Intelligent detection method and system for health state of wind generating set in intelligent wind field

Abstract

The invention discloses an intelligent detection method and system for the health state of a wind generating set in an intelligent wind field. The method comprises the steps that vibration, acoustic emission, temperature gradient and electromagnetic noise multi-mode signals of a main bearing, a gearbox, a generator rotor and a tower drum flange are collected; generating a multi-dimensional state feature vector through time-frequency domain joint feature extraction; inputting a graph attention network model constructed based on a mechanical transmission and thermal coupling relationship, and outputting a local health index; and in combination with working condition parameters such as wind speed, power and yaw angle, a Bayesian fusion inference model is utilized to calculate an overall health score and divide state grades. According to the invention, early weak fault high-sensitivity identification can be realized, and the evaluation precision and the engineering deployability are improved.

Description

Technical Field

[0001] This invention relates to the fields of new energy and intelligent operation and maintenance technology, and in particular to a method and system for intelligent detection of the health status of wind turbine generators in smart wind farms. Background Technology

[0002] With the deepening of energy structure transformation, wind power installed capacity continues to grow rapidly. In complex and ever-changing operating environments, newly built wind turbines are still in the performance break-in phase, facing operational risks such as abnormal component wear and excessive vibration. Meanwhile, older turbines, due to material aging and fatigue accumulation, have a gradually increasing probability of failure. To ensure the safe and efficient operation of wind farms, there is an urgent need for an accurate and efficient health status detection method to monitor the real-time operating status of wind turbine generators.

[0003] Currently, vibration signal analysis and temperature monitoring are commonly used to assess the condition of wind turbine generators. However, existing detection methods still have certain limitations in terms of multi-source heterogeneous data fusion and early weak fault feature extraction. They usually rely on multiple sensors working together to improve diagnostic accuracy. Therefore, it is necessary to develop a highly integrated and adaptable intelligent detection method and system to achieve comprehensive, accurate, and long-term monitoring of the health status of wind turbine generators. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent detection of the health status of wind turbine generator sets in smart wind farms, which solves the problems mentioned in the background art.

[0005] This invention is implemented as follows: a smart wind farm wind turbine generator health status intelligent detection method and system, comprising: collecting multimodal sensing signals generated by multiple key components of the wind turbine generator during operation, wherein the multimodal sensing signals include vibration acceleration signals, acoustic emission signals, temperature gradient signals, and electromagnetic noise signals, and the key components include the main bearing, gearbox, generator rotor, and tower connecting flange; performing time-frequency domain joint feature extraction on the multimodal sensing signals to generate multidimensional state feature vectors corresponding to each key component, wherein the time-frequency domain joint feature extraction includes performing wavelet packet decomposition on the vibration acceleration signals to obtain the energy entropy coefficient, performing short-time Fourier transform on the acoustic emission signals to extract the center frequency offset, calculating the difference in heat conduction rate between adjacent measuring points on the temperature gradient signals, and performing Hilbert-Huang transform on the electromagnetic noise signals to obtain the instantaneous amplitude fluctuation sequence; and inputting the multidimensional state feature vectors into... In the pre-trained graph attention network model, the nodes of the graph attention network model correspond to each key component, and the edge weights are dynamically set according to the mechanical transmission path and thermal coupling relationship between the components. The state information of adjacent nodes is aggregated through a multi-layer graph attention mechanism to output the local health index of each key component. Based on the local health index and the overall operating condition parameters of the wind turbine generator set, including wind speed range, power output level and yaw angle change rate, a Bayesian fusion inference model is constructed to calculate the overall health status score of the wind turbine generator set, and the current operating status is divided into normal, warning or fault according to the preset health level threshold. In the training phase, the graph attention network model uses historical operating datasets for end-to-end optimization. The historical operating datasets contain labeled fault samples and corresponding multimodal sensor signals. The model loss function is composed of the local health index prediction error and the overall health status classification cross-entropy.

