An Adaptive Wind Turbine Monitoring Method and System Based on Vibration Acoustic Fingerprint Fusion Diagnosis

By combining multi-sensor fusion and adaptive processing of vibration and acoustic signals, and utilizing an edge AI quantization model, early identification and accurate diagnosis of wind turbine bearing faults are achieved. This solves the problems of difficult identification and insufficient real-time early warning in existing technologies, and improves the accuracy and efficiency of wind turbine monitoring.

CN123109965APending Publication Date: 2026-11-17THREE GORGES NEW ENERGY WUJIAQU POWER GENERATION CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611297527.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing wind turbine monitoring technologies cannot effectively identify early damage in different parts of wind turbine bearings, are easily affected by environmental noise, and lack real-time fault warning capabilities.

Method used

Vibration and acoustic signature signals are acquired simultaneously using a triaxial accelerometer and microphone array. Through adaptive filtering and preprocessing, acoustic signature features are extracted and fused with vibration signals. An edge AI quantization model is then used for fault identification and classification.

Benefits of technology

It enables early identification and accurate diagnosis of wind turbine bearing faults, reduces false alarm rate, reduces operation and maintenance costs, and improves the real-time performance and accuracy of diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109965A_ABST
    Figure CN123109965A_ABST
Patent Text Reader

Abstract

An adaptive wind turbine monitoring system and method based on vibration and acoustic signature fusion diagnosis is disclosed. The system includes a vibration sensing module, a speed synchronous acquisition module, an adaptive signal processing module, an acoustic signature perception and fusion analysis module, an edge fault diagnosis module, a power supply module, a wireless transmission module, and an early warning control module. The system synchronously acquires wind turbine vibration and acoustic signature signals using a triaxial accelerometer and microphone array. Adaptive filtering preprocessing is performed on the acquired signals, combined with generator speed signals, to remove environmental noise and amplify the target frequency band signal. Acoustic signature features are extracted and encoded into an acoustic signature supervector via a time-delay neural network, which is then weighted and fused with vibration and speed features. The fused feature vector is quantized and input into a model for fault identification, outputting fault types such as bearing inner ring, outer ring, rolling elements, and cage, and three levels: minor, early warning, and alarm. This achieves early warning and accurate diagnosis of wind turbine bearing faults, effectively reducing unplanned downtime losses and maintenance costs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment condition monitoring technology, specifically relating to an adaptive wind turbine monitoring method and system based on vibration and acoustic signature fusion diagnosis. Background Technology

[0002] Wind power generation, as an important component of clean energy, has been widely used in recent years. Taking the Beitashan Wind Farm in Changji area as an example, this wind farm has an installed capacity of 198MW, with 132 Goldwind 1.5MW units. Built in 2015, it has been 10 years since then, and the generators have suffered varying degrees of damage. However, the vibration detection devices installed at the beginning of construction, which only analyze the X and Y axis vibration data of the units, are no longer sufficient to detect bearing damage.

[0003] Existing wind turbine monitoring technologies have the following shortcomings: First, traditional vibration detection only collects biaxial vibration data, providing a single information dimension and making it difficult to accurately identify early damage in different parts of the bearing, such as the inner and outer rings, rolling elements, and cages. Second, the environmental noise in wind farms is complex, and relying solely on vibration signals is easily subject to interference, leading to false alarms or missed alarms. Third, there is a lack of ability to collect and analyze acoustic signals, making it impossible to fully utilize the rich fault information contained in the operating sounds of the equipment. Fourth, traditional monitoring systems adopt a cloud-based analysis mode, resulting in large data transmission volumes and high response delays, making it difficult to achieve real-time fault early warning. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an adaptive wind turbine monitoring method and system based on vibration and acoustic fingerprint fusion diagnosis, which realizes early identification and accurate diagnosis of wind turbine bearing faults through multi-sensor fusion, adaptive signal processing and edge AI quantization model.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis, comprising the following steps: S1: The vibration and acoustic signals of the wind turbine are collected synchronously through a triaxial accelerometer and a microphone array, and the generator speed signal is obtained. S2: Adaptive filtering and preprocessing are performed on the collected vibration and acoustic signals to remove environmental noise and amplify the target frequency band signal; S3: Extract voiceprint features from the preprocessed voiceprint signal, and combine them with vibration signal and rotation speed signal for multimodal fusion to construct a fused feature vector; S4: Input the fused feature vector into the edge AI quantization model for fault identification and classification; S5: Based on the model output results, determine the fault type and fault level of the wind turbine bearing or cage, and output the corresponding early warning or alarm signal.

