An intelligent wind farm fan monitoring system and method based on machine learning and digital twinning

Abstract

The application discloses an intelligent wind farm fan monitoring system and method based on machine learning and digital twinning, and relates to the technical field of fan monitoring.The application is based on dynamic probability digital twinning technology, and a multi-dimensional operation state model of a fan is constructed in real time, and state distribution and uncertainty quantization results are dynamically updated, so that the real-time operation state of the fan is accurately reflected, and the flexibility and accuracy of state modeling are improved; a machine learning module realizes accurate prediction of key operation parameters through a physical constraint neural network and a deep operation network, and evaluates potential fault risks, so that the equipment failure rate is effectively reduced; a monitoring module utilizes real-time visualization technology, can intuitively display the operation state of the fan and calibrate a fault area, and facilitates quick positioning of abnormal problems; and an intelligent optimization control module is based on deep reinforcement learning technology, realizes dynamic adjustment of fan operation parameters, improves the power generation efficiency of the fan, and optimizes operation stability and economy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fan monitoring, and particularly relates to an intelligent wind farm fan monitoring system and method based on machine learning and digital twinning. BACKGROUND

[0002] With the continuous progress of wind power technology, the scale and complexity of wind farms are gradually increasing. In modern large-scale wind farms, there are numerous wind turbines, and the real-time changes of wind turbine operating conditions and environmental conditions pose higher requirements on monitoring systems. Traditional wind turbine monitoring systems mainly rely on fixed rules and parameter settings for state monitoring and fault detection, which has great limitations and is difficult to cope with the complexity of wind turbine operating conditions and the diversity of potential faults.

[0003] The development of digital twinning technology provides new possibilities for intelligent monitoring of wind farms. Digital twinning technology can reflect the operating state and performance of the wind turbine in real time by establishing a digital model of the wind turbine, thereby achieving accurate prediction and diagnosis. In addition, the application of machine learning technology in big data analysis and pattern recognition provides strong support for the intelligent operation and maintenance of wind farms. Through continuous training and optimization of machine learning algorithms, the monitoring system can realize real-time analysis and state prediction of wind turbine operating data, timely detect potential faults, and avoid equipment damage and downtime.

[0004] However, existing intelligent wind farm monitoring systems have not fully combined digital twinning and machine learning technology, and cannot realize the whole life cycle management and fault prediction of wind turbines. SUMMARY

[0005] To solve the above problems, the present application provides an intelligent wind farm fan monitoring system and method based on machine learning and digital twinning, to improve the accuracy and intelligence level of monitoring, realize accurate monitoring and early warning of wind turbine operating state, and improve the overall operation efficiency and safety of wind farms.

[0006] The present application achieves the above-mentioned purposes through the following technical solutions:

[0007] An intelligent wind farm fan monitoring system based on machine learning and digital twinning, the system comprising:

[0008] A multi-source data acquisition module for acquiring wind turbine operating data, environmental data and fault-related data, and performing preprocessing operations and compressed transmission on the acquired data through distributed edge computing nodes;

[0009] A data processing module for constructing and updating a multi-dimensional operating state model of the wind turbine in real time based on dynamic probabilistic digital twinning technology, generating dynamically updated state distribution data and uncertainty quantification results;

[0010] The machine learning module is used to predict key operating parameters of the wind turbine and assess potential failure risks based on an integrated architecture of physically constrained neural networks and deep operational networks. The key operating parameters include mechanical operating parameters, electrical operating parameters, environmental parameters, thermodynamic parameters, and failure-related parameters.

[0011] The monitoring module is used to display the wind turbine's operating status through real-time visualization technology, identify potential fault areas, and provide comprehensive monitoring of the wind turbine's status.

[0012] The intelligent optimization control module is used to dynamically adjust the operating parameters of the wind turbine through real-time analysis and deep reinforcement learning technology;

[0013] The user interaction module provides users with real-time visualization of the wind turbine's operating status, manual control functions, and simulation analysis support through a multi-view interface and virtual reality technology. It also collects user feedback to optimize the system's intelligence.

[0014] As a preferred embodiment of the present invention, the multi-source data acquisition module includes:

[0015] The wind turbine operation data acquisition unit is used to collect sensor data inside the wind turbine in real time, including speed, vibration, temperature, current, voltage and power generation.

[0016] The environmental data acquisition unit is used to collect external environmental sensor data in real time, including wind speed, wind direction, temperature and humidity, atmospheric pressure and light intensity;

[0017] The fault-related data acquisition unit is used to collect historical fault logs and real-time abnormal event tagged data;

[0018] The distributed edge computing node unit is used to perform data filtering, data compression, real-time preprocessing, and anomaly detection on the wind turbine operation data, environmental data, and fault-related data. It combines time synchronization technology to timestamp the collected data to ensure the consistency of multi-source data.

[0019] As a preferred embodiment of the present invention, the data processing module includes:

[0020] The high-order outlier removal unit is used for outlier detection and boundary value removal in multi-source data. The outlier detection specifically uses the DBSCAN density clustering algorithm to detect outliers in high-dimensional space. Combined with temporal characteristics, a variational autoencoder is used to train normal patterns to detect data points with high deviation probability. The boundary value removal specifically uses the box plot method to remove extreme values ​​outside the upper and lower quartiles for numerical data. Combined with the real-time updated wind turbine status distribution, new outliers are dynamically removed.

[0021] The feature extraction unit is used to extract the time-frequency characteristics of the wind turbine vibration signal using short-time Fourier transform, extract the instantaneous time-domain characteristics using wavelet transform, and extract the statistical quantities of mean, variance, and skewness to describe the overall trend of the signal.

[0022] The dimensionality optimization unit is used to reduce the dimensionality of the high-dimensional features extracted by the feature extraction unit, reduce redundancy, and retain core information. This includes kernel PCA dimensionality reduction and temporal embedding. Kernel PCA dimensionality reduction uses nonlinear mapping to embed data into a high-dimensional feature space and extract principal components. Temporal embedding uses an autoregressive model or window aggregation technique to convert time series features into embedded representations and capture temporal correlations.

[0023] The dynamic probabilistic digital twin unit is used to construct a multi-dimensional operating state model of the wind turbine and dynamically update it using real-time data to generate operating state characteristic data and uncertainty quantification results.

