A smart diagnostic device and method for blade icing based on acoustic and vibration feature fusion

Abstract

This invention belongs to the field of wind power equipment condition monitoring and fault diagnosis technology, specifically relating to an intelligent diagnostic device and method for blade icing based on acoustic-vibration feature fusion. It includes: a multimodal sensing unit for acquiring vibration response signals and aeroacoustic signals; a signal conditioning and synchronous acquisition unit connected to the multimodal sensing unit for conditioning and synchronous analog-to-digital conversion of the vibration response signals and aeroacoustic signals to obtain synchronized digital vibration and digital sound signals; and an edge computing and diagnostic unit connected to the signal conditioning and synchronous acquisition unit. This invention synchronously acquires and deeply fuses vibration signals reflecting the blade's structural state with specific frequency band sound signals reflecting the aerodynamic state of the blade surface, constructing an intelligent diagnostic system based on acoustic-vibration coherence analysis and machine learning models. This achieves high-precision, robust online diagnosis and early warning of blade icing conditions.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment condition monitoring and fault diagnosis technology, specifically relating to an intelligent diagnostic device and method for blade icing based on acoustic and vibration feature fusion. Background Technology

[0002] In cold and humid regions, icing on wind turbine blades is one of the primary threats to the safe, stable, and efficient operation of wind farms. Ice buildup significantly disrupts the aerodynamic shape of the blades, leading to a reduction in wind energy capture efficiency of over 20%, and causing severe mass and aerodynamic imbalances. This greatly increases the fatigue load on the blades, hub, and tower, and in extreme cases, can result in major accidents such as blade breakage and turbine shutdown, causing significant power generation losses and safety risks.

[0003] Therefore, early and accurate diagnosis of blade icing is a prerequisite for efficient de-icing and ensuring unit safety. Current diagnostic technologies mainly rely on monitoring a single physical quantity. Existing blade icing diagnostic technologies mainly suffer from the following bottlenecks: vibration-based diagnostic methods are easily affected by operating conditions such as wind speed and rotational speed, have low sensitivity to early thin ice, and have difficulty distinguishing between "modal changes caused by icing" and "response changes caused by changes in operating conditions"; temperature / image-based methods have limitations such as "low temperature does not equal icing", are easily affected by sunlight and weather, and cannot identify transparent ice, making it difficult to achieve quantitative online diagnosis; single-sensor solutions have poor fault tolerance and lack deep fusion of multi-source heterogeneous information, failing to utilize the inherent physical coupling relationship between blade vibration and aerodynamic noise during the icing process, resulting in significant bottlenecks in diagnostic accuracy and system robustness. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent diagnostic device and method for blade icing based on acoustic-vibration feature fusion. By synchronously collecting and deeply fusing vibration signals reflecting the structural state of the blade with specific frequency band sound signals reflecting the aerodynamic state of the blade surface, an intelligent diagnostic system based on acoustic-vibration coherence analysis and machine learning model is constructed, thereby achieving high-precision and robust online diagnosis and early warning of blade icing status.

[0005] The specific technical solution adopted by this invention is as follows: A smart diagnostic device for blade icing based on acoustic and vibration feature fusion, comprising: A multimodal sensing unit is used to acquire vibration response signals and aeroacoustic signals; The signal conditioning and synchronous acquisition unit is connected to the multimodal sensing unit and is used to condition and synchronously convert the vibration response signal and aeroacoustic signal to obtain synchronous vibration digital signal and sound digital signal. The edge computing and diagnostic unit is connected to the signal conditioning and synchronous acquisition unit and has a built-in signal preprocessing module, feature extraction and fusion module, and intelligent diagnostic module. The signal preprocessing module is used to preprocess the vibration digital signal and the sound digital signal. The feature extraction and fusion module is used to extract single-channel multidimensional features from the preprocessed vibration digital signal and sound digital signal, respectively. Calculate the coherence features of the vibration digital signal and the sound digital signal within a preset frequency band, and fuse the single-channel features and coherence features to form a multi-dimensional feature vector; The intelligent diagnostic module is used to input the multidimensional feature vector into a pre-trained machine learning diagnostic model to obtain the classification results of the leaf icing state and the corresponding confidence level.