[0006] Secondly, this invention provides an intelligent health status detection system for wind turbine generator sets in smart wind farms, including a multimodal sensor array, an edge computing unit, a communication module, a central analysis server, and a database. The multimodal sensor array is fixedly installed on the main bearing housing, gearbox housing, generator stator winding ends, and tower flange connection of the wind turbine generator set. Vibration acceleration sensors are vertically attached to the bearing housing surface via bolts. Acoustic emission sensors are magnetically attached to the outer wall of the gearbox and cover the lubricating oil return channel area using a magnetic base. Temperature sensors are embedded inside the generator winding insulation layer with a spacing not exceeding 15 cm. Electromagnetic noise probes are connected to the generator output terminal block via shielded cables. The edge computing unit is integrated in the control cabinet at the bottom of the tower and connected to the multimodal sensor array via an industrial Ethernet switch for monitoring... The raw sensor signals undergo anti-aliasing filtering, sampling rate unification, and preliminary feature extraction. The generated multi-dimensional state feature vectors are then uploaded via the communication module. The communication module, employing a 5G industrial module, is deployed on an internal support frame within the nacelle and connected to the wind farm's SCADA system via a fiber optic ring network. It transmits data output from the edge computing unit to the central analysis server in real time. The central analysis server is equipped with a GPU accelerator card and stores the graph attention network model and the Bayesian fusion inference model. After receiving multi-dimensional state feature vectors from multiple wind turbine generators, it performs graph structure construction, attention weight calculation, and health status score derivation. The database, deployed in the wind farm's local data center, uses a time-series database format to store historical multimodal sensor signals, model inference results, and maintenance records. It supports indexing and querying by timestamp, generator number, and component type.

[0007] Thirdly, the present invention provides a health status detection device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the intelligent health status detection method for wind turbine generator sets in the smart wind farm described in the first aspect.

[0008] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the intelligent health status detection method for wind turbine generator sets in the first aspect described above.

[0009] Fifthly, the present invention provides a computer program product that, when running on a health status detection device, causes the health status detection device to execute the intelligent health status detection method for smart wind farm wind turbine generator sets described in the first aspect.

[0010] The advantages of this invention compared to existing technologies are as follows: In this invention, the synchronous acquisition of multimodal sensing signals covers four physical dimensions: mechanical, acoustic, thermal, and electromagnetic, avoiding the missed detection of weak early fault features by a single sensor; the time-frequency domain joint feature extraction process employs adaptive mathematical transformation methods for different signal types, preserving the non-stationary characteristics during fault evolution; the graph attention network model utilizes the inherent mechanical topology of the wind turbine generator to construct graph node relationships, enabling health status assessment to not only rely on the intrinsic signal but also incorporate the abnormal propagation effects of adjacent components, enhancing the contextual awareness capability of diagnosis; the Bayesian fusion inference model introduces operating condition parameters as conditional variables, effectively suppressing the interference of normal operating condition disturbances such as high wind speeds or pitch maneuvers on health scoring; the entire system completes front-end signal processing through edge computing units, reducing communication bandwidth requirements, while the central analysis server uses a unified model to centrally evaluate all units in the field, avoiding the inconsistency in evaluation standards caused by dispersed models in traditional methods. This method requires no additional dedicated fault diagnosis hardware; it relies solely on existing sensor layouts for signal enhancement and intelligent fusion to achieve highly sensitive identification of early faults such as bearing micro-pitting, gear micro-cracks, and localized winding overheating. This significantly improves the practicality and engineering deployability of wind turbine health status monitoring. (See attached figures.) Figure 1 This is a schematic diagram of the intelligent health status detection system for wind turbine generator sets in a smart wind farm, provided in an embodiment of the present invention.