[0024] As a preferred embodiment of the present invention, the method for constructing the multidimensional operating state model of the wind turbine includes:

[0025] Initial distribution modeling: Initialize the distribution of the wind turbine's operating state using a Gaussian mixture model:

[0026]

[0027] In the formula, P(x|θ) represents the probability density of data point x under model parameters θ; θ is the set of parameters of the Gaussian mixture model, including the mean μ. i , covariance matrix ∑ i and weight w i k is the number of Gaussian components; w i Let the weight of the i-th Gaussian component satisfy the following condition: This indicates that the i-th Gaussian component has a mean value μ. i The sum of the covariance matrix ∑ i ;

[0028] Dynamic covariance adjustment: Introducing an uncertainty matrix U, the covariance matrix is ​​dynamically adjusted.

[0029] ∑'=∑+λ·U;

[0030] In the formula, ∑ is the current covariance matrix, which is used to describe the multidimensional distribution of the wind turbine's operating status; ∑' is the updated covariance matrix; λ is a dynamic adjustment factor, which is used to control the weight of the uncertainty matrix U on the adjustment of the covariance matrix;

[0031] Importance sampling optimization: Optimize the model's weight updates using importance sampling.

[0032]

[0033] In the formula, w (i) The updated weights for particle i; Let P(D) be the weight of particle i at time t-1; t |θ (i) D represents the likelihood function value of particle i; t θ represents the real-time observation data at the current time t; (i) Let be the model parameters for particle i.

[0034] As a preferred embodiment of the present invention, the machine learning module includes:

[0035] The data preprocessing unit receives the state distribution data and uncertainty quantification results from the data processing module, converts them into a vector representation acceptable to the neural network, and performs normalization processing.

[0036] The operating parameter prediction unit is used to construct a physically constrained neural network, using the blade motion equations as constraints to predict the key operating parameters of the wind turbine. The loss function of the physically constrained neural network is expressed as:

[0037] ξ=ξ data +γξ physics ;

[0038] In the formula, ξ is the total loss function of the physically constrained neural network; data The prediction error loss based on observation data; ξ physics γ represents the residual loss of the partial differential equation based on physical constraints; γ is the tradeoff factor.

[0039] The fault risk assessment unit is used to extract fault-related features from the state distribution through a deep operational network to perform fault risk assessment.

[0040] Deep operational networks map the multidimensional inputs of wind turbine status to failure risk probabilities:

[0041] b = κ(u,a);

[0042] In the formula, b is the predicted output of fault risk; κ represents the operating network, which is used to extract nonlinear mapping relationships from input features; u is the state feature vector; and a is the operating variable, which describes the external environmental influencing factors.

[0043] Based on the current state distribution, predict the probability of a failure occurring:

[0044] P(FailureD t )=∫P(Failureθ)·P(θD t )dθ;

[0045] In the formula, P(FailureD)t ) indicates that in the current data D t Under given model parameters θ, the conditional probability of wind turbine failure is given; P(Failureθ) represents the probability of wind turbine failure given model parameters θ; P(θD) represents the conditional probability of wind turbine failure given model parameters θ. t (Based on current data D) t The posterior distribution of θ represents the probability density of the model parameters θ.

[0046] The Monte Carlo integral is used for numerical solution, and the predicted results of future operating parameters and the failure risk assessment report are output.

[0047] As a preferred embodiment of the present invention, the monitoring module includes:

[0048] The multi-mode data visualization unit includes time-series charts and thermal imaging maps. The time-series charts display the historical trends and real-time changes of key operating parameters, while the thermal imaging maps of key components of the fan are generated based on temperature data to intuitively identify possible abnormal areas.

[0049] The 3D modeling unit is used to generate a high-resolution 3D model of the wind turbine by combining digital twin technology, which dynamically reflects the current status of the wind turbine;

[0050] The alarm and status classification unit is used to classify the wind turbine status into "normal", "abnormal" or "emergency" status based on key operating parameters and fault risk assessment results. In the "emergency" status, a fault alarm is triggered to highlight the risk area and provide operating suggestions.

[0051] The status report generation unit is used to generate status reports periodically, providing a report format that combines charts and text, including operating parameter trends, health status of key components, and fault warning information.

[0052] As a preferred embodiment of the present invention, the method for implementing the intelligent optimization control module includes:

[0053] The data processing module obtains multidimensional operating state models and uncertainty quantification results as the basis for control optimization; the machine learning module obtains key operating parameter predictions and fault risk assessment results to set optimization objectives.

[0054] Construct a reinforcement learning model and define the state, action and reward function: the state refers to the current operating parameters of the wind turbine and environmental data, the action refers to adjusting the operating parameters of the wind turbine, and the reward function should take into account the wind turbine's power generation efficiency, operating stability and failure risk.

[0055] Optimization strategies are generated through deep Q-learning or policy gradient algorithms. The wind turbine operating parameters are adjusted in real time according to the optimization strategies, and real-time data is continuously collected to update the multi-dimensional operating status model, forming a closed-loop optimization mechanism.

[0056] The dynamically optimized control commands are output and directly applied to the operation of the wind turbine.

[0057] It monitors changes in key operating parameters in real time after optimization, evaluates the optimization effect, and provides optimization suggestions for record, supporting adjustments to long-term operating strategies.

[0058] A wind turbine monitoring method for an intelligent wind farm wind turbine monitoring system based on machine learning and digital twins, the method comprising:

[0059] Collect wind turbine operation data, environmental data, and fault-related data. Perform preprocessing and compression operations on the collected data through distributed edge computing nodes. The preprocessing operations include data filtering, data compression, real-time preprocessing, and anomaly detection. Combined with time synchronization technology, timestamps are added to the collected data to ensure the consistency of multi-source data.

[0060] Based on dynamic probabilistic digital twin technology, a multi-dimensional operating state model of the wind turbine is constructed and updated in real time, generating dynamically updated state distribution data and uncertainty quantification results;

[0061] Based on the integrated architecture of physical constraint neural network and deep operation network, the key operating parameters of wind turbines are predicted and potential failure risks are assessed.

