Icing detection method and system for blades of wind generating set
By integrating multimodal perception and edge intelligent detection technologies, and fusing data from acoustic signatures, infrared sensors, and environmental sensors, the system employs VMD and lightweight ResNet models for real-time classification of icing conditions. This solves the reliability and cost-effectiveness issues of wind turbine blade icing detection, achieving efficient and low-cost icing detection.
Patent Information
- Application Number
- CN202511390753.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-26
- Publication Date
- 2026-01-13
AI Technical Summary
Existing wind turbine blade icing detection technologies suffer from problems such as limited detection methods, severe environmental interference, high false alarm rate, poor real-time performance of algorithms, and high deployment and maintenance costs, making it difficult to achieve reliable, efficient, and low-cost applications.
A multimodal perception system was constructed, which integrates data from acoustic fingerprints, infrared sensors, and environmental sensors. The VMD algorithm was used to extract anti-interference features, and a lightweight ResNet model was used for real-time classification of icing conditions. A greedy algorithm was combined to optimize sampling and communication strategies to achieve intelligent edge detection.
It achieves accurate perception of multi-dimensional icing status, reduces false alarm rate, meets real-time early warning requirements, and reduces system power consumption and deployment costs, making it economically feasible for large-scale application.
Smart Images
Figure CN121322320A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power generation technology, specifically to a method and system for detecting icing on wind turbine blades. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Wind power is a widely used clean and renewable energy source. In cold and humid climates, wind turbine blades are highly susceptible to icing. Icing alters the aerodynamic shape of the blades, leading to a significant decrease in the aerodynamic efficiency of the unit and a loss of power generation. At the same time, icing can cause uneven blade loads and rotational mass imbalances, resulting in increased unit vibration. This not only accelerates component fatigue damage but may also lead to serious safety accidents such as shutdowns and tower collapses.
[0004] For detecting icing on wind turbine blades, existing technologies mainly rely on information from a single sensor or simple logical judgments, which have the following limitations: Wind turbine blade icing is a critical issue affecting the operating efficiency and safety of wind turbines. Existing technologies have the following shortcomings: The current detection methods are limited and cannot comprehensively perceive the icing state. Existing technologies mostly use single types of sensors, such as vibration sensors or infrared thermal imagers. Vibration sensors indirectly infer mass changes by analyzing changes in the natural frequency of the blades, but they are not sensitive to the initial stage of icing or asymmetric icing. Infrared thermal imagers can directly observe surface temperature anomalies, but they cannot detect the mass changes and aerodynamic noise changes caused by icing. Icing is a complex process involving changes in multiple physical fields such as mass, morphology, and temperature. Relying on a single information source cannot build a comprehensive and accurate perception model, resulting in insufficient detection reliability.
[0005] Severe environmental interference leads to a high false alarm rate. The harsh operating environment of wind turbines makes sensor signals susceptible to strong interference from complex environmental factors. For example, signals collected by acoustic or vibration sensors are mixed with wind noise caused by strong winds, resulting in an extremely low signal-to-noise ratio. Infrared thermal imaging is severely affected by factors such as ambient temperature, sunlight, and rain, causing huge fluctuations in temperature readings and making it difficult to effectively distinguish between minute temperature changes caused by icing and environmental thermal noise. Current technologies lack effective environmental noise suppression and compensation mechanisms, resulting in a persistently high false alarm rate under conditions of wind speed changes and rain / snow.
[0006] The algorithm suffers from poor real-time performance, failing to meet the demands of online early warning. To handle complex signal analysis tasks (such as time-frequency analysis and image recognition), existing solutions typically employ a mode of remotely transmitting sensor data to a cloud server for processing. This mode is limited by network bandwidth, latency, and the queuing and scheduling of cloud computing resources, resulting in system response delays often reaching several seconds or even minutes. This latency cannot meet the requirements for real-time monitoring and rapid early warning of icing, thus missing the optimal intervention opportunity to prevent ice accumulation.
[0007] The high cost of system deployment and maintenance makes large-scale application difficult. To achieve multi-dimensional perception, some solutions attempt to integrate multiple sensors, but this leads to a sharp increase in system complexity, power consumption, and cost. Data synchronization, calibration, and maintenance of multi-sensor systems are cumbersome. Furthermore, cloud-based computing solutions not only incur continuous communication costs, but their high power consumption also makes the power supply costs for deployment in remote wind farms prohibitively high.
[0008] In summary, existing wind turbine blade icing detection technologies have inherent defects in terms of sensing dimensions, anti-interference ability, real-time performance, and economy, making it difficult to achieve reliable, efficient, and low-cost applications in real-world industrial scenarios. Summary of the Invention
[0009] To address the technical problems mentioned above, this invention provides a method and system for detecting icing on wind turbine blades, constructing an intelligent edge detection system for blade icing that integrates multimodal perception, adaptive anti-interference, edge intelligent decision-making, and dynamic power consumption management. By fusing acoustic signature, infrared, and environmental sensor data, and utilizing the VMD algorithm to extract anti-interference features, a lightweight ResNet model with injected environmental parameters and integrated attention mechanism is employed to achieve real-time and accurate classification of icing status directly at the terminal. Simultaneously, a greedy algorithm dynamically optimizes sampling and communication strategies, ultimately achieving a highly reliable and scalable icing detection solution with extremely low latency and power consumption.