[0006] The multimodal sensing unit includes at least one low-frequency acceleration sensor and at least one directional microphone array. The low-frequency acceleration sensor is installed at the root of the wind turbine blade or at a corresponding position on the hub to collect vibration response signals during wind turbine blade operation. The directional microphone array is installed on the inner wall of the wind turbine nacelle with its main lobe aligned with the blade rotation plane to collect aerodynamic sound signals generated when the wind turbine blade rotates.

[0007] The single-channel multidimensional features include time-domain features, frequency-domain features, and time-frequency-domain features; The time-domain features are calculated directly from the amplitude variation sequence of the vibration digital signal and the sound digital signal over time, and are used to reflect the statistical characteristics and waveform morphology of the signal. For the preprocessed vibration digital signal and sound digital signal, perform the following operations respectively: Calculate the root mean square value of the amplitude at all sampling points of the signal; Find the maximum amplitude among all sampling points of the signal as the peak value; Based on the peak value and the root mean square value, the peak factor is obtained; The frequency domain features are obtained by performing spectral analysis on the signal and extracting the spectral centroid from the frequency distribution. The fundamental frequency and its harmonics are determined based on the passing frequency of the wind turbine blades; local peaks near each harmonic of the fundamental frequency and harmonics are searched in the power spectrum, and a narrow frequency band is preset with each harmonic frequency as the center; the energy in each narrow frequency band is calculated as the total harmonic energy of each harmonic component. Based on the total harmonic energy and the frequency range of the narrow band, the frequency component energy ratio of each harmonic component is obtained through analysis and calculation.

[0008] The time-frequency domain features can simultaneously reflect the energy distribution of the signal in time and frequency, and wavelet packet energy entropy is extracted using wavelet packet decomposition. The specific steps are as follows: Select a wavelet basis function and a decomposition level, and perform multi-level wavelet packet decomposition on the preprocessed signal. After decomposition, multiple frequency bands with equal bandwidth are obtained, and each frequency band corresponds to a set of wavelet packet coefficients. The energy of a frequency band is obtained based on wavelets within each band. The total energy of all frequency bands is obtained by merging the energy of all frequency bands. Based on the energy of the frequency band and the total energy of the frequency band, obtain the energy distribution probability of the frequency band; Wavelet calculations are performed on the energy distribution probability of each frequency band to obtain the wavelet packet energy entropy.

[0009] When calculating coherent features, the feature extraction and fusion module uses the following formula: in, Let x be the cross-power spectral density of the vibration digital signal and y be the sound digital signal. and These are the self-power spectral densities of the vibration signal x and the sound signal y, respectively.

[0010] The machine learning diagnostic model in the intelligent diagnostic module is one of support vector machine, random forest or lightweight neural network, and is trained offline using historical data. It is used to learn different mapping patterns of icing and changes in operating conditions in the fusion feature space.

[0011] It also includes an early warning and communication unit, which is connected to the edge computing and diagnostic unit, and is used to generate and report multi-level early warning information based on the diagnostic results and confidence levels.

[0012] A smart diagnostic method for blade icing based on acoustic and vibration feature fusion, applied to the aforementioned device, includes the following steps: S1: Data Synchronization Acquisition: During the operation of the wind turbine, the vibration sensing subunit and the acoustic sensing subunit are synchronously triggered to acquire the original vibration signal and the original sound signal for a period of time with the set synchronization accuracy. S2: Signal preprocessing and feature engineering: The acquired raw vibration signal and raw sound signal are preprocessed, and single-channel features are extracted respectively. At the same time, the coherence features of the two signals in the preset frequency band are calculated, and the single-channel features and coherence features are fused to form a fused feature vector. S3: Intelligent Diagnosis and Decision Making: Input the fused feature vector into a pre-trained machine learning diagnostic model to obtain the classification results of the leaf icing status and the diagnostic confidence level; S4: Early Warning and Information Reporting: Based on the diagnostic results and confidence level, trigger the corresponding level of early warning and report the diagnostic report through the communication network.

