has long-term adaptive evolution capability: It adopts the Elastic Weight Consolidation (EWC) algorithm for incremental learning. When the new data accumulates to 200 sets and the error exceeds 5%, it triggers an update, taking into account both the preservation of historical knowledge and the adaptation to degradation trends. Edge real-time processing (less than 150ms) and cloud collaborative optimization form a closed loop. After 18 months of continuous operation, the prediction accuracy is maintained at more than 95%, avoiding catastrophic forgetting and ensuring the long-term reliability of the system.

[0083] 5. Excellent engineering practicality and economic value: Non-contact acoustic sensor is easy to install and requires no downtime, making it suitable for complex field environments; after actual deployment, unplanned downtime is reduced by 80%, maintenance costs are reduced by 35%, and spare parts inventory is optimized and reduced by 20%, providing high-value technical support for the full life cycle asset management of wind farms.

[0084] Example 2 The wind turbine acoustic signature feature dimensionality reduction and evolution analysis system of the present invention includes: The acquisition module is used to acquire multi-channel acoustic fingerprint signals from wind turbines and record the speed and load parameters of wind turbines. The extraction module is used to perform improved MFCC time-frequency joint feature extraction on the multi-channel acoustic signature signal of the wind turbine generator to obtain a three-dimensional feature tensor. ; Input module, used to input the three-dimensional feature tensor The input is fed into a 3D-CNN network to obtain a 128-dimensional deep feature vector; The compensation module is used to compensate the 128-dimensional depth feature vector using the wind turbine's speed and load parameters. The inference module is used to input the compensated 128-dimensional deep feature vector into the Attention-LSTM network to obtain the evolution trajectory of the voiceprint features; The prediction module is used to quantify the degradation stage and predict the critical point based on the evolution trajectory of the voiceprint features.

[0085] The module division in this embodiment is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in each embodiment of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0086] Example 3 A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the wind turbine acoustic signature feature dimensionality reduction and evolution analysis method, for example including: acquiring multi-channel acoustic signature signals from the wind turbine, recording the wind turbine's rotational speed and load parameters; and performing improved MFCC time-frequency joint feature analysis on the multi-channel acoustic signature signals of the wind turbine to obtain a three-dimensional feature tensor. ; the three-dimensional feature tensor The input is fed into a 3D-CNN network to obtain a 128-dimensional depth feature vector. This 128-dimensional depth feature vector is then compensated using the wind turbine's rotational speed and load parameters. The compensated 128-dimensional depth feature vector is then input into an Attention-LSTM network to obtain the voiceprint feature evolution trajectory. Degradation stage quantization and critical point prediction are performed based on this trajectory. The memory may include main memory, such as high-speed random access memory, or it may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which can be an industry-standard architecture bus, a peripheral component interconnection standard bus, or an extended industry-standard architecture bus. The bus can be categorized as an address bus, data bus, and control bus. The memory stores programs; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and it provides instructions and data to the processor.

[0087] Example 4 A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the wind turbine acoustic signature feature dimensionality reduction and evolution analysis method, for example including: acquiring multi-channel acoustic signature signals from the wind turbine, recording the wind turbine's rotational speed and load parameters; and performing improved MFCC time-frequency joint feature analysis on the multi-channel acoustic signature signals of the wind turbine to obtain a three-dimensional feature tensor. ; the three-dimensional feature tensor The input is fed into a 3D-CNN network to obtain a 128-dimensional depth feature vector. This 128-dimensional depth feature vector is then compensated using the wind turbine's rotational speed and load parameters. The compensated 128-dimensional depth feature vector is then input into an Attention-LSTM network to obtain the voiceprint feature evolution trajectory. Based on this trajectory, degradation stage quantization and critical point prediction are performed. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include read-only memory (ROM), hard disk, flash memory, optical disk, magnetic disk, etc.