[0010] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for detecting icing on wind turbine blades, comprising the following steps: Acquire and preprocess the voiceprint signal, infrared thermal imaging signal and environmental parameters, determine the environmental interference compensation factor based on the environmental parameters, and perform adaptive weighted fusion processing on the voiceprint signal and infrared thermal imaging signal to obtain the fused multimodal feature vector. The acoustic signature signal in the multimodal feature vector is processed by the variational mode decomposition algorithm to obtain several intrinsic mode functions. The acoustic signature features are extracted from them, including high-frequency energy ratio and low-frequency kurtosis. Based on the infrared thermal imaging signal in the multimodal feature vector, the temperature gradient index, which characterizes the degree of abnormality in the temperature distribution on the blade surface, is calculated. The voiceprint features, temperature gradient index and environmental parameters are used to construct a feature vector, and the prediction model is used to output the icing probability value. The icing threshold is dynamically adjusted based on the current environmental parameters, and the icing probability value is compared with the dynamically adjusted threshold. If the probability value is greater than the threshold, the blades are determined to be icy, and an early warning message is generated.
[0011] Furthermore, adaptive weighted fusion processing is performed on the voiceprint signal and the infrared thermal imaging signal, specifically as follows: Based on environmental parameters, the interference intensity coefficient is estimated using a linear regression model. The weighting coefficients of acoustic signature features and infrared features during fusion are dynamically adjusted based on the interference intensity coefficient. Based on the formula: The environmental vector and the voiceprint / infrared signal are weighted and fused into the main feature vector, where, The softmax function is used to normalize the weight coefficients. , This is the original sensor feature vector. This represents the environmental feature vector.
[0012] Furthermore, voiceprint features are extracted, specifically: The variational mode decomposition algorithm is used to process the acoustic signature signal in the multimodal feature vector to obtain... One intrinsic mode function; from Among the intrinsic mode functions, the intrinsic mode function representing the high-frequency component is selected, its energy in the set frequency band is calculated, and the high-frequency energy ratio is obtained by comparing it with the total energy of the entire frequency band. from Among the intrinsic mode functions, the intrinsic mode function representing the main vibration frequency band is selected, and its kurtosis value in the time domain is calculated as the low-frequency kurtosis.
[0013] Furthermore, the temperature gradient index is calculated as follows: the temperature distribution matrix is obtained from the infrared thermal imaging signal in the multimodal feature vector, the temperature difference matrix is determined by combining it with the reference temperature, and the maximum and minimum values of the temperature difference matrix are obtained. The temperature gradient index, which characterizes the degree of abnormality in the temperature distribution on the blade surface, is obtained by the ratio of the difference between the maximum and minimum values to the reference temperature.
[0014] Furthermore, the prediction model is an improved ResNet-18 model with an integrated attention mechanism. It receives feature vectors containing high-frequency energy ratio, low-frequency kurtosis, temperature gradient exponent, and environmental parameters through the input layer. The model parameters are compressed to 8-bit precision for deployment on edge computing devices.
[0015] Furthermore, the icing determination threshold is dynamically adjusted based on the current environmental parameters, as shown in the following formula: ; in, Environmental sensitivity coefficient, This indicates the adjusted threshold. Indicates the baseline threshold. This represents the norm of the environment vector.
[0016] Furthermore, during the acquisition of voiceprint signals, infrared thermal imaging signals, and environmental parameters, dynamic power consumption optimization is performed using a greedy optimization algorithm, specifically as follows: Establish a system energy consumption model with sensor sampling rate, communication interval, and sleep duration as variables; With the detection performance requirements as constraints, the optimal parameter combination that minimizes the total energy consumption of the system is found through greedy optimization algorithm iteratively, and the working state of the sensor is dynamically configured accordingly.
[0017] A second aspect of the present invention provides an icing detection system for wind turbine blades, comprising: The multimodal sensor module is configured to acquire acoustic signature signals, infrared thermal imaging signals, and environmental parameters. The signal processing module is configured to: preprocess the acquired acoustic signature signal, infrared thermal imaging signal and environmental parameters, determine the environmental interference compensation factor based on the environmental parameters, and perform adaptive weighted fusion processing on the acoustic signature signal and infrared thermal imaging signal to obtain the fused multimodal feature vector. The signal processing module is also configured to: process the acoustic signature signal in the multimodal feature vector using the variational mode decomposition algorithm to obtain several intrinsic mode functions, extract the acoustic signature features, including high frequency energy ratio and low frequency kurtosis; and calculate the temperature gradient index characterizing the degree of abnormality in the temperature distribution on the blade surface based on the infrared thermal imaging signal in the multimodal feature vector. The signal processing module is also configured to: construct a feature vector from the acoustic signature features, temperature gradient index and environmental parameters, and output the icing probability value using the prediction model; The signal processing module is also configured to dynamically adjust the icing judgment threshold according to the current environmental parameters, compare the icing probability value with the dynamically adjusted threshold, and if the probability value is greater than the threshold, determine that the blade is icing and generate an early warning message.
