Intelligent temperature measurement and early warning system for thermal imagery of fan spindle based on double-branch neural network

Abstract

The invention relates to the field of monitoring and intelligent operation and maintenance of wind power generation equipment, and discloses a double-branch neural network-based intelligent temperature measurement and early warning system for thermal imagery of a fan spindle. Comprising a customized thermal imaging acquisition module, a multi-dimensional data sensing module, a data preprocessing module, a dual-branch neural network fusion calculation module and a dynamic threshold early warning and interaction module. The customized thermal imaging acquisition module acquires an infrared thermal image and a visible light image, and the multi-dimensional data sensing module synchronously acquires vibration and working condition parameters; the data preprocessing module executes deblurring, filtering and edge detection; the double-branch neural network fusion calculation module fuses the space and time sequence characteristics and outputs a main shaft temperature prediction value; and the dynamic threshold value early warning and interaction module calculates a dynamic temperature threshold value according to a working condition correction formula and implements graded early warning and load reduction control. Through multi-source data fusion and dynamic threshold adjustment, the effect of improving the temperature measurement precision of the fan main shaft and the fault early warning reliability is achieved.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment monitoring and intelligent operation and maintenance, specifically to a dual-branch neural network-based intelligent temperature measurement and early warning system for wind turbine main shaft thermal imaging. Background Technology

[0002] During long-term operation, the main shaft and its bearings of a wind turbine generator bear the main mechanical load of the entire unit and are crucial transmission components connecting the impeller and the generator. Under continuous operation under high load, variable speed, and complex environments, the bearings and sealing structures are highly susceptible to abnormal temperature rise due to frictional heat, poor lubrication, or assembly deviations. If not detected and addressed promptly, this can easily lead to grease carbonization, bearing burn-out, or even main shaft breakage, resulting in unit downtime and significant economic losses. Therefore, high-precision, real-time, and reliable temperature monitoring and early warning of the wind turbine main shaft are of significant engineering importance.

[0003] Existing methods for monitoring the temperature of wind turbine main shafts mainly fall into two categories: contact temperature sensing and non-contact infrared thermometry. Contact temperature sensors are typically installed on the inner wall of the bearing housing or in the lubrication lines. While they can directly reflect local temperature changes, they are prone to poor contact or signal drift under high speed and strong vibration conditions, and are difficult to install and maintain. Non-contact infrared thermometry, although capable of long-distance measurement, is subject to interference from strong light reflection, solar radiation, and dust obstruction inside the wind turbine nacelle, resulting in large fluctuations in temperature readings and making it difficult to reflect the overall thermal state of the main shaft. Furthermore, traditional systems often use fixed thresholds to determine temperature anomalies, failing to fully consider the dynamic changes in external operating conditions such as wind speed, load, and ambient temperature, leading to frequent false alarms and missed alarms.

[0004] With the development of artificial intelligence and image recognition technology, some studies have attempted to introduce convolutional neural networks (CNNs) or long short-term memory networks (LSTMs) to analyze thermal imaging data in order to improve prediction accuracy. However, existing neural network models are mostly designed for single data sources (such as infrared images or vibration signals), lack the ability to fuse multi-source heterogeneous information, fail to effectively integrate environmental parameters and operating condition data, and do not solve problems such as asynchronous data timing, complex illumination changes, and unclear operating condition coupling relationships in the wind turbine field environment. Therefore, in practical engineering applications, there are still defects such as poor model stability and insufficient adaptability. To address this, we propose a dual-branch neural network-based intelligent temperature measurement and early warning system for wind turbine main shaft thermal imaging. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a dual-branch neural network-based intelligent thermal imaging temperature measurement and early warning system for wind turbine main shafts, thereby solving the technical problems existing in the prior art.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A dual-branch neural network-based intelligent temperature measurement and early warning system for wind turbine main shaft thermal imaging includes: a customized thermal imaging acquisition module for real-time acquisition of infrared thermal image data and visible light images of the surface lubricating grease of the wind turbine main shaft; the customized thermal imaging acquisition module is equipped with a vibration-damping bracket and a narrow-band filter lens to reduce the interference of nacelle vibration and solar radiation on the temperature measurement accuracy. The multi-dimensional data sensing module is used to synchronously collect operating condition data such as engine room vibration signals, wind speed, solar radiation intensity, ambient temperature and humidity, spindle speed, load torque, and lubricating oil pressure, and synchronize the data with the thermal imaging data through a unified timestamp. The data preprocessing module is used to perform preprocessing operations on the collected multi-source data, including: Perform rotation speed-adaptive deblurring on the infrared thermal image data; Kalman filtering is applied to the vibration signal to remove high-frequency noise; Perform edge detection on visible light images to identify areas of dried grease or contamination. A dual-branch neural network fusion computation module is used to perform feature extraction and fusion computation on preprocessed data, and includes: The dynamic feature branch is used to extract spatial features and instantaneous correlation features from infrared thermal imaging data, visible light images and vibration time-series signals. The dynamic feature branch includes a convolutional neural network layer and a gated recurrent unit layer, and outputs a dynamic feature vector. The operating condition feature branch is used to extract time-series trends and multi-parameter coupled features from environmental parameters and operating condition parameters. The operating condition feature branch includes a long short-term memory network layer and outputs an operating condition feature vector. The attention fusion layer is used to calculate weights based on the similarity between dynamic feature vectors and operating condition feature vectors, and to perform weighted fusion of features in the spindle bearing and sealing areas to generate a fused feature vector. The temperature output layer is used to input the fused feature vector into the fully connected neural network to calculate and output the real-time temperature values ​​of the inner and outer rings and seals of the spindle bearing. The dynamic threshold warning and interaction module is used to calculate the dynamic temperature threshold based on the temperature output value and real-time operating parameters, and to trigger graded warnings according to temperature deviations. Specifically, it includes: An early warning is triggered when the temperature exceeds the dynamic threshold by 2-3℃. A mid-term warning is triggered when the temperature exceeds the dynamic threshold by 3-5℃. When the temperature exceeds the dynamic threshold by 5°C, a severe warning is triggered and a load reduction command is sent to the wind turbine control system. At the same time, the early warning information will be pushed to the operation and maintenance platform and mobile terminals to realize remote monitoring and maintenance linkage.