[0011] The attached figures are labeled as follows: 1. Multimodal sensor array; 2. Edge computing unit; 3. Communication module; 4. Central analysis server; 5. Database. Detailed implementation methods are described below. Specific embodiments of the present invention combined with Figure 1 The structural diagram of the intelligent health status detection system for wind turbine generators in a smart wind farm, as shown, is explained in detail. Figure 1 As shown, the system consists of five main parts: a multimodal sensor array 1, an edge computing unit 2, a communication module 3, a central analysis server 4, and a database 5. The parts are connected by wired or wireless means to form a complete technical link from field signal acquisition, local preprocessing, remote transmission to centralized analysis and data archiving.

[0012] A multimodal sensor array 1 is deployed on key mechanical and electrical components from the bottom of the wind turbine tower to the interior of the nacelle. This array includes four types of sensing devices: vibration acceleration sensors, acoustic emission sensors, temperature sensors, and electromagnetic noise probes, which are respectively installed on the main bearing housing, gearbox housing, generator stator winding ends, and tower flange connections. The vibration acceleration sensor is vertically fastened to the outer surface of the main bearing housing with M8 bolts. Its sensitive axis is arranged radially to capture high-frequency impact vibrations caused by micro-pitting or localized spalling during the rotation of the main bearing. The acoustic emission sensor uses a magnetic base to adhere to the outer wall of the gearbox housing, and its coverage area is located near the lubricating oil return channel to effectively receive transient elastic wave signals generated by the propagation of micro-cracks during gear meshing. The temperature sensor is a PT100 platinum resistance type, embedded in the inner side of the generator stator winding insulation layer. Adjacent measuring points are evenly distributed along the circumference of the winding, with the spacing controlled within 15 cm, to monitor abnormal local temperature rise in the winding. The electromagnetic noise probe is connected to the generator output terminal block through a shielded twisted-pair cable. The probe shell is grounded to pick up broadband electromagnetic interference signals excited by stator / rotor air gap magnetic field distortion or winding turn short circuits. All four types of sensors are led out through industrial-grade shielded cables and connected to the signal conditioning module in the control cabinet at the bottom of the tower.

[0013] The raw analog signals output by the multimodal sensor array 1 are transmitted to the edge computing unit 2 via shielded cables. The edge computing unit 2 is an embedded industrial computer installed inside the control cabinet at the bottom of the tower. Its input ports are connected to the signal conditioning modules in the multimodal sensor array 1 via an industrial Ethernet switch. The edge computing unit 2 incorporates an anti-aliasing low-pass filter, setting cutoff frequencies of 10 kHz and 200 kHz for the vibration acceleration signal and acoustic emission signal respectively to prevent high-frequency noise aliasing. Simultaneously, a high-precision analog-to-digital converter uniformly samples all analog signals to a 48 kHz sampling rate and performs timestamp alignment to ensure the synchronization of the multimodal signals in the time dimension. Subsequently, edge computing unit 2 performs preliminary feature extraction: It performs three-level wavelet packet decomposition on the vibration acceleration signal, selects the db4 wavelet basis function, calculates the energy of each sub-band, and obtains the energy entropy coefficient after normalization; it performs short-time Fourier transform on the acoustic emission signal, with a window length of 1024 points and an overlap rate of 75%, extracting the shift of the spectral centroid over time as the center frequency shift feature; for the temperature gradient signal, based on the real-time temperature values ​​measured by two adjacent temperature sensors, combined with the physical distance between the measuring points and the material's thermal conductivity parameters, it calculates the difference in heat conduction rate per unit time; for the electromagnetic noise signal, it uses empirical mode decomposition to decompose it into several intrinsic mode functions, and then performs Hilbert transform on the dominant modes to obtain the instantaneous amplitude fluctuation sequence. The above processing results are encapsulated into a multi-dimensional state feature vector, with each dimension corresponding to a specific feature parameter. The vector has 28 dimensions and contains 7 feature indicators for each of the 4 key components.