[0018] A third aspect of the present invention provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the above-described method for detecting icing on wind turbine blades.
[0019] A fourth aspect of the present invention provides an electronic device comprising at least one processor and a memory connected to the processor, the memory being used to store a computer program; the processor being used to execute the computer program, enabling the electronic device to implement the above-described method for detecting icing on wind turbine blades.
[0020] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects: 1. By simultaneously collecting multimodal data such as acoustic vibration, infrared thermal imaging, and environmental parameters, a multidimensional feature vector was constructed. This vector can comprehensively perceive the icing state from multiple physical levels, such as mass change, surface morphology, and temperature distribution. This overcomes the limitation of the perception dimension of a single sensor, greatly reduces the false negative rate, and improves the accuracy of state determination.
[0021] 2. By introducing environmental parameters as compensation factors into the feature fusion and decision-making process, and through an adaptive weight adjustment algorithm and a dynamic threshold determination mechanism, the system can compensate for the interference of wind noise and environmental temperature and humidity fluctuations on soundprints and infrared signals in real time, giving the system strong environmental robustness and significantly reducing false alarms caused by severe weather.
[0022] 3. By adopting the more computationally efficient VMD algorithm and the improved lightweight ResNet-18 model, and with the help of optimization techniques such as model quantization, this solution deploys the core algorithm on the edge computing terminal, realizing local real-time data processing and decision-making, completely avoiding the latency caused by cloud transmission, with an end-to-end response time of less than 1 second, and can provide timely and effective early warning information for operation and maintenance personnel.
[0023] 4. By employing a dynamic power consumption optimization algorithm to intelligently adjust the sensor sampling rate and communication sleep strategy, the overall system power consumption is significantly reduced, extending the device's battery life. Combined with a highly integrated and lightweight edge computing solution, the reliance on communication networks and cloud resources is reduced, effectively controlling the hardware, deployment, and maintenance costs of the entire system, making it economically feasible for large-scale application in wind farms. Attached Figure Description
[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0025] Figure 1This is a schematic diagram of the overall process of the wind turbine blade icing detection method provided in one or more embodiments of the present invention. Detailed Implementation
[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0028] As described in the background section, existing wind turbine blade icing detection technologies have inherent defects in terms of sensing dimensions, anti-interference capabilities, real-time performance, and economy, making it difficult to achieve reliable, efficient, and low-cost applications in actual industrial scenarios.
[0029] The following embodiments present a method and system for detecting icing on wind turbine blades, constructing an intelligent edge detection system for blade icing that integrates multimodal perception, adaptive anti-interference, edge intelligent decision-making, and dynamic power consumption management. By fusing acoustic signature, infrared, and environmental sensor data, and utilizing the VMD algorithm to extract anti-interference features, a lightweight ResNet model with injected environmental parameters and integrated attention mechanism is employed to achieve real-time and accurate classification of icing status directly at the terminal. Simultaneously, a greedy algorithm dynamically optimizes sampling and communication strategies, ultimately achieving a highly reliable and scalable icing detection solution with extremely low latency and power consumption.
[0030] Example 1: like Figure 1 As shown, the method for detecting icing on wind turbine blades includes the following steps: Acquire and preprocess the voiceprint signal, infrared thermal imaging signal and environmental parameters, determine the environmental interference compensation factor based on the environmental parameters, and perform adaptive weighted fusion processing on the voiceprint signal and infrared thermal imaging signal to obtain the fused multimodal feature vector. The acoustic signature signal in the multimodal feature vector is processed by the variational mode decomposition algorithm to obtain several intrinsic mode functions. The acoustic signature features are extracted from them, including high-frequency energy ratio and low-frequency kurtosis. Based on the infrared thermal imaging signal in the multimodal feature vector, the temperature gradient index, which characterizes the degree of abnormality in the temperature distribution on the blade surface, is calculated. The voiceprint features, temperature gradient index and environmental parameters are used to construct a feature vector, and the prediction model is used to output the icing probability value. The icing threshold is dynamically adjusted based on the current environmental parameters, and the icing probability value is compared with the dynamically adjusted threshold. If the probability value is greater than the threshold, the blades are determined to be icy, and an early warning message is generated.
[0031] (1) Multimodal data acquisition and preprocessing.
[0032] Objective: To eliminate environmental noise interference by synchronously collecting blade status data through acoustic fingerprint sensors, infrared thermal imaging, and environmental parameters (temperature, humidity, wind speed).