[0007] Preferably, the customized thermal imaging acquisition module includes an infrared thermal imaging acquisition unit and a visible light acquisition unit; The infrared thermal image acquisition unit is a vibration-resistant infrared thermal imager, which is installed on the inner wall of the cabin and fixedly connected to the support structure through a vibration damping bracket. The infrared thermal imager has a resolution of 1280×1024 and a frame rate of 100Hz. An 850nm narrowband filter lens is installed at the front of the lens to shield infrared interference caused by direct sunlight. The visible light acquisition unit and the infrared thermal imaging acquisition unit are coaxially mounted with a resolution of 1920×1080. The acquired visible light image covers the surface of the main shaft bearing and sealing area of ​​the fan. The field of view deviation between the visible light acquisition unit and the infrared thermal imaging acquisition unit does not exceed 1° to ensure the spatial registration accuracy of the two types of images in subsequent data fusion. The infrared thermal imaging acquisition unit operates in an ambient temperature range of -40℃ to 70℃ and is equipped with a dustproof and waterproof housing with a protection level of not less than IP67.

[0008] Preferably, the multi-dimensional data sensing module includes a vibration detection unit, an environmental detection unit, and a working condition data interface unit; The vibration detection unit includes a triaxial accelerometer, which is installed on the nacelle shell near the main shaft to collect vibration acceleration data of the nacelle in the X, Y, and Z directions. The measurement accuracy of the triaxial accelerometer is not less than ±0.01g, and the sampling frequency is not less than 100Hz. The environmental monitoring unit includes a wind speed and direction sensor, a solar radiation sensor, and a temperature and humidity sensor. The wind speed and direction sensor is installed on an unobstructed area on the top of the nacelle, with a measurement accuracy of ±0.2 m / s; the solar radiation sensor has a measurement accuracy of ±10 W / m. 2 The temperature measurement accuracy of the temperature and humidity sensor is ±0.5℃, and the humidity measurement accuracy is ±3%RH.

[0009] Preferably, the operating condition data interface unit is connected to the fan control system via a CAN bus to acquire operating condition parameters such as spindle speed, load torque, and lubricating oil pressure. The interface unit has a UTC timestamp function to ensure that the operating condition data and thermal imaging data are synchronized within 0.1 seconds. Each sensor output of the multi-dimensional data sensing module is connected to the data acquisition controller via a signal conditioning circuit. The controller is used to uniformly manage the sampling frequency and timestamp allocation, and transmits the acquired data to the data preprocessing module via wired or wireless communication.

[0010] Preferably, the data preprocessing module includes an infrared thermal image preprocessing unit, a vibration data filtering unit, and a visible light image processing unit. The infrared thermal image preprocessing unit performs adaptive deblurring on the image based on the spindle rotation speed. When the spindle rotation speed is 10 to 15 r / min, a 3×3 Gaussian deblurring kernel is used with a standard deviation of 1.2; when the spindle rotation speed is 15 to 20 r / min, a 5×5 Gaussian deblurring kernel is used with a standard deviation of 1.5. By calculating the motion vectors of adjacent frames, the pixel displacement caused by the spindle rotation is corrected, and the corrected infrared thermal image data is output. The vibration data filtering unit uses the Kalman filtering algorithm to filter the input vibration signal in real time at a sampling frequency of 100Hz. The state equation and observation equation are established based on the vibration amplitude and temperature change, respectively. The filtering cutoff frequency is set to within 10Hz to remove high-frequency vibration interference and retain low-frequency components.

[0011] Preferably, the visible light image processing unit includes an image enhancement subunit and an edge detection subunit. The image enhancement subunit performs histogram equalization processing on the input image to improve the grayscale contrast between the grease and metal regions. The edge detection subunit uses the Canny algorithm with a threshold range of 100 to 200 to extract the boundary contour of the grease region and output a feature map containing dry cracks or dirt marks. The outputs of the infrared thermal image preprocessing unit, the vibration data filtering unit, and the visible light image processing unit are transmitted to the dual-branch neural network fusion calculation module through the data synchronization interface. The sampling time interval is no more than 0.1 seconds, and the timestamp synchronization error is no more than 10 milliseconds.

[0012] Preferably, the dual-branch neural network fusion computing module includes a dynamic feature branch, a working condition feature branch, an attention fusion layer, and a temperature output layer; the dynamic feature branch includes a convolutional neural network subunit and a gated recurrent unit subunit; The convolutional neural network subunit has three convolutional layers with 64, 128 and 256 kernels respectively. The kernel size is 3×3 and the stride is 1. Each convolutional layer is followed by a 2×2 max pooling layer to reduce dimensionality. The activation function of the convolutional layer is ReLU and the output is flattened to form a 256-dimensional spatial feature vector. The gated recurrent unit subunit includes a gated recurrent network with 128 hidden units and a dropout probability of 0.2. It is used to receive the time series data of vibration signals and extract time-series correlation features, and output a 64-dimensional time-series feature vector. The outputs of the convolutional neural network subunit and the gated recurrent unit subunit are concatenated to form a 320-dimensional dynamic feature vector.

[0013] Preferably, the operating condition feature branch includes two layers of long short-term memory network, with 64 hidden units in each layer and a dropout probability of 0.2. The first layer is used to extract the time variation trend of a single parameter, and the second layer is used to extract multi-parameter coupled features, outputting a 160-dimensional operating condition feature vector. The operating condition feature branch processes the input time-series data of wind speed, solar radiation, temperature and humidity, spindle speed, load torque and lubricating oil pressure, with a sampling period of no more than 0.5 seconds. The attention fusion layer calculates the correlation matrix between the dynamic feature vector and the working condition feature vector based on the additive attention mechanism. It then performs a weighted summation of highly correlated features using normalized weights to generate a 480-dimensional fusion feature vector, where the feature weights for the bearing and sealing areas are no less than 0.6. The temperature output layer consists of two fully connected layers. The first layer has 256 nodes and uses ReLU as the activation function. The second layer has 32 nodes and uses a linear function to output the temperature values ​​of the inner and outer rings of the spindle bearing and the seals. The temperature output layer is trained using mean squared error as the loss function, and the temperature calculation delay does not exceed 50 milliseconds.