[0088] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and disclosure of the invention. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0093] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

[0094] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators, characterized in that, comprising: collecting multi-channel voiceprint signals of the wind turbine, and recording the rotational speed and load operating condition parameters of the wind turbine; Improved MFCC time-frequency joint features were applied to the multi-channel acoustic signature signal of the wind turbine to obtain a three-dimensional feature tensor. ; The three-dimensional feature tensor The input is fed into a 3D-CNN network to obtain a 128-dimensional deep feature vector; compensating the 128-dimensional deep feature vector through the rotational speed and load operating condition parameters of the wind turbine; inputting the compensated 128-dimensional deep feature vector into an Attention-LSTM network to obtain a voiceprint feature evolution trajectory; performing degradation stage quantification and critical point prediction based on the voiceprint feature evolution trajectory.

2. The method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators according to claim 1, characterized in that, The multi-channel voiceprint signals of the wind turbine are collected by bone conduction sensors and ultrasonic sensor arrays deployed near the gearbox bearing seat, the generator end cover, and the main shaft bearing seat.

3. The method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators according to claim 1, characterized in that, calculating the Mel-frequency cepstral coefficient MFCC of each frame of voiceprint signal, and extracting the first-order difference ΔMFCC and the second-order difference Δ²MFCC; Stack MFCC, ΔMFCC, and Δ²MFCC along the channel dimension to generate a three-dimensional feature tensor. Where T is the number of time frames, F is the Mel frequency dimension, and C is the number of channels, C=3.

4. The method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators according to claim 1, characterized in that, The 3D-CNN network structure comprises: Conv3D(32,k=(3,3,3))→BatchNorm→ReLU→MaxPool3D(2,2,1)→Conv3D(64,k=(3,3,3))→BatchNorm→ReLU→MaxPool3D(2,2,1)→Conv3D(128,k=(3,3,3))→BatchNorm→ReLU→GlobalAveragePool3D→Flatten→Dense(256,ReLU)→Dropout(0.3)→Dense(128,ReLU).

5. The method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators according to claim 1, characterized in that, The LSTM network adopts a 2-layer stacked structure, each layer contains 128 hidden units, and the input sequence length corresponds to 30 days of continuous monitoring data; the attention mechanism adopts multi-head self-attention.

6. The method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators according to claim 1, characterized in that, Based on the voiceprint feature evolution trajectory, a health index H(t)∈[0,1] is defined, and the model output is mapped to the health index interval through a Sigmoid function, where H(t)>0.85 is the normal period, 0.6<H(t)≤0.85 is the slow degradation period, 0.3<H(t)≤0.6 is the accelerated degradation period, and H(t)≤0.3 is the critical failure period; Calculate the rate of change of health indicators v(t) = dH(t) / dt. When v(t) exceeds a preset threshold... If the condition persists for more than 3 time steps, it is determined to be a critical point for stage transition, and an early warning message is output.

7. The method for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators according to claim 1, characterized in that, further comprising: outputting and displaying the health index degradation curve of large components of the wind turbine, the degradation stage division result, and the prediction information of stage transition critical points.

8. A system for dimensionality reduction and evolution analysis of acoustic signature features of wind turbine generators, characterized in that, comprising: a collection module, configured to collect multi-channel voiceprint signals of the wind turbine, and record the rotational speed and load operating condition parameters of the wind turbine; The extraction module is used to perform improved MFCC time-frequency joint feature extraction on the multi-channel acoustic signature signal of the wind turbine generator to obtain a three-dimensional feature tensor. ; Input module, used to input the three-dimensional feature tensor The input is fed into a 3D-CNN network to obtain a 128-dimensional deep feature vector; a compensation module, configured to compensate the 128-dimensional deep feature vector through the rotational speed and load operating condition parameters of the wind turbine; an evolution module, configured to input the compensated 128-dimensional deep feature vector into an Attention-LSTM network to obtain a voiceprint feature evolution trajectory; a prediction module, configured to perform degradation stage quantification and critical point prediction based on the voiceprint feature evolution trajectory.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, the steps of the method for dimensionality reduction and evolution analysis of wind turbine voiceprint features according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, the steps of the method for dimensionality reduction and evolution analysis of wind turbine voiceprint features according to any one of claims 1 to 7 are implemented.

Images