[0033] Step 1.1 Acoustic signal acquisition: A wideband acoustic sensor (100Hz-80kHz) was used to acquire blade vibration and aerodynamic noise signals; sampling frequency This satisfies the Nyquist sampling theorem.
[0034] Step 1.2 Infrared Thermal Imaging Data Acquisition: Use a non-contact infrared sensor (640 x 480 resolution) to acquire the temperature distribution matrix on the blade surface. (Unit: °C) Temperature anomaly area is defined as: (in (This represents the average temperature under icing-free conditions).
[0035] Step 1.3 Environmental Parameter Fusion: Input Temperature and Humidity and wind speed Construct environmental feature vectors Environmental factors (such as wind noise introduced by high wind speeds or temperature fluctuations caused by temperature and humidity) can distort sensor signals. Fusion technology can compensate for these disturbances in real time. The model dynamically adjusts the thresholds, weights, and filtering parameters of signal processing based on environmental data to achieve adaptive denoising and improve system robustness.
[0036] The specific steps for environmental parameter fusion are as follows: Fusion Step 1: Acquire the raw signal and extract environmental parameters as compensation input; Fusion Step 2: Calculate the environmental interference compensation factor and use a linear regression model to estimate the interference intensity, such as the interference intensity coefficient. (in (Assessment weights, determined using historical data). Fusion Step 3: The environmental vector and the voiceprint / infrared signal are weighted and fused into the main feature vector. The fusion formula is as follows: ; in, (The softmax function is used to normalize the weights, ensuring...) ), This is the original sensor feature vector.
[0037] This step is used to address environmental interference issues: in high wind speeds... Down, To reduce the weight of the acoustic signature signal and suppress wind noise, the temperature gradient threshold is adjusted to correct the infrared data during temperature and humidity fluctuations. This mechanism dynamically filters out noise components through adaptive weighting. Experiments show that the signal-to-noise ratio is improved by 15%-25% after fusion, significantly reducing the false alarm rate.
[0038] Traditional filtering (such as fixed-threshold low-pass filtering) assumes that the noise is stable or has known statistical characteristics, but the noise at the wind turbine site is non-stationary, and its intensity and characteristics are strongly correlated with the environment: For example, wind noise is directly proportional to wind speed; the higher the wind speed, the greater the noise, and it may also excite resonance at specific frequencies. For example, infrared temperature interference: ambient temperature, humidity, and solar radiation directly determine the heat exchange process on the blade surface, and their fluctuations can completely mask the minute temperature changes caused by icing.
[0039] Therefore, noise in this scheme is a variable driven by environmental parameters. Directly utilizing these driving factors (environmental data) to suppress noise is the most direct and effective way to solve the problem at its root. Achieving active compensation through noise reduction via environmental parameter fusion allows for the prediction of noise's impact based on environmental parameters before it even occurs, enabling proactive mitigation or adjustment.
[0040] (2) Voiceprint signal decomposition and feature extraction based on VMD.
[0041] Objective: To reduce computational complexity and extract icing features by replacing traditional EEMD (Empirical Mode Decomposition) with Variational Mode Decomposition (VMD). This includes the following steps.
[0042] Step 2.1 Decomposition using the VMD algorithm: Decompose the voiceprint signal Decompose the K intrinsic mode functions (IMFs) using the variational mode decomposition (VMD) algorithm: ; in, For the first One modal function, This is the residual term. Modal number. Typically, a balance parameter of 3-5 is chosen. The default value is 2000. Step 2.2 Feature extraction; Step 2.2.1 Calculate the High Frequency Energy Ratio (HER): The High Frequency Energy Ratio (HER) is used to quantify the energy proportion of high-frequency components in a signal, reflecting the high-frequency characteristics of the signal. First, for each mode function... Calculate its energy spectral density Then calculate the ratio of high-frequency energy to full-frequency energy: ; in, The denominator represents the energy spectral density at frequency f; the denominator represents the total energy across the entire frequency band; the numerator represents the total energy in the high-frequency band (5-20kHz), which corresponds to high-frequency characteristics such as cracking; HER is calculated by analyzing all mode functions. The energy spectrum was calculated.
[0043] Step 2.2.2 Calculate the low-frequency kurtosis (LKC): Low-frequency kurtosis (LKC) is used to quantify the peak level of a signal, helping to detect low-frequency anomalies caused by icing. First, from each mode function... Extract the low-frequency component and calculate its kurtosis: ; in, Indicates the first Modal functions Sample values in the low and mid-frequency range, and These are the mean and standard deviation of the low-frequency signal, respectively. This represents the number of sample points. This formula quantifies the peak strength of the signal, used to detect low-frequency anomalies caused by icing. Both HER and LKC are based on per-mode function. Calculated. Specifically: HER: First, calculate the energy spectrum of each mode function, especially the energy in the high-frequency part, and then compare it with the energy of the whole frequency band to obtain the high-frequency energy ratio.