[0014] Preferably, the dynamic threshold warning and interaction module includes a threshold calculation unit, a warning judgment unit, a control linkage unit, and an information interaction unit; the threshold calculation unit determines the dynamic temperature threshold based on the spindle material parameters and real-time working condition data, the dynamic temperature threshold is obtained by superimposing a base temperature threshold and a correction coefficient, the base temperature threshold is the recommended upper limit of the bearing material; The correction factor is calculated using the following formula:

[0015] in, This represents the percentage change in load torque. This represents the change in wind speed. This refers to the deviation in lubricating oil pressure. , , These are correction factors for torque, wind speed, and lubricating oil pressure, with values ​​ranging from 0.05 to 0.25, -0.3 to 0.05, and -0.4 to 0.1, respectively. The threshold calculation unit updates the dynamic temperature threshold in real time during each sampling period.

[0016] Preferably, the early warning determination unit calculates the difference between the real-time temperature value from the temperature output layer and the dynamic temperature threshold, and triggers a graded early warning based on the temperature difference range; A temperature difference between 2°C and 3°C is designated as a Level 1 warning. A temperature difference between 3°C and 5°C is designated as a Level II warning. A temperature difference greater than 5℃ is designated as a Level III warning. The early warning determination unit encodes the early warning status at each level into an independent identification signal and outputs it to the control linkage unit; The control linkage unit is communicatively connected to the wind turbine control system, receives the warning indicator signal output by the warning judgment unit, and generates a load reduction control command in the third-level warning state to limit the wind turbine power output to within 30% of the rated power. The information interaction unit includes a communication interface and a data push terminal. The communication interface uses the MODBUS protocol to connect with the wind farm monitoring system and uploads temperature data and early warning levels to the operation and maintenance platform. At the same time, it sends early warning information to the mobile terminal through the wireless communication module, supporting text and graphical display, with an update cycle of no more than 1 second. The units of the dynamic threshold warning and interaction module are interconnected through a system bus to form a closed-loop signal path from temperature detection to warning response.

[0017] In summary, the present invention has the following main beneficial effects: By setting up a customized thermal imaging acquisition module and a multi-dimensional data sensing module, synchronous acquisition of infrared thermal images, visible light images, vibration signals, and environmental operating parameters of the wind turbine main shaft is achieved. The customized thermal imaging acquisition module is equipped with a vibration-damping bracket and a narrow-band filter lens, effectively eliminating interference from nacelle vibration and solar radiation. The multi-dimensional data sensing module synchronizes the three-axis acceleration, wind speed and direction, solar radiation, and temperature and humidity signals with operating information such as main shaft speed, load torque, and lubricating oil pressure in a unified time frame. Through the above-mentioned multi-source synchronous acquisition structure, the spatiotemporal consistency of temperature measurement data and environmental adaptability are ensured, enabling the main shaft temperature monitoring results to have high accuracy and stability under different operating conditions.

[0018] By constructing a dual-branch neural network fusion computing structure comprising a dynamic feature branch and an operating condition feature branch, spatial features of infrared thermal images and visible light images, temporal features of vibration signals, and trend features of operating condition parameters are extracted and fused through branching and attention fusion. The convolutional neural network and gated recurrent units of the dynamic feature branch achieve coupled modeling of image and temporal information, while the long short-term memory network of the operating condition feature branch extracts multi-parameter coupling patterns. The attention fusion layer assigns high weights to the bearing and sealing areas based on feature similarity. Through this dual-branch fusion structure, the temperature distribution of key parts of the spindle can be accurately predicted under complex operating conditions, and the interpretability of features and the generalization ability of the model can be improved.

[0019] By introducing a dynamic threshold early warning and interaction structure based on a working condition correction formula, the dynamic temperature threshold, including correction factors for load torque, wind speed, and lubricating oil pressure, is calculated in real time. This, combined with a tiered judgment logic, enables early warning and load reduction control. The early warning judgment unit eliminates transient interference through continuous sampling verification. In a severe early warning state, the control linkage unit issues a load reduction command to the wind turbine control system, and the information interaction unit simultaneously pushes the early warning results to the operation and maintenance platform and mobile terminals. Through this dynamic threshold calculation and tiered response mechanism, the system achieves real-time prevention of main shaft overheating risks throughout the entire wind turbine operation cycle, improving system safety and operation and maintenance timeliness. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 refer to Figure 1 A dual-branch neural network-based intelligent temperature measurement and early warning system for wind turbine main shaft thermal imaging includes: a customized thermal imaging acquisition module for real-time acquisition of infrared thermal image data and visible light images of the surface lubricating grease of the wind turbine main shaft; the customized thermal imaging acquisition module is equipped with a vibration-damping bracket and a narrow-band filter lens to reduce the interference of nacelle vibration and solar radiation on the temperature measurement accuracy. The multi-dimensional data sensing module is used to synchronously collect operating condition data such as engine room vibration signals, wind speed, solar radiation intensity, ambient temperature and humidity, spindle speed, load torque, and lubricating oil pressure, and synchronize the data with the thermal imaging data through a unified timestamp. The data preprocessing module is used to perform preprocessing operations on the collected multi-source data, including: Perform rotation speed-adaptive deblurring on the infrared thermal image data; Kalman filtering is applied to the vibration signal to remove high-frequency noise; Perform edge detection on visible light images to identify areas of dried grease or contamination. A dual-branch neural network fusion computation module is used to perform feature extraction and fusion computation on preprocessed data, and includes: The dynamic feature branch is used to extract spatial features and instantaneous correlation features from infrared thermal imaging data, visible light images and vibration time-series signals. The dynamic feature branch includes a convolutional neural network layer and a gated recurrent unit layer, and outputs a dynamic feature vector. The operating condition feature branch is used to extract time-series trends and multi-parameter coupled features from environmental parameters and operating condition parameters. The operating condition feature branch includes a long short-term memory network layer and outputs an operating condition feature vector. The attention fusion layer is used to calculate weights based on the similarity between dynamic feature vectors and operating condition feature vectors, and to perform weighted fusion of features in the spindle bearing and sealing areas to generate a fused feature vector. The temperature output layer is used to input the fused feature vector into the fully connected neural network to calculate and output the real-time temperature values ​​of the inner and outer rings and seals of the spindle bearing. The dynamic threshold warning and interaction module is used to calculate the dynamic temperature threshold based on the temperature output value and real-time operating parameters, and to trigger graded warnings according to temperature deviations. Specifically, it includes: An early warning is triggered when the temperature exceeds the dynamic threshold by 2-3℃. A mid-term warning is triggered when the temperature exceeds the dynamic threshold by 3-5℃. When the temperature exceeds the dynamic threshold by 5°C, a severe warning is triggered and a load reduction command is sent to the wind turbine control system. At the same time, the early warning information will be pushed to the operation and maintenance platform and mobile terminals to realize remote monitoring and maintenance linkage.