[0014] Edge computing unit 2 sends the generated multi-dimensional state feature vector to communication module 3 via an industrial Ethernet interface. Communication module 3 is an industrial-grade communication module supporting the 5G NR standard, installed on a bracket inside the nacelle. Its input is connected to the industrial Ethernet switch inside the tower via fiber optic patch cords, and its output is connected to the backbone network of the wind farm's SCADA system via a fiber optic ring network deployed within the wind farm. Communication module 3 uses the MQTT protocol to encapsulate and upload feature vector data packets. The packet header includes the turbine number, timestamp, and data checksum to ensure integrity and traceability during transmission. Under normal operating conditions, the data upload cycle is once every 10 seconds. When edge computing unit 2 detects that any feature indicator exceeds a preset threshold, it triggers an event-driven mechanism, immediately uploading the current feature vector and attaching an alarm flag.

[0015] Central Analysis Server 4 is deployed in the wind farm's control center room, equipped with an NVIDIA A100 GPU accelerator card and 64 GB of memory, running a Linux operating system. This server maintains a real-time connection with Communication Module 3 via a fiber optic ring network, receiving multi-dimensional state feature vectors from all wind turbine generators across the farm. Central Analysis Server 4 internally loads a pre-trained graph attention network model and a Bayesian fusion inference model. The graph attention network model has a fixed number of four nodes, corresponding to the four key components: the main bearing, gearbox, generator rotor, and tower connecting flange. The edges between nodes are constructed based on the actual mechanical transmission paths and thermal coupling relationships of the wind turbine generators: the main bearing and gearbox have a direct mechanical connection, with an initial edge weight of 0.8; the gearbox and generator rotor are connected via a high-speed shaft, with an edge weight of 0.75; the tower connecting flange and main bearing are structurally rigidly connected and have a thermal conduction path, with an edge weight of 0.6; no edges are set between other non-directly related components. During each inference process, the central analysis server 4 first initializes the feature representation of each node based on the received feature vector, and then executes a three-layer graph attention mechanism: the first layer aggregates information from directly adjacent nodes, the second layer introduces the influence of second-order neighborhoods, and the third layer performs global context awareness. Each layer uses the LeakyReLU activation function and a learnable attention coefficient matrix. Finally, four local health indices are output, with values ​​ranging from 0 to 1. The lower the value, the worse the component's health status.

[0016] After obtaining the local health index, the central analysis server 4 calls the Bayesian fusion inference model to calculate the overall health status score. This model uses the local health index as the observed variable and the current wind speed range, power output level, and yaw angle change rate as conditional variables. The wind speed range is divided into four levels: [0, 5) m / s, [5, 12) m / s, [12, 20) m / s, and ≥20 m / s, provided in real time by the SCADA system. The power output level is expressed as a percentage of rated power and is divided into three categories: low load (<30%), medium load (30%~70%), and high load (>70%). The yaw angle change rate is obtained by differentiating the yaw encoder signal in the cabin, with units of ° / s. Based on a conditional probability table obtained from historical fault sample statistics, the Bayesian model uses the maximum a posteriori probability criterion to derive the probability distribution of the overall health status belonging to the three categories of "normal," "warning," or "fault," and selects the category with the highest probability as the current status judgment result. The overall health status score is defined as P(normal)×1.0 + P(warning)×0.6 + P(fault)×0.2. A warning is triggered when the score is below 0.7, and a fault is determined when the score is below 0.4.

[0017] Central analysis server 4 writes the local health index, overall health status score, status classification results, and original feature vectors obtained from each inference into database 5. Database 5 is deployed in the server cluster of the wind farm's local data center, using the InfluxDB time-series database format. Each record contains fields such as timestamp, unit number, component type, feature vector dimensions, local health index, operating condition parameters, overall score, and status label. Database 5 establishes a composite index, supporting fast queries by time range, unit ID list, or component type. For example, it can retrieve "the local health index change trend of all #12 unit main bearings from June 1st to June 30th, 2024". In addition, database 5 also stores historical operating datasets used for model training, including labeled fault samples and their corresponding multimodal sensor signal raw data, for subsequent model iteration and optimization.