[0044] LKC: Extracts the low-frequency portion of each mode function, calculates the kurtosis of that portion, and quantizes the peak value of the signal.
[0045] Step 2.2.3 Infrared Signal Processing: Calculating the Temperature Gradient Index (TGI) The infrared signal is processed, and the Temperature Gradient Index (TGI) is defined as follows: ; in, Represents the temperature difference matrix. This represents the reference temperature, used for normalization and quantification of temperature gradient anomalies.
[0046] Reference temperature The results are obtained by infrared measurement under conditions of no icing and no abnormalities. They can be manually set as a reference value for the average temperature of the equipment surface or the ambient temperature, or they can be automatically determined by the average of multiple samples.
[0047] When the infrared sensor collects the temperature distribution of a two-dimensional region, the temperature difference matrix... Temperature of each pixel within the region Compared with reference temperature The difference consists of: ; If the infrared signal is only a single-point temperature sequence, the deviation from the reference temperature is calculated in the time dimension, and an equivalent temperature difference matrix is constructed using the time points. Under regional temperature distribution conditions, the temperature difference matrix reflects the differences in heat distribution at different locations; under single-point temperature conditions, the temperature difference matrix reflects the differences in temperature changes over time.
[0048] (3) Icing state determination of lightweight deep learning models.
[0049] Objective: To achieve real-time classification of icing conditions by fusing voiceprint features, temperature gradients, and environmental parameters using an improved ResNet model.
[0050] Step 3.1 Feature Vector Construction: Input feature vector (6-dimensional); in, Represents the input vector (6-dimensional). Indicates feature extraction; Represents the environment vector.
[0051] Step 3.2 Improve the ResNet model: For scenarios involving icing state determination, the standard ResNet model was improved in both architecture and parameters to enhance its robustness and real-time performance in low-sample, high-noise environments.
[0052] Step 3.2.1 improves the model architecture by simplifying the deep structure of standard ResNet (usually 50+ layers) to an 18-layer version, reducing the number of residual blocks (from multiple blocks to 3), and adding an attention mechanism module (SEBlock) to each residual block to dynamically highlight icing-related features (such as temperature gradients and low-frequency kurtosis). The formula is as follows: ; in, Indicates residual output, represents the input features; SE represents the squeeze-excitation module, used for channel attention weight calculation.
[0053] Step 3.2.1 parameter improvements: Environmental parameters were introduced as auxiliary inputs, expanded into a 6-dimensional vector at the input layer, and the learning rate (reduced from 0.001 to 0.0005) and batch size (reduced from 32 to 16) were optimized to adapt to the memory limitations of embedded devices. Additionally, a dropout layer (rate 0.3) was added to prevent overfitting. These improvements enable the model to handle "icing state determination" scenarios: the attention mechanism focuses on multimodal fusion features (such as temperature anomalies under high wind speeds), the shallow architecture reduces computation (floating-point operations are reduced from hundreds of millions to millions), and ensures accurate differentiation between healthy and icing states under noise interference (AUC > 0.95 in experiments).
[0054] Step 3.2.2 Design the network structure: Set up an input layer (6-dimensional), hidden layers (3 residual blocks, each containing 2 convolutional layers, ReLU activation), and an output layer (Sigmoid outputting the icing probability).
[0055] Step 3.2.3 Develop a training strategy, using binary cross-entropy loss as the loss function, and adding an L2 regularization term to prevent overfitting: ; in, Indicates the loss value. This represents the true label (0 / 1). This represents the predicted probability (0-1).
[0056] Step 3.2.4 addresses the high latency issue: Existing methods rely on cloud processing, resulting in latency >5s. This solution implements edge computing using a lightweight ResNet model, eliminating the need for cloud data transmission; the simplified architecture reduces inference time (single inference <0.5s), and combined with dynamic power optimization (adjusting the sampling rate to reduce data volume), end-to-end latency is controlled to <1s. Specific mechanisms include model parameter compression (quantization to 8-bit) and batch processing optimization, ensuring real-time response to alert requirements on low-power hardware.
[0057] Step 3.3 Adaptive threshold determination: The icing determination threshold is dynamically adjusted based on environmental parameters, as shown in the following formula: ; in, This indicates the adjusted threshold, which defaults to 0.5 based on prior knowledge. This represents the baseline threshold, with a value of 0.5. This represents the environmental sensitivity coefficient, with a value chosen between 0.1 and 0.5. This represents the norm of the environment vector.
[0058] Based on the norm of the real-time environment vector The decision threshold is dynamically adjusted, and the probability value output by the model is compared with the dynamic threshold. The comparison ultimately leads to a binary decision of health or icing, triggering an early warning.
[0059] (4) Dynamic power consumption optimization algorithm: Objective: To extend device battery life by balancing sensor sampling rate and communication power consumption using a greedy optimization algorithm. This includes the following steps.