[0023] Customized thermal imaging acquisition module: The customized thermal imaging acquisition module includes an infrared thermal imaging acquisition unit and a visible light acquisition unit. These two units work together to achieve multi-spectral imaging and simultaneous acquisition of temperature information in the main shaft area of ​​the wind turbine. The infrared thermal imaging acquisition unit uses a vibration-resistant infrared thermal imager with a high-strength aluminum alloy shell and elastic vibration-damping washers at the lens connection point. It is fixedly mounted on a rigid frame on the inner wall of the nacelle via a vibration-damping bracket. This mounting position maintains a distance of 80-100mm from the central axis of the main shaft, ensuring complete coverage of the main shaft bearing and sealing area without interfering with the main shaft rotation.

[0024] To reduce the impact of wind vibration and electromagnetic interference on imaging during cabin operation, the infrared thermal imager is connected to the cabin support structure through multiple layers of rubber vibration damping pads and aluminum-based fixing plates, forming a dual-layer structure that balances mechanical vibration isolation and thermal conductivity. This allows the camera to maintain stable operation in a temperature environment ranging from -40℃ to 70℃. The overall protection level is IP67, which can prevent dust, water vapor and oil mist from entering the imaging cavity and ensure long-term reliable operation in high humidity and high salt spray environments.

[0025] The core imaging device of the infrared thermal imager uses an uncooled focal plane detector with a resolution of 640×512 to 1280×1024 and a frame rate of no less than 100Hz to ensure that continuous thermal image sequences can still be obtained under the condition of low-speed rotation of the main shaft (10-20r / min). A narrowband filter lens with a center wavelength of 850nm and a bandwidth of 10nm is installed at the front of the lens. This filter lens is used to suppress infrared interference caused by direct sunlight and reflection of external visible light, thereby ensuring that the infrared radiation signal comes only from the radiation energy of the main shaft surface itself. This filter lens is connected to the thermal imager barrel through a threaded detachable bayonet, which can be cleaned or replaced during maintenance.

[0026] The visible light acquisition unit is a high-definition industrial camera, coaxially mounted with the infrared thermal imager via a metal mounting bracket. The optical axes of both are factory-calibrated to have a parallel deviation of no more than 1°, and the difference in field of view is controlled within 0.5°. This coaxial mounting structure enables pixel-level spatial registration of the two types of images (infrared thermal image and visible light image) during subsequent data fusion, avoiding feature misalignment caused by parallax.

[0027] The visible light acquisition unit has a resolution of 1920×1080 and a frame rate of 60Hz. The lens is a low-distortion wide-angle lens with a focal length of 5mm to cover the surface of the outer ring of the spindle bearing, the sealing end cap, and the adjacent grease area. Its imaging area completely overlaps with the thermal imaging area of ​​the infrared thermal imager, so that the system can obtain dual-modal image information of the same spatial position at the same time.

[0028] When the visible light acquisition unit is installed inside the cabin, it is equipped with an anti-reflective coated glass cover. The cover has a light transmittance of over 90%, which can prevent local overexposure caused by reflected light from cabin lighting or electrical equipment. At the same time, its image signal is connected to the signal conditioning module through a shielded data cable to avoid the influence of electromagnetic noise generated by the high-voltage motor in the cabin on the image signal.

[0029] Both the infrared thermal imaging unit and the visible light acquisition unit have a built-in unified UTC timestamp module, and acquisition is initiated by a synchronization trigger signal sent by the main control board. The trigger delay between the two is no more than 5ms.

[0030] The synchronization mechanism ensures that the time segments corresponding to the infrared image and the visible light image are consistent at any angle position of the spindle rotation. In addition, during the system debugging phase, multi-point temperature calibration is performed using a standard blackbody radiation source to establish a mapping relationship between infrared grayscale values ​​and actual temperature, and this relationship is stored in the lookup table of the main control chip to ensure that the temperature measurement accuracy reaches ±0.5℃.

[0031] Multi-dimensional data perception module: The multi-dimensional data sensing module includes a vibration detection unit, an environmental detection unit, and a working condition data interface unit. This module works collaboratively with multiple sensors and a bus interface to achieve comprehensive sensing of the engine room's operating status, external environmental parameters, and spindle working conditions. It also synchronizes the data with the image data output from the infrared thermal imaging module, providing complete input data for subsequent neural network feature fusion.

[0032] The vibration detection unit includes at least one triaxial accelerometer for real-time acquisition of vibration acceleration signals of the nacelle in the X, Y, and Z directions. The triaxial accelerometer preferably employs a MEMS structure, offering high sensitivity and strong shock resistance. The sensor is mounted on the inner wall of the nacelle shell near the main shaft bearing seat using a stainless steel mounting bracket. Its mounting surface is perpendicular to the main shaft axis, ensuring strict correspondence between the three axes and the nacelle's spatial coordinate axes. To avoid resonance errors in the sensor caused by naval vibrations, a silicone damping pad with a thickness of 3-5mm is installed between the sensor base and the inner wall of the naval cabin. The data cable uses twisted-pair shielded cable to reduce electromagnetic interference. The triaxial accelerometer has a measurement range of ±16g, a measurement accuracy of no less than ±0.01g, and a sampling frequency of no less than 100Hz. It can completely capture the low-to-medium frequency vibration characteristics caused by spindle imbalance, bearing wear, and loose couplings. The vibration signal is amplified by the signal conditioning circuit and converted from analog to digital before being input to the data acquisition controller for time registration with the thermal imaging data.