[0018] The components form a closed-loop system through the aforementioned connections and data flow. The multimodal sensor array 1 is fixed to the physical structure of the wind turbine generator and connected to the edge computing unit 2 via a shielded cable. The edge computing unit 2 connects to the communication module 3 via an industrial Ethernet switch. The communication module 3 communicates with the central analysis server 4 via a fiber optic ring network. The central analysis server 4 interacts with the database 5 via a high-speed network link. The entire system is physically deployed in three layers: a field layer (multimodal sensor array 1 and edge computing unit 2), a transmission layer (communication module 3), and a central layer (central analysis server 4 and database 5). Data exchange between these layers is achieved through standardized interface protocols. During the model training phase, the central analysis server 4 reads historical operational datasets from the database 5. This dataset contains at least 500 labeled samples, each covering multimodal sensor signals and corresponding status labels within 72 hours prior to the fault occurrence. During training, the loss function consists of two parts: the mean squared error between the predicted local health index and the actual degradation curve, and the cross-entropy loss between the overall health status classification result and the true label. The two are weighted and summed and then backpropagated end-to-end through the Adam optimizer until the validation set loss converges.

[0019] In actual operation, taking a 1.5 MW doubly-fed wind turbine as an example, when early micropitting occurs in the gearbox input shaft bearing, the vibration acceleration signal energy increases significantly in the 3–5 kHz frequency band, and the wavelet packet energy entropy coefficient increases from 0.32 to 0.48. Simultaneously, due to poor lubrication leading to increased friction, the center frequency of the acoustic emission signal shifts from 120 kHz to 145 kHz; the difference in thermal conductivity between adjacent temperature measurement points increases from 0.8 W / (m·K) to 1.5 W / (m·K); and the standard deviation of the instantaneous amplitude fluctuation of electromagnetic noise increases from 0.12 V to 0.21 V. These characteristics are extracted by the edge computing unit 2 and uploaded to the central analysis server 4. In the graph attention network model, the gearbox node receives minor anomaly information (local health index 0.85) from the main bearing node and calculates a local health index of 0.62 based on its own characteristics. The Bayesian model, considering operating parameters such as a current wind speed of 8.3 m / s (medium wind speed zone), power output of 65% of rated value, and yaw angle change rate of 0.5° / s, calculates an overall health status score of 0.68, classifying it as a "warning" state, and stores this result in database 5. After viewing the abnormal status alert of the unit through the SCADA human-machine interface, maintenance personnel can retrieve relevant feature trend charts from database 5 to arrange planned maintenance and prevent further deterioration of the fault.

[0020] The above implementation process fully embodies the entire workflow of this invention, from signal acquisition, feature extraction, graph structure modeling, working condition fusion to state determination and data storage. The physical connections, data interfaces, processing logic, and deployment locations between each component are clearly defined, enabling those skilled in the art to reproduce the technical solution of this invention based on the content of this specification. To better enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0021] When the wind turbine is in operation, the multimodal sensor array 1 continuously collects physical signals from key components: the vibration acceleration sensor is fixed to the outer surface of the main bearing housing, with its radial sensitive axis perpendicular to the plane of rotation of the main shaft, so that the transient impact caused by bearing micro-pitting generates energy concentration in the 3–5 kHz frequency band; the acoustic emission sensor is adsorbed near the lubricating oil return channel of the gearbox housing, and the elastic wave energy released by the expansion of micro-cracks on the gear meshing surface is mainly concentrated in the 100–200 kHz frequency band, and the magnetic base ensures that the acoustic impedance between the sensor and the housing is matched, reducing signal attenuation; the temperature sensor is embedded in the inner side of the generator stator winding insulation layer, and the distance between adjacent measuring points is controlled within 15 cm, so that the temperature rise gradient caused by local inter-turn short circuit can be accurately captured; the electromagnetic noise probe is connected to the generator output terminal block through a shielded twisted pair cable, and its grounded shell suppresses common-mode interference, thereby effectively picking up broadband electromagnetic disturbance signals excited by air gap eccentricity or winding insulation degradation. The above four types of signals are transmitted to the signal conditioning module in the control cabinet at the bottom of the tower via industrial-grade shielded cables, and then connected to the edge computing unit 2.