[0060] Step 4.1 Energy consumption modeling: Based on the sampling rate Communication interval Hibernation duration Construct the energy consumption function: ; in, The energy consumption coefficient is determined experimentally. This shows the total energy consumption (unit: W). Indicates the sampling rate (kHz). Indicates the communication interval. Indicates the duration of hibernation.
[0061] Step 4.2 Greedy Optimization: Iteratively adjust parameters from top to bottom, selecting the combination that minimizes energy consumption at each step. Constraints kHz, s, s (constraints are calculated through analysis of experimental and historical data).
[0062] The greedy optimization algorithm runs continuously, dynamically optimizing the sensor sampling rate, communication interval, and sleep duration based on the current state (such as whether icing is detected or the severity of the environment), minimizing the total system power consumption and extending the device's battery life while ensuring performance.
[0063] Verification experiment.
[0064] Based on a typical wind turbine testing environment, a GW155-4.5MW unit from a certain brand was used as the test platform to simulate winter wind farm conditions (ambient temperature -10℃ to 5℃, wind speed 6 to 15 m / s, humidity 70% to 90%). Test data were derived from a combination of laboratory simulation and field verification, including 500 healthy samples and 300 icing samples (thickness 1 to 10 mm). Cross-validation was used to evaluate performance. All parameters were calibrated according to the methods described in this scheme, and the hardware used was an STM32H743 processor and a LoRa communication module.
[0065] In actual deployment, the device is installed at the 3 o'clock position at the blade root (50-100cm from the blade root), using a magnetic base and industrial adhesive for double fixation, ensuring a tensile strength greater than 500N and meeting IP67 protection standards. The test unit is a GW155-4.5MW from a certain brand, with one multimodal sensor module (integrating acoustic fingerprint unit, infrared unit, and environmental parameter unit) installed at the root of each blade. The sampling rate is set to 192kHz (dynamically adjustable to 96-384kHz according to wind speed), and the sampling duration is 10s / cycle. Data interaction is achieved through LoRa wireless communication, uploading an icing status summary (including probability p, threshold θ, and key feature vectors) once per hour. In low-power mode (sampling rate reduced to 96kHz, communication interval 3600s, sleep duration 8s), the device's battery life exceeds 6 months (based on a 2000mAh battery test). The host computer system supports remote parameter reconfiguration, such as adjusting the fusion weight β (default 0.5, dynamically updated according to environmental interference intensity α). This configuration ensures convenient contactless installation without interrupting wind turbine operation, with a deployment time of less than 30 minutes per blade.
[0066] In field tests, this solution successfully detected icing as thin as 1 mm (based on a combined criterion of low-frequency kurtosis LKC > 4.5 and temperature gradient exponent TGI > 0.15), providing an early warning 2 hours earlier than traditional methods (traditional infrared methods require icing accumulation of more than 5 mm to trigger). Under strong interference conditions with wind speeds of 12 m / s, the false alarm rate was controlled at <5% (based on statistics from 1000 sets of simulated data; the signal-to-noise ratio was improved by 20 dB after fusing the environmental vector e); the response time was end-to-end latency <1 s, including data acquisition (0.2 s), VMD decomposition and feature extraction (0.3 s), improved ResNet inference (0.2 s), and threshold adjustment (0.1 s). For example, in an icing simulation experiment, when the ambient temperature T_env = -5℃ and the wind speed V_wind = 10m / s, the model outputs an icing probability p = 0.92 and a dynamic threshold θ = 0.65 (baseline θ_base = 0.5, sensitivity coefficient γ = 0.3), triggering an immediate warning and reporting an event packet (including timestamp, feature summary, and location coordinates). These effects are achieved through multimodal fusion and edge computing, ensuring robustness and real-time performance under complex wind fields.
[0067] To verify the superiority of this scheme, its performance was compared with that of traditional infrared methods (relying on single thermal imaging and cloud processing) and acoustic fingerprint sensor methods (single audio analysis without environmental fusion). Data were obtained from the same test environment (wind speed 8-12 m / s, ice thickness 1-5 mm, sample size 800 groups), and AUC, F1 score and other indicators were used for evaluation. The results are shown in Table 1.
[0068] Table 1 Comparison of evaluation results under the same test environment
[0069] The results show that the proposed scheme significantly outperforms the baseline method in terms of detection sensitivity, real-time performance, and cost-effectiveness. This is mainly attributed to the low-complexity feature extraction of VMD decomposition, the lightweight design of the improved ResNet, and the greedy algorithm with dynamic power consumption optimization (reducing energy consumption by 50%).
[0070] Significantly improves the comprehensiveness and accuracy of detection: This solution constructs a multi-dimensional feature vector by simultaneously collecting multi-modal data such as acoustic vibration, infrared thermal imaging and environmental parameters. It can comprehensively perceive the icing state from multiple physical levels such as mass change, surface morphology and temperature distribution, overcome the limitation of the limited perception dimension of a single sensor, greatly reduce the false negative rate and improve the accuracy of state determination.