[0033] The environmental monitoring unit includes wind speed and direction sensors, a solar radiation sensor, and temperature and humidity sensors. This unit measures aerodynamic, radiative, and meteorological conditions in the wind turbine's operating environment, providing necessary input for subsequent temperature threshold correction and model training. The wind speed and direction sensors are installed on the unobstructed area of ​​the nacelle top, fixed to the casing with a metal bracket to ensure unobstructed sensor blades and accurate reflection of inflow wind speed and direction. The measurement accuracy is ±0.2 m / s, the output signal is an RS485 digital signal, and the sampling period is 1 second. When the wind speed change exceeds 2 m / s, the system automatically marks the infrared thermal image data for that time period for subsequent image removal to eliminate image disturbances caused by strong winds. The solar radiation sensor is installed on the nacelle top near the wind direction sensor, with the probe facing the sky, and its measurement range is 0-2000 W / m². 2 The measurement accuracy is ±10W / m 2 The sensor is used to capture changes in solar radiation intensity and prevent thermal drift errors caused by direct sunlight on the infrared lens. The real-time measured radiation data will be used as one of the input features of the neural network to correct the temperature drift caused by radiation during the model fusion stage.

[0034] Temperature and humidity sensors are installed inside the cabin ventilation openings to monitor the air temperature and humidity of the internal environment. The temperature measurement accuracy is ±0.5℃, and the humidity measurement accuracy is ±3%RH. The data is used to correct the baseline of the internal ambient temperature of the thermal imager to avoid temperature measurement deviations caused by high humidity conditions. The signal outputs of all environmental sensors are centrally fed into the data acquisition controller through the signal conditioning module and recorded with a unified UTC timestamp to ensure consistency with the infrared and visible light data acquisition time.

[0035] The operating condition data interface unit is connected to the fan control system via a CAN bus to read operating parameters such as spindle speed, load torque, and lubricating oil pressure in real time. This interface unit has a built-in UTC clock chip with time stamping and data buffering functions. Each piece of operating condition data has a timestamp field with a time error of no more than 0.1 seconds, enabling frame-level synchronization with infrared thermal imaging and vibration signals. The data interface supports the CAN2.0B protocol with a communication rate of 500kbps and CRC check function, ensuring data reliability even in environments with strong electromagnetic interference. The sampling period is 0.5 seconds. All operating condition parameters are sent to the data acquisition controller via the bus and stored in a buffer before being transmitted to the data preprocessing module via an Ethernet interface.

[0036] The sampling of each sub-unit in the multi-dimensional data perception module is uniformly scheduled and controlled by the main control board. When the main control board sends a unified synchronization trigger signal, the vibration detection unit, the environmental detection unit, and the working condition data interface unit simultaneously perform sampling and embed the timestamp information into the data frame header. The multi-source data after time synchronization is uniformly received by the preprocessing module, so that infrared, visible light, vibration, environmental and working condition information are aligned on the same time dimension, providing a complete dataset for subsequent feature fusion operations.

[0037] Data preprocessing module: The data preprocessing module is located between the multi-dimensional data perception module and the dual-branch neural network fusion computing module. It is used to standardize and suppress noise in the acquired multi-source data such as infrared thermal images, visible light images, and vibration signals. The module consists of three parts: an infrared thermal image preprocessing unit, a vibration data filtering unit, and a visible light image processing unit. The multi-source data processed by this module all have a unified UTC timestamp and are transmitted to the dual-branch neural network module through a synchronization interface to ensure the consistency of data temporal, spatial, and statistical characteristics.

[0038] The infrared thermal imaging preprocessing unit is used to eliminate image blurring caused by the rotation of the wind turbine main shaft and vibration of the nacelle, improving the pixel stability of temperature calculation. It receives raw infrared frame data from a customized thermal imaging acquisition module and dynamically adjusts the image deblurring parameters based on the main shaft rotation speed signal. When the main shaft rotation speed is in the range of 10–15 r / min, the system uses a 3×3 Gaussian deblurring kernel with a standard deviation σ set to 1.2; when the main shaft rotation speed is in the range of 15–20 r / min, it uses a 5×5 Gaussian deblurring kernel with a standard deviation σ set to 1.5. The Gaussian kernel convolution operation is performed by the embedded GPU to achieve real-time image filtering. To correct the pixel plane displacement caused by the principal axis rotation, the infrared thermal image preprocessing unit calculates the optical flow vector field of two adjacent frames, and uses an improved Lucas-Kanade algorithm to perform integral compensation on the displacement field, re-registering the image to the principal axis reference coordinate system; the registered image is then processed by grayscale normalization and brightness equalization, and the output is a 256×256 pixel normalized thermal image matrix for use as input to the subsequent neural network convolutional layer; Through adaptive deblurring and motion compensation processing, the stability and comparability of thermal imaging data can be maintained under different rotational speeds and vibration conditions.

[0039] The vibration data filtering unit is used to suppress noise and preserve features of vibration signals.

[0040] This unit adopts a real-time filtering structure based on the Kalman filter algorithm. The state equation and observation equation are expressed as follows:

[0041] in, For the first The true state vector of vibration at any given time. The signal measured by the sensor, Here is the state transition matrix. For the observation matrix, , The system noise and observation noise are represented; the system noise covariance matrix and observation noise covariance matrix are based on the real-time temperature change. The rate of change of vibration amplitude is dynamically updated.

[0042] The filtering frequency is set to 100Hz, and the filtering cutoff frequency is within 10Hz. High-frequency components above this frequency are identified as noise and attenuated, thus preserving the low-frequency vibration characteristics of the bearing. The filtered data is used to calculate the effective value of vibration acceleration and spectral characteristic parameters through envelope analysis, which are then used as the timing signal input to the neural network. The filtering unit uses an independent MCU chip to implement hardware-level operations, ensuring that the delay does not exceed 10 milliseconds, thus avoiding the real-time performance issues that occur in software filtering.

[0043] The visible light image processing unit includes an image enhancement subunit and an edge detection subunit; The image enhancement subunit performs histogram equalization on the original image to improve the grayscale contrast between the grease and metal areas, making the boundaries clear and distinguishable. The edge detection subunit employs the Canny operator with a threshold range of 100-200, and eliminates isolated pseudo-edge points using a dual-threshold concatenation algorithm. The final output is a grayscale-normalized edge feature map. In the feature map, the system uses a binary mask to mark areas of grease cracking and contamination, providing salient feature information for the neural network. The entire image processing is completed by an image signal processor (ISP), with an output frame rate of 30 frames per second, ensuring a one-to-one temporal correspondence with the infrared image.