[0006] The adaptive filtering and preprocessing in S2 includes: pre-emphasis processing of the voiceprint signal to enhance high-frequency components and reduce low-frequency interference. The pre-emphasis formula is as follows: ; in This is the nth sampling point of the original speech signal. These are the sampling points for the pre-emphasized speech signal. The pre-emphasis coefficient ranges from [0.95, 0.98]. Then, a Hanning window is used for frame segmentation and windowing to divide the non-stationary speech signal into short-time stationary frames. The Hanning window function is: ; The Hanning window function value, where n is the index of the sampling point within the window. The number of sampling points per frame after framing.

[0007] In step S3, extracting voiceprint features includes: converting the time-domain signal into a frequency-domain signal using Fourier transform to obtain amplitude information at each frequency point; filtering the frequency-domain power spectrum using a Mel filter bank, where the conversion relationship between Mel frequency and linear frequency is: ; ; Based on this, the center frequency of the Mel filter is determined. The power spectrum is weighted and summed using the weights of the Mel-triangle filter. The weighted summation formula is as follows: ; The formula for obtaining FBank features is: ; in This is the result of the Fourier transform of the speech frame at the k-th frequency point. Let be the weight of the j-th Mel-triangle filter at the k-th frequency point; then, frame-level encoding is performed using a time-delay neural network, with the following formula: ; The TDNN encoded output of the t-th frame of speech; This is the weight matrix; It is the bias vector; Size of the time context window; The Sigmoid activation function is used to aggregate the variable-length frame-level coded features into a speaker hypervector, which is then processed by mean averaging. , with standard deviation splicing ,in The total number of frames; linear transformation and batch normalization are performed on the voiceprint hypervector to obtain the standardized voiceprint perception vector: ; in This is for batch normalization operations.

[0008] The multimodal fusion in S3 is as follows: the voiceprint sensing vector is fused with vibration features and rotational speed features by weighted averaging to obtain a fused feature vector. The weighted averaging fusion formula is: And satisfy ,in The weighting coefficients for the k-th model and its features. Let be the voiceprint recognition score of the k-th model; the cosine similarity function for voiceprint matching is: ; in Final voiceprint embedding vector.

[0009] The edge quantization model described in S4 is a CNN-LSTM-Attention hybrid quantization model. Before inputting the model, the fused feature vector is preprocessed with INT8 quantization to convert the high-dimensional floating-point feature vector into a low-precision integer vector. The quantization formula is as follows: ; in To fuse floating-point eigenvalues ​​in the feature vector, The quantized 8-bit integer feature value. The maximum value of all feature values ​​in the eigenvector is merged. The minimum value of all eigenvalues ​​in the fused eigenvector.

[0010] The edge quantization model described in S4 is based on a lightweight deep learning network, extracting local fault features from the fused features. The model inference formula is as follows: ; in , These are the quantized network weights and biases, respectively. This is the quantized input feature vector.

[0011] The fault types in S5 include: normal state, generator bearing inner ring damage, generator bearing outer ring damage, generator bearing rolling element damage, and generator cage damage; the fault levels include: minor, warning, and alarm; the model output is a probability vector. By decoding the formula: ; in, The label corresponding to the maximum value in the probability vector is selected for the selection function, and is used as the predicted fault type or fault level.