[0013] Furthermore, in step S4, the tiered early warning is generated according to preset rules, which include: Red alert rule: If the diagnosis is "thick ice", a red alert is triggered immediately, regardless of the confidence level. Yellow alert rule: If the diagnosis is "thin ice" and the confidence level is greater than the preset threshold, a yellow alert is triggered; Blue alert rule: If the diagnosis is "thin ice" but the confidence level is less than or equal to the preset threshold, a blue alert is triggered to prompt attention; No warning rule: If the diagnosis is "no ice", no warning will be triggered, and it will only be recorded and archived as a normal operating status. The rule matching engine supports dynamic configuration and updates of rules, and can adjust the warning thresholds and level classifications according to the actual needs of the wind farm.

[0014] The technical effects achieved by this invention are as follows: This invention integrates vibration features that are sensitive to changes in mass and acoustic features that are sensitive to changes in aerodynamic shape, and in particular introduces coherent features that characterize the physical relationship between the two, which can capture the weak dynamic and aeroacoustic coupling changes caused by early thin ice, effectively improving the recognition accuracy.

[0015] This invention employs a directional acoustic array and advanced digital filtering technology to effectively suppress ambient background noise. The machine learning model makes decisions by fusing multi-dimensional features, reducing sensitivity to disturbances of single features, resulting in a system with strong overall anti-interference capabilities.

[0016] This invention provides inherent redundancy through a multi-sensor architecture. When a single sensor fails, the system can degrade to using another sensor signal combined with a historical model for a rough diagnosis, resulting in a high fault tolerance rate. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structural components of the intelligent diagnostic device for blade icing; Figure 2 This is a schematic diagram illustrating the principle of acoustic and vibration signal feature fusion. Figure 3 This is a schematic diagram illustrating the working logic and early warning classification of the intelligent diagnostic model; Figure 4 This is the overall flowchart of the intelligent diagnostic method for leaf icing. Detailed Implementation