[0022] Edge computing unit 2 first applies anti-aliasing low-pass filtering to each analog signal, with the cutoff frequency of the vibration signal set to 10 kHz and the acoustic emission signal set to 200 kHz to prevent high-frequency components from folding into the effective frequency band. Then, it uses a high-precision analog-to-digital converter to uniformly sample all signals to 48 kHz and adds precise timestamps to each channel based on a hardware synchronous triggering mechanism to achieve alignment of multimodal signals on a millisecond-level time scale. Based on this, edge computing unit 2 performs targeted feature extraction: for vibration acceleration signals, a three-level wavelet packet decomposition is performed using the db4 wavelet basis, dividing the frequency band into 8 sub-bands, calculating the energy proportion of each sub-band and obtaining the normalized energy entropy, which is sensitive to impact faults; for acoustic emission signals, a short-time Fourier transform with a 1024-point Hanning window and 75% overlap is used to calculate the position of the spectral centroid frame by frame, and its offset over time reflects the crack propagation rate; for temperature signals, the real-time temperature values ​​T1 and T2 of two adjacent measuring points are used, combined with the physical distance d between the measuring points and the thermal conductivity λ of the copper winding, according to the formula Δq = λ·|T1 The difference in heat flux density per unit area is calculated using T2| / d. An abnormal increase in this value indicates local overheating. For electromagnetic noise signals, several intrinsic mode functions (IMFs) are first obtained through empirical mode decomposition. The two IMFs with the highest energy proportions are selected as the dominant modes, and then a Hilbert transform is applied to them to obtain an instantaneous amplitude sequence. The standard deviation of this sequence characterizes the instability of the electromagnetic disturbance. Finally, seven features are extracted from each of the four types of components to form a 28-dimensional state feature vector, which is encapsulated by edge computing unit 2 and sent to communication module 3 via industrial Ethernet.

[0023] Communication module 3 is deployed on a support frame inside the nacelle. Its input is connected to an industrial Ethernet switch inside the tower via fiber optic patch cords, and its output is connected to the wind farm's fiber optic ring network. Under normal operating conditions, communication module 3 uploads feature vectors every 10 seconds using the MQTT protocol. The data packet header includes the unit number, UTC timestamp, and CRC32 checksum. When edge computing unit 2 detects that any feature exceeds a dynamic threshold (e.g., vibration energy entropy > 0.45), it immediately triggers an event-driven upload mechanism, adding an alarm flag to ensure that abnormal data is transmitted first. This mechanism ensures both the continuity of routine monitoring and meets the real-time response requirements for early faults.

[0024] After receiving the feature vectors of all units in the field, the central analysis server 4 first constructs a graph structure: with the main bearing, gearbox, generator rotor, and tower flange as four nodes, edges are set according to the mechanical connection relationship—the main bearing and gearbox are directly coupled through the low-speed shaft, and are assigned an initial weight of 0.8; the gearbox and generator rotor are connected through the high-speed shaft, and the weight is set to 0.75; the tower flange and main bearing are rigidly connected in the tower structure and have a heat conduction path, and the weight is set to 0.6; no edges are set for other node pairs without direct physical connection. Subsequently, the graph attention network performs three layers of information propagation: in the first layer, each node aggregates the feature representations of its first-order neighboring nodes, and the attention coefficient is obtained by concatenating the learnable parameter matrix and node features and then activating it with LeakyReLU; the second layer introduces a second-order neighborhood, for example, the generator rotor node can indirectly obtain the state information of the main bearing; the third layer achieves context awareness through global pooling operation, and finally outputs four local health indices in the range of 0–1. This process utilizes the inherent topology of wind turbine units, so that gearbox anomalies are not only determined by their own acoustic emission and vibration characteristics, but also affected by the weak vibrations transmitted by the main bearing, thereby improving the ability to identify coupled faults.