[0012] Another aspect of the present invention provides an adaptive wind turbine monitoring system based on vibration acoustic signature fusion diagnosis, the system being implemented based on the method and comprising: The vibration sensing module is used to collect triaxial vibration signals from the front and rear shafts of the wind turbine generator; The speed synchronization acquisition module is used to acquire the generator speed signal; An adaptive signal processing module is used for adaptive filtering and preprocessing of vibration signals and acoustic signals; The voiceprint perception and fusion analysis module is used to extract voiceprint features and fuse them with vibration signals and rotation speed signals in a multimodal manner to construct a fused feature vector; The edge fault diagnosis module is used to input the fused feature vector into the edge quantization model for fault identification and classification; The power supply module is used to supply power to the various modules of the system; The wireless transmission module is used to transmit diagnostic results and alarm information; The early warning control module is used to output corresponding early warning and alarm signals based on the fault identification results.

[0013] The adaptive signal processing module synchronously processes vibration data, sound data, and rotational speed data through a multi-channel data processing module. It performs adaptive processing on the vibration data and sound data respectively, and eliminates noise and interference through filtering algorithms to ensure the output target frequency band.

[0014] The edge fault diagnosis module is based on a lightweight deep learning network. It decodes the probability vector of the inference output to restore the fault type and fault level, and ensures the accuracy of the results through threshold verification.

[0015] The main beneficial effects of this invention are as follows: 1. By identifying anomalies through subtle changes in vibration and acoustic signatures, sudden wind turbine shutdowns can be avoided, significantly reducing losses from unplanned shutdowns. Furthermore, continuous monitoring of the changing trends of vibration and acoustic signature characteristics can issue mild warnings when initial signs of faults appear, providing maintenance personnel with sufficient time to respond and effectively preventing power generation losses caused by sudden shutdowns.

[0016] 2. This invention employs multi-sensor fusion judgment and eliminates environmental noise through adaptive filtering, resulting in a low false alarm rate and more objective and reliable diagnostic results. By using pre-emphasis processing and frame-by-frame windowing technology, it effectively suppresses complex environmental noise in wind farms. Combined with a multi-modal weighted fusion strategy, it can achieve a diagnostic effect with high accuracy and low false alarm rate.

[0017] 3. This invention can reduce the number of times maintenance personnel need to climb to heights and conduct on-site inspections, avoid frequent disassembly leading to prolonged equipment downtime, extend the service life of components, and significantly reduce overall maintenance costs; by replacing manual periodic inspections with real-time edge diagnostics, the frequency of climbing to heights can be significantly reduced, blind disassembly and inspection can be avoided, and maintenance costs can be saved.

[0018] 4. By quantizing floating-point model parameters into low-precision integers, this invention can effectively compress the model size and improve inference speed, completing a single diagnosis within milliseconds, fully meeting the response requirements of real-time online monitoring of wind turbines. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0021] like Figure 1 A novel adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis is described, characterized by the following steps: S1: The vibration and acoustic signals of the wind turbine are collected synchronously through a triaxial accelerometer and a microphone array, and the generator speed signal is obtained. S2: Adaptive filtering and preprocessing are performed on the collected vibration and acoustic signals to remove environmental noise and amplify the target frequency band signal; S3: Extract voiceprint features from the preprocessed voiceprint signal, and combine them with vibration signal and rotation speed signal for multimodal fusion to construct a fused feature vector; S4: Input the fused feature vector into the edge AI quantization model for fault identification and classification; S5: Based on the model output results, determine the fault type and fault level of the wind turbine bearing or cage, and output the corresponding early warning or alarm signal.

[0022] Furthermore, the adaptive filtering and preprocessing in S2 includes: pre-emphasis processing of the voiceprint signal to enhance high-frequency components and reduce low-frequency interference. The pre-emphasis formula is as follows: ; in This is the nth sampling point of the original speech signal. These are the sampling points for the pre-emphasized speech signal. The pre-emphasis coefficient ranges from [0.95, 0.98]. Then, a Hanning window is used for frame segmentation and windowing to divide the non-stationary speech signal into short-time stationary frames. The Hanning window function is: ; The Hanning window function value, where n is the index of the sampling point within the window. The number of sampling points per frame after framing.