[0018] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0019] like Figures 1-3 As shown, a smart diagnostic device for blade icing based on acoustic and vibration feature fusion includes: A multimodal sensing unit is used to collect vibration response signals and aerodynamic sound signals. The multimodal sensing unit includes at least one low-frequency acceleration sensor and at least one directional microphone array. The low-frequency acceleration sensor is installed at the root of the wind turbine blade or at the corresponding position of the hub to collect vibration response signals when the wind turbine blade is running. The directional microphone array is installed on the inner wall of the wind turbine nacelle with the main lobe direction aligned with the blade rotation plane to collect aerodynamic sound signals generated when the wind turbine blade rotates. The signal conditioning and synchronous acquisition unit is electrically connected to the vibration sensing subunit and the acoustic sensing subunit, respectively, and is used to condition and synchronously convert the vibration response signal and the aeroacoustic signal into digital signals to obtain synchronous vibration digital signals and sound digital signals. The signal conditioning and synchronous acquisition unit includes: an independent signal conditioning circuit for amplifying and anti-aliasing filtering the original analog signal. The high-precision synchronous analog-to-digital converter uses a global synchronization clock to ensure that the sampling time of the vibration and acoustic channels are strictly aligned, with a synchronization error of less than 10 microseconds. The sampling frequency is set according to the Nyquist theorem, and the acoustic channel sampling rate is not less than 4kHz. The edge computing and diagnostic unit is connected to the signal conditioning and synchronous acquisition unit, and has built-in signal preprocessing module, feature extraction and fusion module and intelligent diagnostic module. The signal preprocessing module is used to preprocess the vibration digital signal and the sound digital signal. The feature extraction and fusion module is used to extract single-channel multidimensional features from the preprocessed vibration digital signal and sound digital signal, respectively; Single-channel multidimensional features include time-domain features, frequency-domain features, and time-frequency-domain features; Time-domain characteristics are calculated directly from the amplitude variation sequence of vibration digital signals and sound digital signals over time, and are used to reflect the statistical characteristics and waveform morphology of the signals; For the preprocessed vibration digital signal and sound digital signal, perform the following operations respectively: Squaring the amplitude of all sampling points of the signal yields the squared value of each point; then calculating the arithmetic mean of all squared values; finally, taking the square root of the mean value gives the root mean square value. Find the maximum amplitude among all sampling points of the signal and take it as the peak value; then calculate the ratio of the peak value to the root mean square value, and the result is the peak factor. Frequency domain features are extracted from the frequency distribution of a signal through spectral analysis. The specific steps are as follows: Based on the power spectrum, the frequency value of each frequency point is multiplied by the power value of that frequency point, and the sum is accumulated for all frequency points; at the same time, the power values ​​of all frequency points are accumulated; finally, the frequency-power weighted accumulated value is divided by the power accumulated value, and the result is the spectral centroid. Time-frequency domain features can simultaneously reflect the energy distribution of a signal in time and frequency. Wavelet packet decomposition is used to extract wavelet packet energy entropy. The specific steps are as follows: Select a wavelet basis function and a decomposition level, and perform multi-level wavelet packet decomposition on the preprocessed signal. After decomposition, multiple frequency bands with equal bandwidth are obtained, and each frequency band corresponds to a set of wavelet packet coefficients. The wavelet packet coefficients in each frequency band are squared and then summed to obtain the energy of that frequency band. The total energy is obtained by summing the energies of all frequency bands. Divide the energy of each frequency band by the total energy to obtain the energy distribution probability of that frequency band. The energy distribution probability of each frequency band is logarithmically calculated, base 2, and then multiplied by the energy distribution probability itself. The negative value is then summed over all frequency bands, and the result is the wavelet packet energy entropy. The coherence characteristics of the vibration digital signal and the sound digital signal within a preset frequency band are calculated. The coherence function is defined as follows: in, Let x be the cross-power spectral density of the vibration digital signal and y be the sound digital signal. and These are the self-power spectral densities of the vibration signal x and the sound signal y, respectively. And the single-channel features and coherent features are fused to form a multi-dimensional feature vector; The intelligent diagnosis module is used to input multi-dimensional feature vectors into a pre-trained machine learning diagnosis model to obtain the classification results of the icing status of the blades and the corresponding confidence scores. The machine learning diagnosis model in the intelligent diagnosis module is one of support vector machine, random forest or lightweight neural network, and is trained offline through historical data to learn different mapping patterns of icing and working condition changes in the fused feature space. The early warning and communication unit is connected to the edge computing and diagnostic unit to receive diagnostic results and confidence levels, generate graded early warning information according to preset strategies, and send the early warning information and raw feature data to the wind farm central monitoring system or cloud platform through wired or wireless communication.

[0020] like Figure 4 As shown, a smart diagnostic method for blade icing based on acoustic and vibration feature fusion includes the following steps: S1: Data Synchronization Acquisition: During the operation of the wind turbine, the vibration sensing subunit and the acoustic sensing subunit are synchronously triggered to acquire the original vibration signal and the original sound signal for a period of time with the set synchronization accuracy. S2: Signal preprocessing and feature engineering: The acquired raw vibration signal and raw sound signal are preprocessed, and single-channel features are extracted separately. At the same time, the coherence features of the two signals in the preset frequency band are calculated, and the single-channel features and coherence features are fused to form a fused feature vector. S3: Intelligent Diagnosis and Decision Making: Input the fused feature vector into the pre-trained machine learning diagnostic model to obtain the classification results of the leaf icing status and the diagnostic confidence level; S4: Early Warning and Information Reporting: Based on the diagnostic results and confidence level, trigger the corresponding level of early warning. The tiered early warning is generated according to preset rules, which include: Red alert rule: If the diagnosis is "thick ice", a red alert is triggered immediately, regardless of the confidence level. Yellow alert rule: If the diagnosis is "thin ice" and the confidence level is greater than the preset threshold, a yellow alert is triggered; Blue alert rule: If the diagnosis is "thin ice" but the confidence level is less than or equal to the preset threshold, a blue alert is triggered to prompt attention; No warning rule: If the diagnosis is "no ice", no warning will be triggered, and it will only be recorded and archived as a normal operating status; the preset threshold is 85%.

[0021] The rule matching engine supports dynamic configuration and updates of rules, and can adjust the warning threshold and level classification according to the actual needs of the wind field; The diagnostic report is then submitted via a communication network.