[0062] The wind turbine's operating status is displayed through real-time visualization technology, potential fault areas are identified, and comprehensive monitoring of the wind turbine's status is provided.

[0063] By using real-time analysis and deep reinforcement learning techniques, the operating parameters of the wind turbine are dynamically adjusted.

[0064] Through multi-view interfaces and virtual reality technology, the system provides users with real-time visualization of wind turbine operating status, manual control functions, and simulation analysis support, while collecting user feedback to optimize system intelligence.

[0065] As a preferred embodiment of the present invention, the method for constructing and updating a multi-dimensional operating state model of the wind turbine in real time based on dynamic probabilistic digital twin technology, and generating dynamically updated state distribution data and uncertainty quantification results, includes:

[0066] The DBSCAN density clustering algorithm is used to detect outliers in high-dimensional space. Combined with temporal characteristics, a variational autoencoder is used to train normal patterns and detect data points with high deviation probability. At the same time, the box plot method is used to remove extreme values ​​outside the upper and lower quartiles.

[0067] The short-time Fourier transform is used to extract the time-frequency characteristics of the wind turbine vibration signal, the wavelet transform is used to extract the instantaneous time-domain characteristics, and the statistical quantities of mean, variance, and skewness are extracted to describe the overall trend of the signal.

[0068] The extracted high-dimensional features are subjected to dimensionality reduction processing, including kernel PCA dimensionality reduction and temporal embedding, to convert time series features into embedded representations to capture temporal correlations;

[0069] A multi-dimensional operating status model of the wind turbine is constructed, and the model status is updated through real-time data to generate operating status characteristic data and uncertainty quantification results.

[0070] As a preferred embodiment of the present invention, the method for predicting key operating parameters of a wind turbine and assessing potential failure risks based on the integrated architecture of a physically constrained neural network and a deep operational network includes:

[0071] Operating parameter prediction: A physical constraint neural network is constructed, using the blade motion equation as a constraint condition to predict the key operating parameters of the wind turbine. The loss function of the physical constraint neural network includes prediction error loss and partial differential equation residual loss based on physical constraints.

[0072] Fault risk assessment extracts fault-related features from the state distribution through a deep operational network and predicts the conditional probability of wind turbine failure. Monte Carlo integration is used to numerically solve for future operating parameters and outputs a fault risk assessment report.

[0073] The method of displaying the wind turbine's operating status through real-time visualization technology, identifying potential fault areas, and providing comprehensive monitoring of the wind turbine's status includes:

[0074] The historical trends and real-time changes of key operating parameters are displayed using time-series charts, and thermal imaging maps of key components of the fan are generated based on temperature data to visually identify potential abnormal areas.

[0075] By combining digital twin technology, a high-resolution 3D model of the wind turbine is generated, which dynamically reflects the current status of the wind turbine;

[0076] Based on key operating parameters and fault risk assessment results, the wind turbine status is classified into "normal", "abnormal" or "emergency" states. In the "emergency" state, a fault alarm is triggered and the risk area is highlighted, while operating suggestions are generated.

[0077] Regularly generate operational status reports, providing a report format that combines charts and text, including operational parameter trends, health status of key components, and fault warning information.

[0078] The beneficial effects of this invention are as follows: The multi-source data acquisition module efficiently collects wind turbine operation data, environmental data, and fault-related data. Edge computing nodes are used for data preprocessing and time synchronization, ensuring the consistency and accuracy of the multi-source data and providing reliable data support for subsequent analysis. The data processing module, based on dynamic probabilistic digital twin technology, constructs a multi-dimensional operating state model of the wind turbine in real time and dynamically updates the state distribution and uncertainty quantification results, thereby accurately reflecting the real-time operating state of the wind turbine and improving the flexibility and accuracy of state modeling. The machine learning module, through physically constrained neural networks and deep operational networks, achieves accurate prediction of key operating parameters and evaluates... The system assesses potential failure risks, providing data-driven support for early warning and optimized wind turbine operation, effectively reducing equipment failure rates. The monitoring module utilizes real-time visualization technology to intuitively display the wind turbine's operating status and pinpoint fault areas, facilitating rapid location of anomalies and improving fault diagnosis and maintenance efficiency. The intelligent optimization control module, based on deep reinforcement learning technology, enables dynamic adjustment of wind turbine operating parameters, improving power generation efficiency while optimizing operational stability and economy. The user interaction module provides real-time visualization, multi-view interfaces, and simulation analysis functions, offering a convenient interactive platform for user operation and feedback optimization, significantly enhancing the system's intelligence level and user experience. Attached Figure Description

[0079] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0080] in:

[0081] Figure 1 This is a system module diagram of the present invention;

[0082] Figure 2 This is a list view of fans in an embodiment of the present invention;

[0083] Figure 3 This is a columnar view of the fan in an embodiment of the present invention;

[0084] Figure 4 This is a cross-sectional view of a single fan in an embodiment of the present invention;

[0085] Figure 5 This is a schematic diagram of the fan control command interface in an embodiment of the present invention;

[0086] Figure 6 This is a flowchart of the method of the present invention. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0088] like Figure 1 As shown, this is one embodiment of the present invention, which provides an intelligent wind farm turbine monitoring system based on machine learning and digital twins, comprising:

[0089] (1) Multi-source data acquisition module

[0090] It is used to collect wind turbine operation data, environmental data and fault-related data, and to perform preprocessing and compression transmission of the collected data through distributed edge computing nodes;

[0091] In one embodiment, the multi-source data acquisition module includes:

[0092] The wind turbine operation data acquisition unit is used to collect sensor data inside the wind turbine in real time, including parameters such as speed, vibration, temperature, current, voltage, and power generation; high-precision sensors are arranged in the core components of the wind turbine (such as gearbox, main shaft, generator) and blades;

[0093] The environmental data acquisition unit deploys weather stations or integrates environmental sensors in the wind farm environment to collect external environmental sensor data in real time, including parameters such as wind speed, wind direction, temperature and humidity, atmospheric pressure, and light intensity.