[0071] Effectively suppressing environmental interference and significantly reducing false alarm rate: This solution innovatively introduces environmental parameters as compensation factors into the feature fusion and decision-making process. Through adaptive weight adjustment algorithm and dynamic threshold determination mechanism, it can compensate for the interference of wind noise and environmental temperature and humidity fluctuations on soundprint and infrared signals in real time, giving the system strong environmental robustness and significantly reducing false alarms caused by severe weather.
[0072] Achieving millisecond-level real-time response to meet online early warning requirements: By adopting the computationally more efficient VMD algorithm and the improved lightweight ResNet-18 model, along with optimization techniques such as model quantization, this solution deploys the core algorithm on the edge computing terminal, realizing local real-time data processing and judgment, completely avoiding the latency caused by cloud transmission, with an end-to-end response time of less than 1 second, providing timely and effective early warning information for operation and maintenance personnel.
[0073] This solution reduces system deployment and maintenance costs, facilitating large-scale applications: Through dynamic power optimization algorithms, it intelligently adjusts sensor sampling rates and communication sleep strategies, significantly reducing overall system power consumption and extending device battery life. Combined with a highly integrated and lightweight edge computing solution, it reduces reliance on communication networks and cloud resources, effectively controlling the hardware, deployment, and maintenance costs of the entire system, making it economically feasible for large-scale application in wind farms.
[0074] Example 2: The icing detection system for wind turbine blades includes: The multimodal sensor module is configured to acquire acoustic signature signals, infrared thermal imaging signals, and environmental parameters. The signal processing module is configured to: preprocess the acquired acoustic signature signal, infrared thermal imaging signal and environmental parameters, determine the environmental interference compensation factor based on the environmental parameters, and perform adaptive weighted fusion processing on the acoustic signature signal and infrared thermal imaging signal to obtain the fused multimodal feature vector. The signal processing module is also configured to: process the acoustic signature signal in the multimodal feature vector using the variational mode decomposition algorithm to obtain several intrinsic mode functions, extract the acoustic signature features, including high frequency energy ratio and low frequency kurtosis; and calculate the temperature gradient index characterizing the degree of abnormality in the temperature distribution on the blade surface based on the infrared thermal imaging signal in the multimodal feature vector. The signal processing module is also configured to: construct a feature vector from the acoustic signature features, temperature gradient index and environmental parameters, and output the icing probability value using the prediction model; The signal processing module is also configured to dynamically adjust the icing judgment threshold according to the current environmental parameters, compare the icing probability value with the dynamically adjusted threshold, and if the probability value is greater than the threshold, determine that the blade is icing and generate an early warning message.
[0075] The hardware component includes a multimodal sensor module and an embedded processing system. The embedded processing system processes the signals acquired by the multimodal sensor module through a signal processing module and generates early warning information. The multimodal sensor module includes: Acoustic unit: MEMS microphone array (signal-to-noise ratio >70dB); Infrared unit: Non-contact infrared sensor (resolution) ); Environmental parameter unit: temperature and humidity sensor (±0.5℃ / ±3%RH accuracy), wind speed sensor (±2% accuracy).
[0076] Embedded processing systems include: Hardware architecture: STM32H743 + FPGA co-processing; Storage unit: NOR Flash (4MB) + SDRAM (64MB); Communication interface: PoE power supply and LoRa wireless communication (transmission distance >10km).
[0077] The anti-interference installation scheme for the above hardware is as follows: Installation location: 3 o'clock position at the root of the blade (the area of greatest stress); Fixing method: Double fixing with magnetic base and industrial adhesive (tensile strength > 500N); Protection design: IP67 protection rating, suitable for temperatures ranging from -40°C to 70°C.
[0078] Example 3: A computer program product includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the aforementioned method for detecting icing on wind turbine blades.
[0079] Example 4: An electronic device includes at least one processor and a memory connected to the processor, the memory being used to store computer programs; the processor is used to execute the computer programs, enabling the electronic device to implement the aforementioned method for detecting icing on wind turbine blades.
[0080] Example 5: A computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the aforementioned method for detecting icing on wind turbine blades, including the following steps.
[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for detecting icing on wind turbine blades, characterized in that, Includes the following steps: Acquire and preprocess the voiceprint signal, infrared thermal imaging signal and environmental parameters, determine the environmental interference compensation factor based on the environmental parameters, and perform adaptive weighted fusion processing on the voiceprint signal and infrared thermal imaging signal to obtain the fused multimodal feature vector. The acoustic signature signal in the multimodal feature vector is processed by the variational mode decomposition algorithm to obtain several intrinsic mode functions. The acoustic signature features are extracted from them, including high-frequency energy ratio and low-frequency kurtosis. Based on the infrared thermal imaging signal in the multimodal feature vector, the temperature gradient index, which characterizes the degree of abnormality in the temperature distribution on the blade surface, is calculated. The voiceprint features, temperature gradient index and environmental parameters are used to construct a feature vector, and the prediction model is used to output the icing probability value. The icing threshold is dynamically adjusted based on the current environmental parameters, and the icing probability value is compared with the dynamically adjusted threshold. If the probability value is greater than the threshold, the blades are determined to be icy, and an early warning message is generated.