[0023] Furthermore, in S3, extracting voiceprint features includes: converting the time-domain signal into a frequency-domain signal using Fourier transform to obtain amplitude information at each frequency point; filtering the frequency-domain power spectrum using a Mel filter bank, where the conversion relationship between Mel frequency and linear frequency is: ; ; Based on this, the center frequency of the Mel filter is determined. The power spectrum is weighted and summed using the weights of the Mel-triangle filter. The weighted summation formula is as follows: ; The formula for obtaining FBank features is: ; in This is the result of the Fourier transform of the speech frame at the k-th frequency point. Let be the weight of the j-th Mel-triangle filter at the k-th frequency point; then, frame-level encoding is performed using a time-delay neural network, with the following formula: ; The TDNN encoded output of the t-th frame of speech; This is the weight matrix; It is the bias vector; Size of the time context window; The Sigmoid activation function is used to aggregate the variable-length frame-level coded features into a speaker hypervector, which is then processed by mean averaging. , with standard deviation splicing ,in The total number of frames; linear transformation and batch normalization are performed on the voiceprint hypervector to obtain the standardized voiceprint perception vector: ; in This is for batch normalization operations.

[0024] Furthermore, the multimodal fusion in S3 is as follows: the voiceprint perception vector is fused with vibration features and rotational speed features by weighted averaging to obtain a fused feature vector. The weighted averaging fusion formula is: And satisfy ,in The weighting coefficients for the k-th model and its features. Let be the voiceprint recognition score of the k-th model; the cosine similarity function for voiceprint matching is: ; in Final voiceprint embedding vector.

[0025] Furthermore, the edge quantization model in S4 is a CNN-LSTM-Attention hybrid quantization model; before inputting it into the model, the fused feature vector is preprocessed with INT8 quantization to convert the high-dimensional floating-point feature vector into a low-precision integer vector. The quantization formula is as follows: ; in To fuse floating-point eigenvalues ​​in the feature vector, The quantized 8-bit integer feature value. The maximum value of all feature values ​​in the eigenvector is merged. The minimum value of all eigenvalues ​​in the fused eigenvector.

[0026] Furthermore, the edge quantization model in S4 is based on a lightweight deep learning network, extracting local fault features from the fused features. The model inference formula is as follows: ; in , These are the quantized network weights and biases, respectively. This is the quantized input feature vector.

[0027] Furthermore, the fault types in S5 include: normal state, generator bearing inner ring damage, generator bearing outer ring damage, generator bearing rolling element damage, and generator cage damage; the fault levels include: minor, warning, and alarm; the model output is a probability vector. By decoding the formula: ; in, The label corresponding to the maximum value in the probability vector is selected for the selection function and used as the predicted fault type or fault level, while a confidence threshold is set. If the maximum probability is less than 0.7, it is determined that accurate diagnosis is impossible, triggering an alarm on the edge device and uploading data to the cloud for further analysis; the fault level determination criteria are: When the characteristic amplitude accounts for less than 15% and the effective vibration value is ≤ 1.2 times the threshold, it is classified as a slight level. A warning level is set when the characteristic amplitude accounts for 15% to 30% and the effective vibration value is ≤ 1.5 times the threshold. An alarm level is set when the characteristic amplitude accounts for more than 30% and the effective vibration value is more than 1.5 times the threshold.