[0094] The fault-related data acquisition unit is used to collect historical fault logs and real-time abnormal event tagged data;

[0095] Industrial IoT protocols (such as MQTT and CoAP) are used to achieve efficient data transmission between sensors and edge computing nodes;

[0096] Distributed edge computing node units are used to filter, compress, preprocess in real time, and detect anomalies in wind turbine operation data, environmental data, and fault-related data. Combined with time synchronization technology (such as NTP protocol), the collected data is timestamped to ensure the consistency of multi-source data. Edge device caching and batch transmission strategies are used to reduce the latency of real-time transmission and network pressure.

[0097] Data filtering: Removes obviously erroneous data, such as missing values ​​or outliers that are outside the sensor's range;

[0098] Data compression: Compress the collected data using algorithms (such as LZW and Huffman coding) to reduce transmission bandwidth consumption;

[0099] Real-time preprocessing: preliminary noise reduction and standardization to reduce the data processing pressure on subsequent modules;

[0100] Anomaly detection: Lightweight machine learning models (such as random forests) are used to initially identify and label potentially anomalous data points.

[0101] (2) Data processing module

[0102] This is used to construct and update a multi-dimensional operating state model of wind turbines in real time based on dynamic probabilistic digital twin technology, and to generate dynamically updated state distribution data and uncertainty quantification results.

[0103] The data processing module includes:

[0104] The high-order outlier removal unit is used for outlier detection and boundary value removal in multi-source data. Specifically, outlier detection uses the DBSCAN density clustering algorithm to detect outliers in high-dimensional space. Combined with temporal characteristics, a variational autoencoder (VAE) is used to train normal patterns to detect data points with high deviation probability. Boundary value removal is specifically for numerical data. The box plot method is used to remove extreme values ​​outside the upper and lower quartiles. Combined with the real-time updated wind turbine status distribution, new outliers are dynamically removed to improve the model's adaptability to real-time changes.

[0105] The feature extraction unit is used to extract the time-frequency characteristics of the wind turbine vibration signal using short-time Fourier transform (STFT), extract the instantaneous time-domain characteristics using wavelet transform, characterize the local variation features of the signal, and extract statistical quantities such as mean, variance, and skewness to describe the overall variation trend of the signal.

[0106] The dimensionality optimization unit is used to reduce the dimensionality of high-dimensional features extracted by the feature extraction unit, reducing redundancy and retaining core information. This includes kernel PCA dimensionality reduction and temporal embedding. Kernel PCA dimensionality reduction embeds data into a high-dimensional feature space through nonlinear mapping to extract principal components. It is suitable for extracting dynamically distributed nonlinear features to enhance the expressive power of the model. Temporal embedding uses autoregressive models or window aggregation techniques to convert time series features into embedded representations and capture temporal correlations.

[0107] The dynamic probabilistic digital twin unit is used to construct a multi-dimensional operating state model of the wind turbine and dynamically update it using real-time data to generate operating state characteristic data and uncertainty quantification results, thereby enhancing the model's ability to describe the operating state.

[0108] In one specific embodiment, the method for constructing a multidimensional operating state model of the wind turbine includes:

[0109] Initial distribution modeling: The distribution of the wind turbine's operating state is initialized using a Gaussian mixture model (GMM).

[0110]

[0111] In the formula, P(xθ) represents the probability density of data point x under model parameters θ; θ is the set of parameters of the Gaussian mixture model, including the mean μ. i , covariance matrix ∑ i and weight w i k is the number of Gaussian components; w i Let the weight of the i-th Gaussian component satisfy the following condition: This indicates that the i-th Gaussian component has a mean value μ. i The sum of the covariance matrix ∑ i ;

[0112] Dynamic covariance adjustment: Introducing an uncertainty matrix U, the covariance matrix is ​​dynamically adjusted.

[0113] ∑'=∑+λ·U;

[0114] In the formula, ∑ is the current covariance matrix, which is used to describe the multidimensional distribution of the wind turbine's operating status; ∑' is the updated covariance matrix; λ is a dynamic adjustment factor, which is used to control the weight of the uncertainty matrix U on the adjustment of the covariance matrix; the uncertainty matrix U reflects the new sources of uncertainty in the data distribution (such as real-time data changes);

[0115] Importance sampling optimization: Optimize the model's weight updates using importance sampling.

[0116]

[0117] In the formula, w (i) The updated weight for particle i reflects its contribution to the current posterior distribution; Let P(D) be the weight of particle i at time t-1; t |θ (i) Let θ be the likelihood function value of particle i, representing the parameter θ. (i) In the current data D t Adaptability under D; t θ represents the real-time observation data at the current time t; (i) Let i be the model parameters for particle i.

[0118] (3) Machine Learning Module

[0119] This is used to predict key operating parameters of wind turbines and assess potential failure risks based on an integrated architecture of Physically Constrained Neural Network (PINN) and DeepONet.

[0120] Key operating parameters include:

[0121] 1) Mechanical operating parameters

[0122] Rotational speed: The rotational speed of the fan blades (unit: rpm), used to reflect the mechanical operating status of the fan;

[0123] Vibration amplitude: The vibration intensity of key components of the wind turbine (such as bearings and gearboxes), used to detect whether there are mechanical abnormalities;

[0124] Yaw angle: The relative angle between the wind turbine and the wind direction, used to assess the wind turbine's response to changes in wind direction;

[0125] 2) Electrical operating parameters

[0126] Power generation: The actual power output of the wind turbine (unit: kW), which directly reflects the power generation efficiency and operating status;

[0127] Current and voltage: Current and voltage of the wind turbine power generation system; abnormal values ​​may indicate electrical system faults.

[0128] Power factor: The ratio of the effective component to the total component of electrical energy output, used to assess power quality;

[0129] 3) Environmental related parameters

[0130] Wind speed and direction: These are the main environmental factors affecting wind turbine operation, directly impacting power generation and blade load.

[0131] Ambient temperature has a significant impact on the performance of fan components (such as bearings and lubricating oil).

[0132] 4) Thermodynamic parameters

[0133] Bearing temperature: The temperature of the bearing, a critical component, used to monitor whether its operating status is normal;

[0134] Gearbox temperature: The temperature of the fan gearbox is used to determine the lubrication status and wear level;

[0135] 5) Fault-related parameters

[0136] Component stress and fatigue life: stress state and life prediction of critical components to assess potential failure risks;

[0137] Transient abnormal parameters, such as load fluctuations and speed jumps, are used to capture abnormal behaviors under non-stationary conditions.