[0025] After obtaining the local health indices, the central analysis server 4 invokes a Bayesian fusion inference model. This model uses four local health indices as observed variables and wind speed range, power load level, and yaw rate of change provided by the SCADA system as conditional variables. For example, when the wind speed is 8.3 m / s (within the [5,12) m / s range), the power output is 65% (medium load), and the yaw rate of change is 0.5° / s (stable yaw), the model queries a pre-stored conditional probability table to calculate P(normal|observation, operating condition), P(warning|observation, operating condition), and P(fault|observation, operating condition), and generates an overall health score according to the weighted formula P(normal)×1.0 + P(warning)×0.6 + P(fault)×0.2. This scoring mechanism, through operating condition constraints, avoids misjudging normal vibrations at high wind speeds as faults and can also identify acoustic emission signals that should be weak but are abnormally enhanced under low load.

[0026] Finally, the central analysis server 4 writes the local health index, overall score, status label, and original feature vector into database 5. Database 5 adopts the InfluxDB time-series format, with each record using a timestamp as the primary key, and establishes a composite index of unit number, component type, and operating parameters, supporting maintenance personnel to quickly retrieve the health index evolution curve of a specific component of a specific unit within a specific time period. Simultaneously, the historical fault samples stored in database 5 are used for model iteration: the central analysis server 4 periodically extracts at least 500 sets of labeled samples (including complete multimodal data from the 72 hours prior to the fault) from the database, using the mean squared error between the predicted local health index value and the actual degradation trajectory, and the overall state classification cross-entropy as the joint loss, and updates the graph attention network and Bayesian model parameters through the Adam optimizer to achieve adaptive system evolution.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method and system for intelligent detection of the health status of a wind turbine generator unit in a smart wind farm, characterized in that, include: Multimodal sensing signals generated by several key components of a wind turbine generator during operation are collected. These signals include vibration acceleration signals, acoustic emission signals, temperature gradient signals, and electromagnetic noise signals. The key components include the main bearing, gearbox, generator rotor, and tower connecting flange. Time-frequency domain joint feature extraction is performed on the multimodal sensing signals to generate multidimensional state feature vectors corresponding to each key component. This extraction includes wavelet packet decomposition of the vibration acceleration signals to obtain the energy entropy coefficient, short-time Fourier transform of the acoustic emission signals to extract the center frequency shift, calculation of the heat conduction rate difference between adjacent measuring points for the temperature gradient signals, and Hilbert-Huang transform of the electromagnetic noise signals to obtain the instantaneous amplitude fluctuation sequence. The multidimensional state feature vectors are then input into a pre-trained graph attention network model. The nodes of the attention network model correspond to key components, and the edge weights are dynamically set according to the mechanical transmission path and thermal coupling relationship between components. The state information of adjacent nodes is aggregated through a multi-layer graph attention mechanism to output the local health index of each key component. Based on the local health index and the overall operating parameters of the wind turbine generator set, including wind speed range, power output level and yaw angle change rate, a Bayesian fusion inference model is constructed to calculate the overall health status score of the wind turbine generator set, and classify the current operating status as normal, warning or fault according to the preset health level threshold. In the training phase, the graph attention network model uses historical operating datasets for end-to-end optimization. The historical operating datasets contain labeled fault samples and corresponding multimodal sensor signals. The model loss function is composed of the local health index prediction error and the overall health status classification cross-entropy.