[0022] A computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned intelligent diagnostic method for blade icing.

[0023] The working principle of this invention is as follows: Synchronous acquisition of multimodal signals: Vibration sensing: Low-frequency acceleration sensors installed at the blade root or hub collect vibration response signals generated by changes in the blade's structural characteristics in real time; Sound perception: A directional microphone array installed in the nacelle with its main lobe aligned with the blades collects aerodynamic sound signals in real time caused by changes in the aerodynamic shape of the blades (such as increased surface roughness or changes in airfoil). Synchronization control: The signal conditioning and synchronous acquisition unit uses a global synchronization clock to ensure that the vibration and sound analog signals are strictly aligned during analog-to-digital conversion, laying the foundation for subsequent correlation analysis; Signal preprocessing and feature engineering: The edge computing unit first performs preprocessing such as filtering and noise reduction on the synchronized digital signal; Single-channel feature extraction: Extracting multi-dimensional features from vibration and sound signals respectively, including: Time-domain characteristics: such as root mean square value, reflecting energy magnitude; peak factor, reflecting impact characteristics; Frequency domain characteristics: such as the centroid of the spectrum, reflecting the dominant frequency shift and harmonic energy ratio, reflecting nonlinear changes; Time-frequency domain characteristics: such as wavelet packet energy entropy, which reflects the complexity of signal energy distribution; Dual-channel coherent feature extraction: Calculate the coherence function of vibration and sound signals within a specific frequency band; the coherence function value reflects the extent to which the sound signal is caused by vibration within a specific frequency band; icing can simultaneously change the vibration mode and the aerodynamic noise generation mechanism, thereby altering this coherence. Intelligent diagnosis and decision-making: Feature fusion: The extracted single-channel features (vibration features + sound features) are concatenated with the dual-channel coherent features to form a high-dimensional fused feature vector; this fused feature vector comprehensively reflects the physical state, aerodynamic state, and intrinsic relationship between the two of the blade. Model inference: The fused feature vector is input into a pre-trained machine learning model in the edge computing unit; this machine learning model has learned from a large amount of historical data and has mastered the different distribution patterns of the fused feature space under the states of "no ice", "thin ice" and "thick ice". Output: The model outputs the current icing state classification result and the corresponding confidence score; Tiered early warning and information reporting: The early warning and communication unit generates early warning information of corresponding levels according to the diagnostic results and confidence level, and in accordance with preset dynamic rules. Finally, the diagnostic reports and early warning information are uploaded to the wind farm central monitoring system or cloud platform via wired or wireless network for operation and maintenance personnel to make decisions.