[0138] In one specific embodiment, the machine learning module includes:

[0139] The data preprocessing unit receives the state distribution data and uncertainty quantification results from the data processing module, converts them into a vector representation acceptable to the neural network, and performs normalization processing.

[0140] The operating parameter prediction unit is used to construct a physically constrained neural network, using the blade motion equations as constraints to predict the key operating parameters of the wind turbine. The loss function of the physically constrained neural network is expressed as:

[0141] ξ=ξ data +γξ physics ;

[0142] In the formula, ξ is the total loss function of the physically constrained neural network; data The prediction error loss based on observation data (data-driven part); ξ physics γ represents the residual loss of physical constraint-based partial differential equations (PDEs) (physical-driven part); γ is a tradeoff factor used to adjust the relative importance of data-driven and physical-driven losses.

[0143] The fault risk assessment unit is used to extract fault-related features from the state distribution through a deep operational network to perform fault risk assessment.

[0144] Deep operational networks map the multidimensional inputs of wind turbine status to failure risk probabilities:

[0145] b = κ(u,a);

[0146] In the formula, b is the predicted output of fault risk, such as fault probability or risk score; κ represents the operating network, which is used to extract nonlinear mapping relationships from input features; u is the state feature vector, which describes the key features of the wind turbine's current operating state; a is the operating variable, which describes the external environmental influencing factors.

[0147] Based on the current state distribution, predict the probability of a failure occurring:

[0148] P(FailureD t )=∫P(Failureθ)·P(θD t )dθ;

[0149] In the formula, P(Failure|D t ) indicates that in the current data D t Under given model parameters θ, the conditional probability of wind turbine failure is given; P(Failureθ) represents the probability of wind turbine failure given model parameters θ; P(θD) represents the conditional probability of wind turbine failure given model parameters θ. t (Based on current data D) t The posterior distribution of θ represents the probability density of the model parameters θ.

[0150] The Monte Carlo integral is used for numerical solution, and the predicted results of future operating parameters and the failure risk assessment report are output.

[0151] (4) Monitoring module

[0152] It is used to display the operating status of wind turbines through real-time visualization technology, identify potential fault areas, and provide comprehensive monitoring of wind turbine status to improve the accuracy of wind turbine operating status monitoring;

[0153] Specifically, the monitoring module includes:

[0154] The multi-mode data visualization unit includes time-series charts and thermal imaging maps. The time-series charts display the historical trends and real-time changes of key operating parameters (such as speed and power generation), while the thermal imaging maps of key wind turbine components are generated based on temperature data to visually identify possible abnormal areas.

[0155] The 3D modeling unit is used to generate a high-resolution 3D model of the wind turbine by combining digital twin technology, which dynamically reflects the current status of the wind turbine;

[0156] The alarm and status classification unit is used to classify the wind turbine status into "normal", "abnormal" or "emergency" status based on key operating parameters and fault risk assessment results. In the "emergency" status, a fault alarm is triggered to highlight the risk area and provide operating suggestions (such as reducing the speed and checking components).

[0157] The status report generation unit is used to generate status reports periodically, which include operating parameter trends, key component health status and fault warning information. It provides a report format that combines charts and text for easy user understanding.

[0158] (5) Intelligent optimization control module

[0159] It is used to dynamically adjust wind turbine operating parameters through real-time analysis and deep reinforcement learning technology to improve power generation efficiency and operational stability;

[0160] Furthermore, the implementation methods of the intelligent optimization control module include:

[0161] The data processing module obtains multidimensional operating state models and uncertainty quantification results as the basis for control optimization; the machine learning module obtains key operating parameter predictions and fault risk assessment results to set optimization objectives (such as minimizing vibration amplitude and maximizing power generation efficiency).

[0162] Construct a reinforcement learning model and define the state, action, and reward function: the state refers to the current operating parameters of the wind turbine and environmental data, the action refers to adjusting the operating parameters of the wind turbine (such as blade angle and speed), and the reward function should take into account the wind turbine's power generation efficiency, operating stability, and failure risk.

[0163] Optimization strategies are generated through Deep Q-Learning or Policy Gradient algorithms. Based on these strategies, wind turbine operating parameters (such as blade angle and power generation) are adjusted in real time. Real-time data is continuously collected to update the multi-dimensional operating state model, forming a closed-loop optimization mechanism.

[0164] The dynamically optimized control commands are output and directly applied to the operation of the wind turbine.

[0165] Real-time monitoring of changes in key operating parameters after optimization, evaluation of optimization effects (such as reduced energy consumption and reduced failure risk), and provision of optimization suggestion records to support adjustments to long-term operating strategies.

[0166] (6) User interaction module

[0167] It is used to provide users with real-time visualization of wind turbine operating status, manual control functions and simulation analysis support through multi-view interface and virtual reality (VR) technology, while collecting user feedback to optimize system intelligence.

[0168] Figures 2-5The diagrams shown are some interface illustrations in embodiments of the present invention. The wind turbine list view allows users to customize the monitoring points and define the display column order of the points. The points can be sorted by maximum or minimum value, allowing users to focus on monitoring specific points. The wind turbine bar view can be used to observe which wind turbines fail to reach their full power output when the wind speed reaches full capacity, which can then be manually recorded by the customer. In the single wind turbine interface, users can view more monitoring information, statistics, and related charts for the current wind turbine, and use corresponding shortcut functions for detailed analysis and refined management. Clicking the wind turbine name in the search box, double-clicking the matrix block, and double-clicking the list will navigate to the single wind turbine view, allowing users to view more monitoring information, statistics, and related charts for the current wind turbine, and use the corresponding functions for the current wind turbine. In the single wind turbine interface, users can choose between "Classic Interface," "Cross-section Interface," and "Benchmark Interface." The default view is "Classic Interface," and clicking the top icon switches to the corresponding interface. In the Classic Interface, users can view more monitoring information, statistics, and related charts for the current wind turbine, and use the corresponding functions for detailed analysis of the current wind turbine. The classic interface is divided into an equipment overview area, a real-time trend curve area, a real-time data area, and an alarm status area. The cross-sectional interface is mainly divided into a basic information area, a status area, a view switching area, a component area, a cross-sectional view area, a measurement point monitoring area, a control area, an alarm / status area, and a basic indicator area. In the single-fan interface, the current fan can be controlled. Interactions are primarily displayed as control commands directly on the screen for easy user access. Clicking the reset, start, stop, lock, or more buttons first brings up a control permission authentication pop-up. After authentication, corresponding operations can be performed. For example, enabling "Equipment Control Monitoring Verification" will bring up a "Monitor Verification" panel. Control commands can only be issued after successful verification. The system supports fan control log recording, ensuring that each issued control command is traceable to the time, person issuing the command, and reason for control. It also supports command execution effect tracking, refreshing and displaying key measurement point information of the fan in real time during the execution of control commands, such as voltage, current, generator speed, and fan status.