2. The intelligent wind farm wind turbine generator unit health state intelligent detection method of claim 1, wherein, The vibration acceleration signal is acquired by a vibration acceleration sensor, which is vertically attached to the surface of the main bearing housing by bolt fastening; the acoustic emission signal is acquired by an acoustic emission sensor, which is magnetically attached to the outer wall of the gearbox housing and covers the lubricating oil return channel area using a magnetic base; the temperature gradient signal is acquired by a temperature sensor, which is embedded in the inner side of the generator stator winding insulation layer with the distance between adjacent measuring points not exceeding 15 cm; the electromagnetic noise signal is acquired by an electromagnetic noise probe, which is connected to the generator output terminal block via a shielded cable. 3.The intelligent wind farm wind turbine generator unit health state intelligent detection method of claim 1, wherein, The multidimensional state feature vector is generated in the edge computing unit, which performs anti-aliasing filtering, sampling rate unification, and timestamp alignment on the original multimodal sensing signal. The anti-aliasing filter cutoff frequencies for the vibration acceleration signal and the acoustic emission signal are 10 kHz and 200 kHz, respectively, and the unified sampling rate is 48 kHz.

4. The intelligent wind farm wind turbine generator unit health state intelligent detection method of claim 1, wherein, In the graph attention network model, there is an edge between the main bearing node and the gearbox node with an initial edge weight of 0.8, an edge between the gearbox node and the generator rotor node with an initial edge weight of 0.75, an edge between the tower connecting flange node and the main bearing node with an initial edge weight of 0.6, and no edges between the remaining node pairs.

5. The intelligent wind farm wind turbine generator unit health state intelligent detection method of claim 1, wherein, In the Bayesian fusion inference model, the wind speed range is divided into four levels: [0, 5) m / s, [5, 12) m / s, [12, 20) m / s, and ≥20 m / s. The power output level is divided into three categories: below 30% of rated power, 30% to 70% of rated power, and above 70% of rated power. The yaw angle change rate is obtained by differentiating the yaw encoder signal in the nacelle.

6. A smart wind farm wind turbine generator unit health state intelligent detection system, characterized in that, The system includes a multimodal sensor array (1), an edge computing unit (2), a communication module (3), a central analysis server (4), and a database (5). The multimodal sensor array (1) is fixedly installed on the main bearing housing, gearbox housing, generator stator winding end, and tower flange connection of the wind turbine generator set, and is used to collect vibration acceleration signals, acoustic emission signals, temperature gradient signals, and electromagnetic noise signals. The edge computing unit (2) is integrated in the control cabinet at the bottom of the tower and is connected to the multimodal sensor array (1) through an industrial Ethernet switch. It is used to perform anti-aliasing filtering, sampling rate unification, and preliminary feature extraction on the original sensor signals, and generate multidimensional signals. The communication module (3) adopts a 5G industrial module, which is deployed on the bracket inside the nacelle and connected to the wind farm SCADA system through an optical fiber ring network. It is used to transmit the multidimensional state feature vector to the central analysis server (4) in real time. The central analysis server (4) is equipped with a GPU acceleration card and stores a graph attention network model and a Bayesian fusion inference model. It is used to perform graph structure construction, attention weight calculation and health status score derivation after receiving the multidimensional state feature vector. The database (5) is deployed in the local data center of the wind farm and uses a time-series database format to store historical multimodal sensor signals, model inference results and operation and maintenance records.

7. The intelligent wind farm wind turbine health condition intelligent detection system of claim 6, wherein, The vibration acceleration sensor in the multimodal sensor array (1) is vertically fastened to the outer surface of the main bearing seat by M8 bolts. The acoustic emission sensor covers an area near the lubricating oil return channel. The temperature sensor is a PT100 platinum resistance type and adjacent measuring points are evenly distributed along the circumference of the winding. The electromagnetic noise probe shell is grounded and connected to the generator output terminal block through a shielded twisted pair cable.

Images