[0044] The outputs of the infrared thermal image preprocessing unit, vibration data filtering unit, and visible light image processing unit are all transmitted to the dual-branch neural network fusion calculation module via the data synchronization interface; the sampling interval of each signal is no greater than 0.1 seconds, and the timestamp synchronization error is no more than 10 milliseconds; the data synchronization interface is implemented based on the TCP / IP protocol, and sorts and caches the data of each channel using the timestamp as an index to ensure that the data sequence input to the neural network corresponds consistently within the same time window.

[0045] The dual-branch neural network fusion calculation module is located between the data preprocessing module and the dynamic threshold warning and interaction module. It is used to extract features and perform fusion calculations on preprocessed infrared thermal image data, visible light images, vibration time series signals and operating parameters, and finally output the real-time temperature values ​​of the inner ring, outer ring and seal of the spindle bearing. The dual-branch neural network fusion calculation module includes four parts: dynamic feature branch, operating condition feature branch, attention fusion layer and temperature output layer.

[0046] The dynamic feature branch is used to extract spatial and instantaneous correlation features from infrared thermal imaging data, visible light images, and vibration time-series signals.

[0047] Dynamic feature branches include convolutional neural network sub-units and gated recurrent unit sub-units; The input to the convolutional neural network (CNN) subunit is a 256×256 pixel infrared thermal image matrix and a grayscale edge image. The CNN subunit contains three convolutional layers with kernel numbers of 64, 128, and 256 respectively, a kernel size of 3×3, and a stride of 1. Each convolutional layer is followed by a 2×2 max-pooling layer to reduce feature dimensionality and improve robustness. The activation function for the convolutional layers is ReLU, which suppresses feature vanishing due to negative gradients. The convolutional output is flattened to form a 256-dimensional feature vector.

[0048] The gated recurrent unit (GRU) sub-unit is used to process vibration time-series signals. It adopts a single-layer gated recurrent network structure with 128 hidden units and a dropout probability of 0.2 to prevent overfitting. The input data is a vibration acceleration sequence processed by Kalman filtering, with a sampling time interval of 0.01 seconds. The GRU uses update and reset gate mechanisms to model short-term and long-term vibration dependencies, outputting a 64-dimensional time-series feature vector.

[0049] The output vectors of the convolutional neural network subunit and the gated recurrent unit subunit are concatenated to form a 320-dimensional dynamic feature vector. This dynamic feature vector is then weighted and fused with the operating condition feature vector in the subsequent attention fusion layer.

[0050] The operating condition feature branch is used to extract the temporal trends and multi-parameter coupling features of environmental parameters and operating condition parameters; the operating condition feature branch includes two layers of long short-term memory network; The first layer of the Long Short-Term Memory (LSTM) network takes as input time-series data wind speed, solar radiation, ambient temperature, ambient humidity, spindle speed, load torque, and lubricating oil pressure. This layer has 64 hidden units with a dropout probability of 0.2, primarily used to extract the temporal trends of each individual parameter.

[0051] The second layer of the Long Short-Term Memory network takes the output features of the first layer as input and uses them to extract the coupling relationship and mutual influence pattern between multiple parameters. The number of hidden units is also 64, and the discard probability is 0.2. The final output is a 160-dimensional feature vector of the working condition; The sampling period of the working condition feature branch is no more than 0.5 seconds to ensure that it is consistent with the time window of the dynamic feature branch.

[0052] The attention fusion layer is used to calculate weighted weights based on the similarity between dynamic feature vectors and working condition feature vectors, thereby achieving deep fusion of multi-source features.

[0053] The attention fusion layer uses an additive attention mechanism to calculate the relevance matrix:

[0054] in, The i-th eigenvector of the dynamic feature branch; Let j be the j-th eigenvector of the working condition feature branch; , This is the weight matrix; This is the weight vector; For bias terms; The attention weight matrix is ​​obtained after the relevance matrix is ​​normalized by Softmax. :

[0055] Using the weight matrix The fused feature vector is obtained by weighted summation of the feature vectors under different operating conditions.

[0056] To increase attention to key areas, the attention fusion layer sets the lower limit of the feature weights corresponding to the bearing and sealing areas to 0.6, preventing the model from ignoring key heat source areas during training; after fusion, a 480-dimensional fused feature vector is output as the input to the temperature output layer.

[0057] The temperature output layer is used to convert the fused feature vector into temperature output values; the temperature output layer consists of two fully connected neural network layers. The first fully connected layer has 256 nodes and uses ReLU activation function to perform non-linear mapping and compress feature dimensions. The second fully connected layer has 32 nodes and uses a linear activation function to output the real-time temperature values ​​of the inner and outer rings and seals of the spindle bearing. The temperature output layer is trained using mean squared error (MSE) as the loss function:

[0058] in, For the actual measured temperature; The model was used to predict temperature. Training employed the Adam optimization algorithm with a learning rate of 0.001, a batch size of 64, and a convergence threshold of 10^-5. The trained model achieved a temperature prediction latency of no more than 50 milliseconds during on-site wind turbine operation.

[0059] Before system deployment, the dual-branch neural network fusion computing module was trained offline using sample data. The training data came from multi-source datasets collected during actual wind turbine operation, including no less than 3,000 hours of infrared thermal images, vibration signals, and operating condition records.

[0060] The dataset was divided into a training set (80%), a validation set (10%), and a test set (10%). During training, the model parameters were automatically adjusted based on the error on the validation set. The final model achieved an average temperature prediction error of ±0.4℃ on the test set.

[0061] After training, the network weight parameters are stored in the neural network acceleration chip on the system's main control board, enabling online inference.

[0062] The dynamic threshold warning and interaction module is located at the output end of the dual-branch neural network fusion calculation module. It is used to calculate the dynamic temperature threshold, perform graded warning judgment, and realize load reduction linkage and information push with the wind turbine control system based on the main shaft bearing temperature value and real-time operating parameters output by the neural network. The dynamic threshold early warning and interaction module includes a threshold calculation unit, an early warning judgment unit, a control linkage unit, and an information interaction unit. These units are sequentially connected via a system bus, forming a closed-loop signal path from temperature data input to early warning response output. The threshold calculation unit determines the dynamic temperature threshold based on the spindle material parameters and real-time operating data. The dynamic temperature threshold is obtained by superimposing a base temperature threshold and a correction coefficient. The base temperature threshold is the recommended upper limit of the bearing material, which can be preset according to the material table (e.g., approximately 90℃ for GCr15 steel and approximately 110℃ for stainless steel). The threshold calculation unit updates the dynamic temperature threshold correction amount in real time according to the following formula:

[0063] in, This represents the percentage change in load torque. This represents the change in wind speed. This refers to the deviation in lubricating oil pressure. , , These are correction coefficients for torque, wind speed, and lubricating oil pressure, with values ​​ranging from 0.05 to 0.25, -0.3 to 0.05, and -0.4 to 0.1, respectively.