[0169] like Figure 6 As shown, another embodiment of the present invention provides a smart wind farm turbine monitoring method based on machine learning and digital twins. Based on the smart wind farm turbine monitoring system based on machine learning and digital twins as described above, the method includes the following steps:

[0170] S1: Collects wind turbine operation data, environmental data, and fault-related data. It performs preprocessing and compression operations on the collected data through distributed edge computing nodes. The preprocessing operations include data filtering, data compression, real-time preprocessing, and anomaly detection. It also adds timestamps to the collected data using time synchronization technology to ensure the consistency of multi-source data.

[0171] S2: Based on dynamic probabilistic digital twin technology, construct and update a multi-dimensional operating state model of the wind turbine in real time, and generate dynamically updated state distribution data and uncertainty quantification results;

[0172] S2.1: Use the DBSCAN density clustering algorithm to detect outliers in high-dimensional space. Combined with temporal characteristics, use a variational autoencoder to train normal patterns and detect data points with high deviation probability. At the same time, use the box plot method to remove extreme values ​​outside the upper and lower quartiles.

[0173] S2.2: Use short-time Fourier transform to extract the time-frequency characteristics of the fan vibration signal, use wavelet transform to extract the instantaneous time-domain characteristics, and extract statistics (mean, variance, skewness) to describe the overall trend of the signal.

[0174] S2.3: Perform dimensionality reduction processing on the extracted high-dimensional features, including kernel PCA dimensionality reduction and temporal embedding, to convert time series features into embedded representations to capture temporal correlations;

[0175] S2.4: Construct a multi-dimensional operating status model of the wind turbine, update the model status through real-time data, and generate operating status characteristic data and uncertainty quantification results.

[0176] S3: Based on the integrated architecture of physical constraint neural network and deep operation network, it predicts the key operating parameters of wind turbines and assesses potential failure risks.

[0177] S3.1: Predicting operating parameters. Construct a physical constraint neural network and use the blade motion equation as a constraint to predict the key operating parameters of the wind turbine. The loss function of the physical constraint neural network includes prediction error loss and physical constraint-based partial differential equation residual loss.

[0178] S3.2: Fault Risk Assessment. This involves extracting fault-related features from the state distribution using a deep operational network, predicting the conditional probability of wind turbine failures, and numerically solving future operating parameters using Monte Carlo integration to output a fault risk assessment report.

[0179] S4: Displays the wind turbine's operating status through real-time visualization technology, identifies potential fault areas, and provides comprehensive monitoring of the wind turbine's status;

[0180] S4.1: Display the historical trends and real-time changes of key operating parameters using time-series charts, and generate thermal imaging maps of key components of the fan based on temperature data to visually identify potential abnormal areas;

[0181] S4.2: Combine digital twin technology to generate a high-resolution 3D model of the wind turbine, dynamically reflecting the current status of the wind turbine;

[0182] S4.3: Based on key operating parameters and fault risk assessment results, classify the wind turbine status into "normal", "abnormal" or "emergency" states. In the "emergency" state, trigger a fault alarm and highlight the risk area, while generating operating suggestions.

[0183] S4.4: Regularly generate operation status reports, providing a report format that combines charts and text, including operation parameter trends, health status of key components, and fault warning information.

[0184] S5: Dynamically adjusts wind turbine operating parameters through real-time analysis and deep reinforcement learning technology;

[0185] S6: Through a multi-view interface and virtual reality technology, it provides users with real-time visualization of the wind turbine's operating status, manual control functions, and simulation analysis support, while collecting user feedback to optimize the system's intelligence.

[0186] In summary, this invention collects wind turbine operation data, environmental data, and fault-related data through a multi-source data acquisition module, and utilizes distributed edge computing nodes for real-time data preprocessing to ensure data consistency and integrity. This module achieves efficient acquisition and processing of multi-type and multi-dimensional data, laying the foundation for subsequent wind turbine operation status modeling and analysis.

[0187] The data processing module employs dynamic probabilistic digital twin technology to construct and update a multi-dimensional operating status model of the wind turbine in real time, generating dynamically updated status distribution data and uncertainty quantification results. This function can accurately reflect the operating status of the wind turbine, promptly capture dynamic changes in operating parameters, and provide a reliable basis for subsequent status prediction and fault assessment.

[0188] The machine learning module, through an integrated architecture of physically constrained neural networks and deep operational networks, enables accurate prediction of key wind turbine operating parameters (such as mechanical operating parameters, electrical operating parameters, environmental parameters, thermodynamic parameters, and fault-related parameters), while also assessing potential failure risks. This module can promptly identify abnormal conditions during wind turbine operation and proactively detect potential failure risks, thereby effectively reducing equipment failure rates and downtime, and improving the overall operating efficiency of the wind farm.

[0189] The monitoring module, through real-time visualization technology, can intuitively display the operating status of the wind turbine, identify potential fault areas, and provide comprehensive monitoring functions for the wind turbine's operating status. By combining multi-mode visualization (such as thermal imaging maps, real-time trend charts, etc.), users can quickly locate fault areas and take corresponding maintenance measures, improving the efficiency of operation and maintenance.

[0190] The intelligent optimization control module combines deep reinforcement learning technology with real-time analysis and multi-dimensional operational status data to dynamically adjust the operating parameters of the wind turbines. By optimizing the wind turbine operating strategy, this module effectively improves the power generation efficiency of the wind turbines, reduces operating costs, and enhances the operational stability of the wind farm.