[0028] like Figure 2 As shown above, an adaptive wind turbine monitoring system based on vibration acoustic signature fusion diagnosis includes: The vibration sensing module uses a triaxial accelerometer, installed at the front and rear bearing housings of the generator, to collect vibration acceleration signals in the X, Y, and Z directions. The sensor range and sensitivity are selected based on the wind turbine model and vibration level. The speed synchronization acquisition module is connected to the generator speed sensor to acquire the generator's real-time speed signal; The adaptive signal processing module synchronously processes vibration, sound, and rotational speed data through a multi-channel data processing module. It performs adaptive processing on the vibration and sound data separately, using filtering algorithms to eliminate noise and interference, ensuring the output reaches the target frequency band. Specifically, it includes a pre-emphasis filter, a Hanning window framing unit, a Fourier transform unit, and an adaptive filtering unit. The voiceprint perception and fusion analysis module includes a microphone array, a Mel filter bank, a time-delay neural network, and a fusion unit. The microphone array is installed near the generator to collect the unit's operating voiceprint signals; the Mel filter bank and the TDNN are used to extract voiceprint features; the fusion unit is used to perform weighted average fusion of voiceprint features, vibration features, and rotational speed features to construct a fused feature vector. The edge fault diagnosis module is based on a lightweight deep learning network (CNN-LSTM-Attention hybrid quantization model) and integrates the INT8 quantization inference engine. It decodes the probability vector output by inference to reconstruct the fault type and level, while threshold verification ensures the accuracy of the results. The module outputs five states: normal, generator bearing inner race damage, generator bearing outer race damage, generator bearing rolling element damage, and generator cage damage. Each damage type corresponds to three levels: minor, warning, and alarm.

[0029] The power supply module is used to supply power to the various modules of the system; The wireless transmission module is used to transmit diagnostic results and alarm information to the wind farm control center or cloud platform. The early warning control module is used to output corresponding early warning or alarm signals based on the fault identification results.

[0030] Furthermore, the adaptive signal processing module synchronously processes vibration data, sound data, and rotational speed data through a multi-channel data processing module, performs adaptive processing on the vibration data and sound data respectively, and eliminates noise and interference through filtering algorithms to ensure the output target frequency band.

[0031] Furthermore, the edge fault diagnosis module is based on a lightweight deep learning network, which decodes the probability vector of the inference output to restore the fault type and fault level, and ensures the accuracy of the results through threshold verification.

[0032] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this invention. It should be understood that the above descriptions are merely specific embodiments of this invention and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis, characterized in that, Includes the following steps: S1: The vibration and acoustic signals of the wind turbine are collected synchronously through a triaxial accelerometer and a microphone array, and the generator speed signal is obtained. S2: Adaptive filtering and preprocessing are performed on the collected vibration and acoustic signals to remove environmental noise and amplify the target frequency band signal; S3: Extract voiceprint features from the preprocessed voiceprint signal, and combine them with vibration signal and rotation speed signal for multimodal fusion to construct a fused feature vector; S4: Input the fused feature vector into the edge AI quantization model for fault identification and classification; S5: Based on the model output results, determine the fault type and fault level of the wind turbine bearing or cage, and output the corresponding early warning or alarm signal.

2. The adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis according to claim 1, characterized in that: The adaptive filtering and preprocessing in S2 includes: pre-emphasis processing of the voiceprint signal to enhance high-frequency components and reduce low-frequency interference. The pre-emphasis formula is as follows: ; in This is the nth sampling point of the original speech signal. These are the sampling points for the pre-emphasized speech signal. The pre-emphasis coefficient ranges from [0.95, 0.98]. Then, a Hanning window is used for frame segmentation and windowing to divide the non-stationary speech signal into short-time stationary frames. The Hanning window function is: ; The Hanning window function value, where n is the index of the sampling point within the window. The number of sampling points per frame after framing.