[0024] Example 1: Embedded Online Diagnostic System Based on Megawatt-Level Wind Turbine Generators In this embodiment, the diagnostic device is specifically applied to a 2.5MW doubly-fed wind turbine generator located in a cold northern region to achieve real-time online monitoring and graded early warning of blade icing status. Detailed implementation method: Hardware deployment and connectivity: Multimodal sensing unit: Two low-frequency ICP-type accelerometers are installed on the flange near the blade root on the inner wall of the turbine hub to monitor the vibration in the direction of blade flapping and oscillation, respectively; at the same time, a directional microphone array consisting of four MEMS microphones is installed inside the rear end cover of the nacelle. The main lobe direction is aligned with the rotor rotation plane through a beamforming algorithm to suppress noise interference from other equipment in the nacelle. Signal conditioning and synchronous acquisition unit: A data acquisition card with 8-channel synchronous acquisition function is used to connect the analog signals from the accelerometer and microphone array; the sampling rate is set to 12.8kHz for the vibration channel and 51.2kHz for the sound channel, with a synchronization accuracy between channels better than 5 microseconds; Edge computing and diagnostic unit: It adopts an industrial-grade embedded industrial control computer built into the cabin control cabinet, running the Linux operating system, and has built-in software modules such as signal preprocessing, feature extraction and fusion, and intelligent diagnosis; Early warning and communication unit: The edge computing unit is connected to the wind farm ring network via an industrial Ethernet switch; Diagnostic procedure execution: Data acquisition: When the fan is running stably above the rated speed, the system automatically triggers synchronous acquisition every 10 minutes to obtain 30 seconds of raw vibration and sound signals; Feature extraction and fusion: Single-channel features: Calculate the root mean square value and peak factor in the time domain for vibration and sound signals respectively; after performing FFT transformation, extract the spectral centroid in the range of 0-1000Hz, and extract the blade passing frequency and its 2nd-5th harmonic energy ratio; perform 3-level wavelet packet decomposition and extract the wavelet packet energy entropy. Coherence characteristics: The focus is on calculating the average coherence coefficient of the vibration signal and the sound signal in the 50Hz-500Hz frequency band, and extracting the peak value of the coherence function and the frequency corresponding to the peak value in this frequency band; Fusion: The extracted vibration features, sound features, and acoustic-vibration coherence features are combined into a 30-dimensional feature vector; Intelligent diagnosis: The fused feature vector is input into a pre-trained random forest diagnostic model, and the model outputs the probability values ​​of three states: "no ice", "thin ice", and "thick ice". Tiered early warning: Based on the model output and confidence level, execute early warning rules: If the probability of outputting "thick ice" is greater than 70%, a red alert will be triggered, and maintenance personnel will be notified to immediately shut down the system for inspection. If the output "thin ice" probability is greater than 85%, a yellow alert will be triggered, and it is recommended to arrange a de-icing plan. If the probability of "thin ice" is between 50% and 85%, a blue alert will be triggered, indicating that attention should be paid to weather conditions and monitoring should be strengthened. Otherwise, record it as normal operating status.

[0026] Example 2: Condition monitoring and diagnostic device for small wind turbine or drone blades This embodiment miniaturizes and lightens the diagnostic device, applying it to the detection of icing on the blades of large industrial drones or small wind turbines. Detailed implementation method: Hardware integration and optimization: Multimodal sensing unit: Due to installation space and load limitations, a miniature MEMS accelerometer (patent type, weight <1g) and a single miniature electret microphone are selected; the microphone extends the acquisition port to the airflow stabilization area near the blade rotation plane through a thin sound guide tube to reduce local wind noise; Signal conditioning and synchronous acquisition unit: integrated into a dedicated signal processing chip the size of a fingernail, containing low-power amplification and filtering circuits and a high-precision ADC; Edge computing and diagnostic unit: It adopts a low-power ARM Cortex-M4 core MCU and runs a lightweight real-time operating system; due to limited computing resources, the machine learning model adopts a pre-trained support vector machine model. Early warning and communication unit: The diagnostic results are sent to the ground station or monitoring center via the drone's own data radio or the 4G / 5G communication module of the small wind turbine; Diagnostic procedure execution: Data acquisition: During aircraft cruise or wind turbine operation, short-term synchronization signals of 2 seconds are acquired at fixed intervals (e.g., every 30 seconds); Feature simplification and fusion: To match the computing power of the MCU, the feature set is simplified: Single-channel features: Only time-domain features (root mean square value, peak factor) and frequency-domain features (the center frequency and energy value of the three frequency bands with the highest energy proportion are extracted by calculating the power spectrum through fast Fourier transform) with low computational requirements. Coherence characteristics: Calculate the average coherence coefficient of vibration and sound signals in the main energy frequency bands; Fusion: Combine the above features into a single 10-dimensional feature vector; Intelligent diagnosis: Input the feature vector into the SVM model and directly output a binary classification result of "iced" or "not iced"; Emergency Response: For drones, if the diagnosis is "icing" and the confidence level exceeds 90%, the MCU directly sends an emergency landing command to the flight control system, and at the same time sends the "icing alarm - request landing" message via data transmission.