[0191] The user interaction module provides users with a real-time visual display of the wind turbine's operating status through a multi-view interface and virtual reality technology, supporting manual control and simulation analysis. Users can intuitively understand the wind turbine's operating status through the interactive interface, quickly execute control commands, and the feedback user data can further optimize the system's intelligence and user experience.

[0192] This system fully integrates digital twin and machine learning technologies, overcoming the limitations of traditional monitoring systems and achieving complete closed-loop management from wind turbine data acquisition, status modeling, fault prediction to dynamic optimization control. Through precise monitoring and prediction capabilities, this invention not only improves the reliability and safety of wind turbines but also reduces manual maintenance costs, providing strong support for the intelligent operation and maintenance of wind farms.

[0193] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart wind farm turbine monitoring system based on machine learning and digital twins, characterized in that, The system includes: a multi-source data acquisition module, which is used to collect wind turbine operation data, environmental data and fault-related data, and performs data filtering, data compression, real-time preprocessing and anomaly detection on the collected data through distributed edge computing nodes, and timestamps the collected data in combination with time synchronization technology to ensure consistency; The data processing module is used to construct and update a multi-dimensional operating state model of the wind turbine in real time based on dynamic probabilistic digital twin technology, generating dynamically updated state distribution data and uncertainty quantification results; the method for constructing the multi-dimensional operating state model of the wind turbine includes: Initial distribution modeling: Initialize the distribution of the wind turbine's operating state using a Gaussian mixture model: In the formula, P(x|θ) represents the probability density of data point x under model parameters θ; θ is the set of parameters of the Gaussian mixture model, including the mean μ. i , covariance matrix ∑ i and weight w i k is the number of Gaussian components; w i Let the weight of the i-th Gaussian component satisfy the following condition: This indicates that the i-th Gaussian component has a mean value μ. i The sum of the covariance matrix ∑ i ; Dynamic covariance adjustment: Introducing an uncertainty matrix U, the covariance matrix is ​​dynamically adjusted. ∑'=∑+λ·U; In the formula, ∑ is the current covariance matrix, which is used to describe the multidimensional distribution of the wind turbine's operating status; ∑' is the updated covariance matrix; λ is a dynamic adjustment factor, which is used to control the weight of the uncertainty matrix U on the adjustment of the covariance matrix; Importance sampling optimization: Optimize the model's weight updates using importance sampling. In the formula, w (i) The updated weights for particle i; Let P(D) be the weight of particle i at time t-1; t |θ (i) D represents the likelihood function value of particle i; t θ represents the real-time observation data at the current time t; (i) For particle i, these are the model parameters; The machine learning module is used to predict key operating parameters of the wind turbine and assess potential failure risks based on an integrated architecture of physically constrained neural networks and deep operational networks. The key operating parameters include mechanical operating parameters, electrical operating parameters, environmental parameters, thermodynamic parameters, and failure-related parameters. The monitoring module is used to display the wind turbine's operating status through real-time visualization technology, identify potential fault areas, and provide comprehensive monitoring of the wind turbine's status. The intelligent optimization control module is used to dynamically adjust the operating parameters of the wind turbine through real-time analysis and deep reinforcement learning technology; The user interaction module is used to provide users with real-time visualization of the wind turbine's operating status, manual control functions, and simulation analysis support through a multi-view interface and virtual reality technology, while collecting user feedback to optimize the system's intelligence. The machine learning module includes: The data preprocessing unit receives the state distribution data and uncertainty quantification results from the data processing module, converts them into a vector representation acceptable to the neural network, and performs normalization processing. The operating parameter prediction unit is used to construct a physically constrained neural network, using the blade motion equations as constraints to predict the key operating parameters of the wind turbine. The loss function of the physically constrained neural network is expressed as: ξ=ξ data +gx physics ; In the formula, ξ is the total loss function of the physically constrained neural network; data The prediction error loss based on observation data; ξ physics γ represents the residual loss of the partial differential equation based on physical constraints; γ is the tradeoff factor. The fault risk assessment unit is used to extract fault-related features from the state distribution through a deep operational network to perform fault risk assessment. Deep operational networks map the multidimensional inputs of wind turbine status to failure risk probabilities: b = κ(u,a); In the formula, b is the predicted output of fault risk; κ represents the operating network, which is used to extract nonlinear mapping relationships from input features; u is the state feature vector; and a is the operating variable, which describes the external environmental influencing factors. Based on the current state distribution, predict the probability of a failure occurring: P(Failure|D t )=∫P(Failure|θ)·P(θ|D t )dθ; In the formula, P(Failure|D t ) indicates that in the current data D t Under given model parameters θ, the conditional probability of wind turbine failure is given; P(Failure|θ) represents the probability of wind turbine failure given model parameters θ; P(θ|D t (Based on current data D) t The posterior distribution of θ represents the probability density of the model parameters θ. The Monte Carlo integral is used for numerical solution, and the predicted results of future operating parameters and the failure risk assessment report are output.

2. The intelligent wind farm turbine monitoring system based on machine learning and digital twins as described in claim 1, characterized in that, The multi-source data acquisition module includes: The wind turbine operation data acquisition unit is used to collect sensor data inside the wind turbine in real time, including speed, vibration, temperature, current, voltage and power generation. The environmental data acquisition unit is used to collect external environmental sensor data in real time, including wind speed, wind direction, temperature and humidity, atmospheric pressure and light intensity; The fault-related data acquisition unit is used to collect historical fault logs and real-time abnormal event tagged data.

3. The intelligent wind farm turbine monitoring system based on machine learning and digital twins as described in claim 1, characterized in that, The data processing module includes: The high-order outlier removal unit is used for outlier detection and boundary value removal in multi-source data. The outlier detection specifically uses the DBSCAN density clustering algorithm to detect outliers in high-dimensional space. Combined with temporal characteristics, a variational autoencoder is used to train normal patterns to detect data points with high deviation probability. The boundary value removal specifically uses the box plot method to remove extreme values ​​outside the upper and lower quartiles for numerical data. Combined with the real-time updated wind turbine status distribution, new outliers are dynamically removed. The feature extraction unit is used to extract the time-frequency characteristics of the wind turbine vibration signal using short-time Fourier transform, extract the instantaneous time-domain characteristics using wavelet transform, and extract the statistical quantities of mean, variance, and skewness to describe the overall trend of the signal. The dimensionality optimization unit is used to reduce the dimensionality of the high-dimensional features extracted by the feature extraction unit, reduce redundancy, and retain core information. This includes kernel PCA dimensionality reduction and temporal embedding. Kernel PCA dimensionality reduction uses nonlinear mapping to embed data into a high-dimensional feature space and extract principal components. Temporal embedding uses an autoregressive model or window aggregation technique to convert time series features into embedded representations and capture temporal correlations. The dynamic probabilistic digital twin unit is used to construct a multi-dimensional operating state model of the wind turbine and dynamically update it using real-time data to generate operating state characteristic data and uncertainty quantification results.