3. The adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis according to claim 1, characterized in that: In step S3, extracting voiceprint features includes: converting the time-domain signal into a frequency-domain signal using Fourier transform to obtain amplitude information at each frequency point; filtering the frequency-domain power spectrum using a Mel filter bank, where the conversion relationship between Mel frequency and linear frequency is: ; ; Based on this, the center frequency of the Mel filter is determined. The power spectrum is weighted and summed using the weights of the Mel-triangle filter. The weighted summation formula is as follows: ; The formula for obtaining FBank features is: ; in This is the result of the Fourier transform of the speech frame at the k-th frequency point. Let be the weight of the j-th Mel-triangle filter at the k-th frequency point; then, frame-level encoding is performed using a time-delay neural network, with the following formula: ; The TDNN encoded output of the t-th frame of speech; This is the weight matrix; It is the bias vector; Size of the time context window; The Sigmoid activation function is used to aggregate the variable-length frame-level coded features into a speaker hypervector, which is then processed by mean averaging. , with standard deviation splicing ,in The total number of frames; linear transformation and batch normalization are performed on the voiceprint hypervector to obtain the standardized voiceprint perception vector: ; in This is for batch normalization operations.

4. The adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis according to claim 1, characterized in that: The multimodal fusion in S3 is as follows: the voiceprint sensing vector is fused with vibration features and rotational speed features by weighted averaging to obtain a fused feature vector. The weighted averaging fusion formula is: And satisfy ,in The weighting coefficients for the k-th model and its features. The voiceprint recognition score of the k-th model; The cosine similarity function for voiceprint matching is: ; in Final voiceprint embedding vector.

5. The adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis according to claim 1, characterized in that: The edge quantization model described in S4 is a CNN-LSTM-Attention hybrid quantization model. Before inputting the model, the fused feature vector is preprocessed with INT8 quantization to convert the high-dimensional floating-point feature vector into a low-precision integer vector. The quantization formula is as follows: ; in To fuse floating-point eigenvalues ​​in the feature vector, The quantized 8-bit integer feature value. The maximum value of all feature values ​​in the eigenvector is merged. The minimum value of all eigenvalues ​​in the fused eigenvector.

6. The adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis according to claim 1, characterized in that: The edge quantization model described in S4 is based on a lightweight deep learning network, extracting local fault features from the fused features. The model inference formula is as follows: ; in , These are the quantized network weights and biases, respectively. This is the quantized input feature vector.

7. The adaptive wind turbine monitoring method based on vibration acoustic signature fusion diagnosis according to claim 1, characterized in that: The fault types in S5 include: normal state, generator bearing inner ring damage, generator bearing outer ring damage, generator bearing rolling element damage, and generator cage damage; the fault levels include: minor, warning, and alarm; the model output is a probability vector. By decoding the formula: ; in, The label corresponding to the maximum value in the probability vector is selected for the selection function, and is used as the predicted fault type or fault level.

8. An adaptive wind turbine monitoring system based on vibration acoustic signature fusion diagnosis, characterized in that, The system is implemented based on the method described in any one of claims 1-7, comprising: The vibration sensing module is used to collect triaxial vibration signals from the front and rear shafts of the wind turbine generator; The speed synchronization acquisition module is used to acquire the generator speed signal; An adaptive signal processing module is used for adaptive filtering and preprocessing of vibration signals and acoustic signals; The voiceprint perception and fusion analysis module is used to extract voiceprint features and fuse them with vibration signals and rotation speed signals in a multimodal manner to construct a fused feature vector; The edge fault diagnosis module is used to input the fused feature vector into the edge quantization model for fault identification and classification; The power supply module is used to supply power to the various modules of the system; The wireless transmission module is used to transmit diagnostic results and alarm information; The early warning control module is used to output corresponding early warning and alarm signals based on the fault identification results.

9. The adaptive wind turbine monitoring system based on vibration acoustic signature fusion diagnosis according to claim 8, characterized in that: The adaptive signal processing module synchronously processes vibration data, sound data, and rotational speed data through a multi-channel data processing module. It performs adaptive processing on the vibration data and sound data respectively, and eliminates noise and interference through filtering algorithms to ensure the output target frequency band.

10. The adaptive wind turbine monitoring system based on vibration acoustic signature fusion diagnosis according to claim 8, characterized in that: The edge fault diagnosis module is based on a lightweight deep learning network. It decodes the probability vector of the inference output to restore the fault type and fault level, and ensures the accuracy of the results through threshold verification.