[0028] Example 3: End-to-end acoustic-vibration feature fusion diagnostic device based on deep learning This embodiment focuses on the intelligent upgrade of the feature extraction and fusion module, using a lightweight neural network to automatically learn deeper fusion features, reducing the workload of manual feature engineering. Detailed implementation method: Hardware deployment: Similar to Example 1, it is deployed on large wind turbine units, but the edge computing units have stronger GPU computing power; Intelligent diagnostic module upgrade: Model Architecture: The machine learning model in the intelligent diagnostic module adopts a two-input lightweight convolutional neural network (CNN) model; Vibration branch: Receives one-dimensional time-series data or two-dimensional time-spectrum diagrams of the preprocessed vibration signal; Audio branch: Receives one-dimensional time-series data or two-dimensional time-spectrum diagram of the pre-processed audio signal; Automatic feature fusion: The network performs convolution operations on vibration and sound data in the first few layers to extract their respective shallow features; in the middle layer of the network, a feature fusion layer is designed to concatenate or add the feature maps extracted by the two branches; this fusion layer not only concatenates single-channel features, but more importantly, through subsequent convolutional layers, the network can automatically learn the high-order nonlinear relationship between vibration and sound features, which is richer in information than manually calculated coherence functions; Output the classification probabilities of "no ice", "thin ice", and "thick ice"; Diagnostic procedure execution: Data acquisition and preprocessing: Vibration and sound signals are acquired simultaneously and preprocessed by DC removal, filtering, normalization, etc., and then they are cut into samples of fixed length. End-to-end diagnostics: Processed vibration and sound samples are directly input into a trained dual-input CNN model; the model automatically completes the entire process of feature extraction, fusion, and classification, and outputs diagnostic results and confidence scores, reducing manual feature engineering and realizing automatic mapping from raw data to diagnostic results; Model Update: Edge computing units can use newly collected and verified data to incrementally learn or fine-tune the model, enabling the model to better adapt to changes such as the environment of a specific wind field and the aging of the unit, thus achieving continuous model optimization. Warning: Subsequent warnings and information reporting rules are the same as in Example 1, and graded warnings are triggered based on the classification results and confidence levels output by the model.

[0030] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A smart diagnostic device for blade icing based on acoustic and vibration feature fusion, characterized in that, include: A multimodal sensing unit is used to acquire vibration response signals and aeroacoustic signals; The signal conditioning and synchronous acquisition unit is connected to the multimodal sensing unit and is used to condition and synchronously convert the vibration response signal and aeroacoustic signal to obtain synchronous vibration digital signal and sound digital signal. The edge computing and diagnostic unit is connected to the signal conditioning and synchronous acquisition unit and has a built-in signal preprocessing module, feature extraction and fusion module, and intelligent diagnostic module. The signal preprocessing module is used to preprocess the vibration digital signal and the sound digital signal. The feature extraction and fusion module is used to extract single-channel multidimensional features from the preprocessed vibration digital signal and sound digital signal, respectively. Calculate the coherence features of the vibration digital signal and the sound digital signal within a preset frequency band, and fuse the single-channel features and coherence features to form a multi-dimensional feature vector; The intelligent diagnostic module is used to input the multidimensional feature vector into a pre-trained machine learning diagnostic model to obtain the classification results of the leaf icing state and the corresponding confidence level.

2. The intelligent diagnostic device for blade icing based on acoustic and vibration feature fusion according to claim 1, characterized in that: The multimodal sensing unit includes at least one low-frequency acceleration sensor and at least one directional microphone array. The low-frequency acceleration sensor is installed at the root of the wind turbine blade or at a corresponding position on the hub to collect vibration response signals during wind turbine blade operation. The directional microphone array is installed on the inner wall of the wind turbine nacelle with its main lobe aligned with the blade rotation plane to collect aerodynamic sound signals generated when the wind turbine blade rotates.

3. The intelligent diagnostic device for blade icing based on acoustic and vibration feature fusion according to claim 1, characterized in that: The single-channel multidimensional features include time-domain features, frequency-domain features, and time-frequency-domain features; The time-domain features are calculated directly from the amplitude variation sequence of the vibration digital signal and the sound digital signal over time, and are used to reflect the statistical characteristics and waveform morphology of the signal. For the preprocessed vibration digital signal and sound digital signal, perform the following operations respectively: Calculate the root mean square value of the amplitude at all sampling points of the signal; Find the maximum amplitude among all sampling points of the signal as the peak value; Based on the peak value and the root mean square value, the peak factor is obtained; The frequency domain features are obtained by performing spectral analysis on the signal and extracting the spectral centroid from the frequency distribution. The fundamental frequency and its harmonics are determined based on the passing frequency of the wind turbine blades; local peaks near each harmonic of the fundamental frequency and harmonics are searched in the power spectrum, and a narrow frequency band is preset with each harmonic frequency as the center; the energy in each narrow frequency band is calculated as the total harmonic energy of each harmonic component. Based on the total harmonic energy and the frequency range of the narrow band, the frequency component energy ratio of each harmonic component is obtained through analysis and calculation.