4. The intelligent wind farm turbine monitoring system based on machine learning and digital twins as described in claim 1, characterized in that, The monitoring module includes: The multi-mode data visualization unit includes time-series charts and thermal imaging maps. The time-series charts display the historical trends and real-time changes of key operating parameters, while the thermal imaging maps of key components of the fan are generated based on temperature data to intuitively identify possible abnormal areas. The 3D modeling unit is used to generate a high-resolution 3D model of the wind turbine by combining digital twin technology, which dynamically reflects the current status of the wind turbine; The alarm and status classification unit is used to classify the wind turbine status into "normal", "abnormal" or "emergency" status based on key operating parameters and fault risk assessment results. In the "emergency" status, a fault alarm is triggered to highlight the risk area and provide operating suggestions. The status report generation unit is used to generate status reports periodically, providing a report format that combines charts and text, including operating parameter trends, health status of key components and fault warning information.

5. The intelligent wind farm turbine monitoring system based on machine learning and digital twins as described in claim 1, characterized in that, The implementation method of the intelligent optimization control module includes: The data processing module obtains multidimensional operating state models and uncertainty quantification results as the basis for control optimization; the machine learning module obtains key operating parameter predictions and fault risk assessment results to set optimization objectives. Construct a reinforcement learning model and define the state, action and reward function: the state refers to the current operating parameters of the wind turbine and environmental data, the action refers to adjusting the operating parameters of the wind turbine, and the reward function should take into account the wind turbine's power generation efficiency, operating stability and failure risk. Optimization strategies are generated through deep Q-learning or policy gradient algorithms. The wind turbine operating parameters are adjusted in real time according to the optimization strategies, and real-time data is continuously collected to update the multi-dimensional operating status model, forming a closed-loop optimization mechanism. The dynamically optimized control commands are output and directly applied to the operation of the wind turbine. It monitors changes in key operating parameters in real time after optimization, evaluates the optimization effect, and provides optimization suggestions for record, supporting adjustments to long-term operating strategies.

6. A wind turbine monitoring method for an intelligent wind farm turbine monitoring system based on machine learning and digital twins according to any one of claims 1-5, characterized in that, The method includes: Collect wind turbine operation data, environmental data, and fault-related data. Perform preprocessing and compression operations on the collected data through distributed edge computing nodes. The preprocessing operations include data filtering, data compression, real-time preprocessing, and anomaly detection. Combined with time synchronization technology, timestamps are added to the collected data to ensure the consistency of multi-source data. Based on dynamic probabilistic digital twin technology, a multidimensional operating state model of the wind turbine is constructed and updated in real time, generating dynamically updated state distribution data and uncertainty quantification results. The method includes: The DBSCAN density clustering algorithm was used to detect outliers in high-dimensional space. Combined with temporal characteristics, a variational autoencoder was used to train normal patterns and detect data points with high deviation probability. At the same time, the box plot method was used to remove extreme values ​​outside the upper and lower quartiles. The time-frequency characteristics of the wind turbine vibration signal were extracted using short-time Fourier transform, and the instantaneous time-domain characteristics were extracted using wavelet transform. The statistical measures of mean, variance, and skewness were extracted to describe the overall trend of the signal. The extracted high-dimensional features are subjected to dimensionality reduction processing, including kernel PCA dimensionality reduction and temporal embedding, to convert time series features into embedded representations to capture temporal correlations; A multi-dimensional operating status model of the wind turbine is constructed, and the model status is updated through real-time data to generate operating status characteristic data and uncertainty quantification results; Based on an integrated architecture of physically constrained neural networks and deep operational networks, key operating parameters of wind turbines are predicted and potential failure risks are assessed. The methods include: Operating parameter prediction: A physical constraint neural network is constructed, using the blade motion equation as a constraint condition to predict the key operating parameters of the wind turbine. The loss function of the physical constraint neural network includes prediction error loss and partial differential equation residual loss based on physical constraints. Fault risk assessment extracts fault-related features from the state distribution through a deep operational network and predicts the conditional probability of wind turbine failure. Monte Carlo integration is used to numerically solve for future operating parameters and outputs a fault risk assessment report. The wind turbine's operating status is displayed through real-time visualization technology, potential fault areas are identified, and comprehensive monitoring of the wind turbine's status is provided. By using real-time analysis and deep reinforcement learning techniques, the operating parameters of the wind turbine are dynamically adjusted. Through multi-view interfaces and virtual reality technology, the system provides users with real-time visualization of wind turbine operating status, manual control functions, and simulation analysis support, while collecting user feedback to optimize system intelligence.

7. The wind turbine monitoring method of the intelligent wind farm wind turbine monitoring system based on machine learning and digital twins as described in claim 6, characterized in that, The method of displaying the wind turbine's operating status through real-time visualization technology, identifying potential fault areas, and providing comprehensive monitoring of the wind turbine's status includes: The historical trends and real-time changes of key operating parameters are displayed using time-series charts, and thermal imaging maps of key components of the fan are generated based on temperature data to visually identify potential abnormal areas. By combining digital twin technology, a high-resolution 3D model of the wind turbine is generated, which dynamically reflects the current status of the wind turbine; Based on key operating parameters and fault risk assessment results, the wind turbine status is classified into "normal", "abnormal" or "emergency" states. In the "emergency" state, a fault alarm is triggered and the risk area is highlighted, while operating suggestions are generated. Regularly generate operational status reports, providing a report format that combines charts and text, including operational parameter trends, health status of key components, and fault warning information.

Images