2. The method for detecting icing on wind turbine blades as described in claim 1, characterized in that, Adaptive weighted fusion processing is performed on the voiceprint signal and the infrared thermal imaging signal, specifically as follows: Based on environmental parameters, the interference intensity coefficient is estimated using a linear regression model. The weighting coefficients of acoustic signature features and infrared features during fusion are dynamically adjusted based on the interference intensity coefficient. Based on the formula: The environmental vector and the voiceprint / infrared signal are weighted and fused into the main feature vector, where, The softmax function is used to normalize the weight coefficients. , This is the original sensor feature vector. This represents the environmental feature vector.
3. The method for detecting icing on wind turbine blades as described in claim 1, characterized in that, Extracting voiceprint features, specifically: The variational mode decomposition algorithm is used to process the acoustic signature signal in the multimodal feature vector to obtain... One intrinsic mode function; from Among the intrinsic mode functions, the intrinsic mode function representing the high-frequency component is selected, its energy in the set frequency band is calculated, and the high-frequency energy ratio is obtained by comparing it with the total energy of the entire frequency band. from Among the intrinsic mode functions, the intrinsic mode function representing the main vibration frequency band is selected, and its kurtosis value in the time domain is calculated as the low-frequency kurtosis.
4. The method for detecting icing on wind turbine blades as described in claim 1, characterized in that, The temperature gradient index is calculated as follows: the temperature distribution matrix is obtained from the infrared thermal imaging signal in the multimodal feature vector, the temperature difference matrix is determined by combining it with the reference temperature, and the maximum and minimum values of the temperature difference matrix are obtained. The temperature gradient index, which characterizes the degree of abnormality in the temperature distribution on the blade surface, is obtained by the ratio of the difference between the maximum and minimum values to the reference temperature.
5. The method for detecting icing on wind turbine blades as described in claim 1, characterized in that, The prediction model is an improved ResNet-18 model with an integrated attention mechanism. It receives feature vectors containing high-frequency energy ratio, low-frequency kurtosis, temperature gradient exponent, and environmental parameters through the input layer. The model parameters are compressed to 8-bit precision for deployment on edge computing devices.
6. The method for detecting icing on wind turbine blades as described in claim 1, characterized in that, The icing determination threshold is dynamically adjusted based on the current environmental parameters, as shown in the following formula: ； in, Environmental sensitivity coefficient, This indicates the adjusted threshold. Indicates the baseline threshold. This represents the norm of the environment vector.
7. The method for detecting icing on wind turbine blades as described in claim 1, characterized in that, During the acquisition of voiceprint signals, infrared thermal imaging signals, and environmental parameters, dynamic power consumption optimization is performed using a greedy optimization algorithm, specifically as follows: Establish a system energy consumption model with sensor sampling rate, communication interval, and sleep duration as variables; With the detection performance requirements as constraints, the optimal parameter combination that minimizes the total energy consumption of the system is found through greedy optimization algorithm iteratively, and the working state of the sensor is dynamically configured accordingly.
8. A system for detecting icing on wind turbine blades, characterized in that, include: The multimodal sensor module is configured to acquire acoustic signature signals, infrared thermal imaging signals, and environmental parameters. The signal processing module is configured to: preprocess the acquired acoustic signature signal, infrared thermal imaging signal and environmental parameters, determine the environmental interference compensation factor based on the environmental parameters, and perform adaptive weighted fusion processing on the acoustic signature signal and infrared thermal imaging signal to obtain the fused multimodal feature vector. The signal processing module is also configured to: process the acoustic signature signal in the multimodal feature vector using the variational mode decomposition algorithm to obtain several intrinsic mode functions, extract the acoustic signature features, including high frequency energy ratio and low frequency kurtosis; and calculate the temperature gradient index characterizing the degree of abnormality in the temperature distribution on the blade surface based on the infrared thermal imaging signal in the multimodal feature vector. The signal processing module is also configured to: construct a feature vector from the acoustic signature features, temperature gradient index and environmental parameters, and output the icing probability value using the prediction model; The signal processing module is also configured to dynamically adjust the icing judgment threshold according to the current environmental parameters, compare the icing probability value with the dynamically adjusted threshold, and if the probability value is greater than the threshold, determine that the blade is icing and generate an early warning message.
9. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to perform the steps in the wind turbine blade icing detection method as claimed in any one of claims 1-7.
10. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, the memory being used to store computer programs; the processor is used to execute the computer programs, enabling the electronic equipment to perform the steps in the wind turbine blade icing detection method as claimed in any one of claims 1-7.