4. The intelligent diagnostic device for blade icing based on acoustic and vibration feature fusion according to claim 3, characterized in that: The time-frequency domain features can simultaneously reflect the energy distribution of the signal in time and frequency, and wavelet packet energy entropy is extracted using wavelet packet decomposition. The specific steps are as follows: Select a wavelet basis function and a decomposition level, and perform multi-level wavelet packet decomposition on the preprocessed signal. After decomposition, multiple frequency bands with equal bandwidth are obtained, and each frequency band corresponds to a set of wavelet packet coefficients. The energy of a frequency band is obtained based on wavelets within each band. The total energy of all frequency bands is obtained by merging the energy of all frequency bands. Based on the energy of the frequency band and the total energy of the frequency band, obtain the energy distribution probability of the frequency band; Wavelet calculations are performed on the energy distribution probability of each frequency band to obtain the wavelet packet energy entropy.

5. The intelligent diagnostic device for blade icing based on acoustic and vibration feature fusion according to claim 1, characterized in that: When calculating coherent features, the feature extraction and fusion module uses the following formula: in, Let x be the cross-power spectral density of the vibration digital signal and y be the sound digital signal. and These are the self-power spectral densities of the vibration signal x and the sound signal y, respectively.

6. The intelligent diagnostic device for blade icing based on acoustic and vibration feature fusion according to claim 1, characterized in that: The machine learning diagnostic model in the intelligent diagnostic module is one of support vector machine, random forest or lightweight neural network, and is trained offline using historical data to learn different mapping patterns of icing and operating condition changes in the fused feature space.

7. The intelligent diagnostic device for blade icing based on acoustic and vibration feature fusion according to claim 1, characterized in that: It also includes an early warning and communication unit, which is connected to the edge computing and diagnostic unit, and is used to generate and report multi-level early warning information based on the diagnostic results and confidence levels.

8. A method for intelligent diagnosis of blade icing based on acoustic and vibration feature fusion, applied to the device as described in any one of claims 1-7, characterized in that, Includes the following steps: S1: Data Synchronization Acquisition: During the operation of the wind turbine, the vibration sensing subunit and the acoustic sensing subunit are synchronously triggered to acquire the original vibration signal and the original sound signal for a period of time with the set synchronization accuracy. S2: Signal preprocessing and feature engineering: The acquired raw vibration signal and raw sound signal are preprocessed, and single-channel features are extracted respectively. At the same time, the coherence features of the two signals in the preset frequency band are calculated, and the single-channel features and coherence features are fused to form a fused feature vector. S3: Intelligent Diagnosis and Decision Making: Input the fused feature vector into a pre-trained machine learning diagnostic model to obtain the classification results of the leaf icing status and the diagnostic confidence level; S4: Early Warning and Information Reporting: Based on the diagnostic results and confidence level, trigger the corresponding level of early warning and report the diagnostic report through the communication network.

9. The method according to claim 8, characterized in that: In step S4, the tiered early warning is generated according to preset rules, which include: Red alert rule: If the diagnosis is "thick ice", a red alert is triggered immediately, regardless of the confidence level. Yellow alert rule: If the diagnosis is "thin ice" and the confidence level is greater than the preset threshold, a yellow alert is triggered; Blue alert rule: If the diagnosis is "thin ice" but the confidence level is less than or equal to the preset threshold, a blue alert is triggered, prompting attention; No warning rule: If the diagnosis is "no ice", no warning will be triggered, and it will only be recorded and archived as a normal operating status. The rule matching engine supports dynamic configuration and updates of rules, and can adjust the warning thresholds and level classifications according to the actual needs of the wind farm.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 8 or 9.

Images