[0064] The threshold calculation unit performs a calculation once per sampling period (0.5 seconds) to update the dynamic temperature threshold in real time. The threshold calculation unit uses FPGA embedded arithmetic logic to achieve fast floating-point calculation, ensuring that the calculation delay does not exceed 10 milliseconds. Compared with the fixed threshold method, the dynamic temperature threshold can be automatically adjusted according to changes in external operating conditions. For example, when the load torque increases or the wind speed decreases, the threshold decreases accordingly, triggering an early warning. When the ambient temperature or lubricating oil pressure increases, the threshold is automatically adjusted upward to avoid false alarms.

[0065] The early warning determination unit is used to calculate the difference between the real-time temperature value from the temperature output layer and the dynamic temperature threshold, and trigger different levels of early warning based on the difference range. The early warning judgment unit calculates the temperature difference. ,in This is the real-time temperature value; For dynamic temperature threshold; according to The corresponding warning signal will be generated for the area in question: when When the temperature is between 2℃ and 3℃, it is classified as a Level 1 warning; when When the temperature is between 3℃ and 5℃, it is classified as a Level II warning. when A temperature above 5℃ is considered a Level 3 warning.

[0066] The warning determination unit includes a comparison logic circuit and a temperature difference storage register. To prevent false judgments caused by transient noise, the warning determination unit only triggers the corresponding level of warning when the temperature difference exceeds the interval threshold for three consecutive sampling cycles (i.e., 1.5 seconds). The warning level is encoded as an independent identification signal: level 1 is 001, level 2 is 010, and level 3 is 100. The encoded signal is sent to the control linkage unit through the output interface.

[0067] The control linkage unit is connected to the wind turbine control system and is used to perform automatic load reduction operations in the event of a severe warning.

[0068] The control linkage unit receives the warning indicator signal from the warning determination unit and generates control commands based on the signal level: When a Level 3 warning signal 100 is received, the control linkage unit generates a load reduction control command to limit the wind turbine power output to within 30% of the rated power. The control command is sent to the wind turbine pitch controller and power controller via the CAN bus, along with a timestamp and safety identification code to prevent false triggering.

[0069] The control linkage unit is equipped with a redundant safety mechanism: if no confirmation feedback signal is received from the wind turbine control system within 3 consecutive sampling cycles, the load reduction command will be automatically resent; if there is still no response, a local audible and visual alarm will be triggered and the event log will be recorded; the control linkage unit uses an ARM processor to implement logic judgment and bus communication, and the processing delay does not exceed 20 milliseconds.

[0070] The information interaction unit is used to enable the interaction between early warning information and the operation and maintenance platform and mobile terminals.

[0071] The information interaction unit includes a communication interface and a data push terminal. The communication interface uses the MODBUS protocol to connect with the wind farm monitoring system and is responsible for uploading temperature data, threshold information, and early warning levels.

[0072] The data push terminal communicates with the operation and maintenance server in real time via a 4G or Ethernet wireless module; the warning information is encapsulated in JSON format and includes fields such as: timestamp, spindle number, real-time temperature, dynamic threshold, warning level, control status, and confirmation status. The mobile terminal application supports both text and graphical display modes, with an update cycle of no more than 1 second. When the system is in a level 3 warning state, the information interaction unit automatically pushes an alarm interface with a red indicator, along with historical temperature curves for manual confirmation. All push records are synchronized to the cloud database for use in operation and maintenance logs and model retraining.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dual-branch neural network-based intelligent thermal imaging temperature measurement and early warning system for wind turbine main shaft, characterized in that, include: A customized thermal imaging acquisition module is used to acquire infrared thermal image data and visible light images of the surface lubricating grease of the wind turbine main shaft in real time. The customized thermal imaging acquisition module is equipped with a vibration-damping bracket and a narrow-band filter lens to reduce the interference of nacelle vibration and solar radiation on temperature measurement accuracy. The multi-dimensional data sensing module is used to synchronously collect operating condition data such as engine room vibration signals, wind speed, solar radiation intensity, ambient temperature and humidity, spindle speed, load torque, and lubricating oil pressure, and synchronize the data with the thermal imaging data through a unified timestamp. The data preprocessing module is used to perform preprocessing operations on the collected multi-source data, including: Perform rotation speed-adaptive deblurring on the infrared thermal image data; Kalman filtering is applied to the vibration signal to remove high-frequency noise; Perform edge detection on visible light images to identify areas of dried grease or contamination. A dual-branch neural network fusion computation module is used to perform feature extraction and fusion computation on preprocessed data, and includes: The dynamic feature branch is used to extract spatial features and instantaneous correlation features from infrared thermal imaging data, visible light images and vibration time-series signals. The dynamic feature branch includes a convolutional neural network layer and a gated recurrent unit layer, and outputs a dynamic feature vector. The operating condition feature branch is used to extract time-series trends and multi-parameter coupled features from environmental parameters and operating condition parameters. The operating condition feature branch includes a long short-term memory network layer and outputs an operating condition feature vector. The attention fusion layer is used to calculate weights based on the similarity between dynamic feature vectors and operating condition feature vectors, and to perform weighted fusion of features in the spindle bearing and sealing areas to generate a fused feature vector. The temperature output layer is used to input the fused feature vector into the fully connected neural network to calculate and output the real-time temperature values ​​of the inner and outer rings and seals of the spindle bearing. The dynamic threshold warning and interaction module is used to calculate the dynamic temperature threshold based on the temperature output value and real-time operating parameters, and to trigger graded warnings according to temperature deviations. Specifically, it includes: An early warning is triggered when the temperature exceeds the dynamic threshold by 2-3℃. A mid-term warning is triggered when the temperature exceeds the dynamic threshold by 3-5℃. When the temperature exceeds the dynamic threshold by 5°C, a severe warning is triggered and a load reduction command is sent to the wind turbine control system. At the same time, the early warning information will be pushed to the operation and maintenance platform and mobile terminals to realize remote monitoring and maintenance linkage.

2. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 1, characterized in that, The customized thermal imaging acquisition module includes an infrared thermal imaging acquisition unit and a visible light acquisition unit; The infrared thermal image acquisition unit is a vibration-resistant infrared thermal imager, which is installed on the inner wall of the cabin and fixedly connected to the support structure through a vibration damping bracket. The infrared thermal imager has a resolution of 1280×1024 and a frame rate of 100Hz. An 850nm narrowband filter lens is installed at the front of the lens to shield infrared interference caused by direct sunlight. The visible light acquisition unit and the infrared thermal imaging acquisition unit are coaxially mounted with a resolution of 1920×1080. The acquired visible light image covers the surface of the main shaft bearing and sealing area of ​​the fan. The field of view deviation between the visible light acquisition unit and the infrared thermal imaging acquisition unit does not exceed 1° to ensure the spatial registration accuracy of the two types of images in subsequent data fusion. The infrared thermal imaging acquisition unit operates in an ambient temperature range of -40℃ to 70℃ and is equipped with a dustproof and waterproof housing with a protection level of not less than IP67.

3. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 2, characterized in that, The multi-dimensional data sensing module includes a vibration detection unit, an environmental detection unit, and a working condition data interface unit. The vibration detection unit includes a triaxial accelerometer, which is installed on the nacelle shell near the main shaft to collect vibration acceleration data of the nacelle in the X, Y, and Z directions. The measurement accuracy of the triaxial accelerometer is not less than ±0.01g, and the sampling frequency is not less than 100Hz. The environmental monitoring unit includes a wind speed and direction sensor, a solar radiation sensor, and a temperature and humidity sensor. The wind speed and direction sensor is installed on an unobstructed area on the top of the nacelle, with a measurement accuracy of ±0.2 m / s; the solar radiation sensor has a measurement accuracy of ±10 W / m. 2 The temperature measurement accuracy of the temperature and humidity sensor is ±0.5℃, and the humidity measurement accuracy is ±3%RH.

4. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 3, characterized in that, The operating condition data interface unit is connected to the fan control system via a CAN bus to acquire operating condition parameters such as spindle speed, load torque, and lubricating oil pressure. The interface unit has a UTC timestamp function to ensure that the operating condition data and thermal imaging data are synchronized within 0.1 seconds. Each sensor output of the multi-dimensional data sensing module is connected to the data acquisition controller via a signal conditioning circuit. The controller is used to uniformly manage the sampling frequency and timestamp allocation, and transmits the acquired data to the data preprocessing module via wired or wireless communication.

5. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 4, characterized in that, The data preprocessing module includes an infrared thermal image preprocessing unit, a vibration data filtering unit, and a visible light image processing unit. The infrared thermal image preprocessing unit performs adaptive deblurring on the image based on the spindle rotation speed. When the spindle rotation speed is 10 to 15 r / min, a 3×3 Gaussian deblurring kernel is used with a standard deviation of 1.2; when the spindle rotation speed is 15 to 20 r / min, a 5×5 Gaussian deblurring kernel is used with a standard deviation of 1.

5. By calculating the motion vectors of adjacent frames, the pixel displacement caused by the spindle rotation is corrected, and the corrected infrared thermal image data is output. The vibration data filtering unit uses the Kalman filtering algorithm to filter the input vibration signal in real time at a sampling frequency of 100Hz. The state equation and observation equation are established based on the vibration amplitude and temperature change, respectively. The filtering cutoff frequency is set to within 10Hz to remove high-frequency vibration interference and retain low-frequency components.

6. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 5, characterized in that, The visible light image processing unit includes an image enhancement subunit and an edge detection subunit. The image enhancement subunit performs histogram equalization processing on the input image to improve the grayscale contrast between the grease and metal areas. The edge detection subunit uses the Canny algorithm with a threshold range of 100 to 200 to extract the boundary contour of the grease area and output a feature map containing dry cracks or dirt marks. The outputs of the infrared thermal image preprocessing unit, the vibration data filtering unit, and the visible light image processing unit are transmitted to the dual-branch neural network fusion calculation module through the data synchronization interface. The sampling time interval is no more than 0.1 seconds, and the timestamp synchronization error is no more than 10 milliseconds.

7. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 6, characterized in that, The dual-branch neural network fusion computing module includes a dynamic feature branch, a working condition feature branch, an attention fusion layer, and a temperature output layer; the dynamic feature branch includes a convolutional neural network subunit and a gated recurrent unit subunit. The convolutional neural network subunit has three convolutional layers with 64, 128 and 256 kernels respectively. The kernel size is 3×3 and the stride is 1. Each convolutional layer is followed by a 2×2 max pooling layer to reduce dimensionality. The activation function of the convolutional layer is ReLU and the output is flattened to form a 256-dimensional spatial feature vector. The gated recurrent unit subunit includes a gated recurrent network with 128 hidden units and a dropout probability of 0.

2. It is used to receive the time series data of vibration signals and extract time-series correlation features, and output a 64-dimensional time-series feature vector. The outputs of the convolutional neural network subunit and the gated recurrent unit subunit are concatenated to form a 320-dimensional dynamic feature vector.

8. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 7, characterized in that, The operating condition feature branch consists of two layers of long short-term memory networks, each with 64 hidden units and a dropout probability of 0.

2. The first layer is used to extract the time variation trend of a single parameter, and the second layer is used to extract the coupled features of multiple parameters, outputting a 160-dimensional operating condition feature vector. The operating condition feature branch processes the time-series data of input wind speed, solar radiation, temperature and humidity, spindle speed, load torque and lubricating oil pressure, with a sampling period of no more than 0.5 seconds. The attention fusion layer calculates the correlation matrix between the dynamic feature vector and the working condition feature vector based on the additive attention mechanism. It then performs a weighted summation of highly correlated features using normalized weights to generate a 480-dimensional fusion feature vector, where the feature weights for the bearing and sealing areas are no less than 0.

6. The temperature output layer consists of two fully connected layers. The first layer has 256 nodes and uses ReLU as the activation function. The second layer has 32 nodes and uses a linear function to output the temperature values ​​of the inner and outer rings of the spindle bearing and the seals. The temperature output layer is trained using mean squared error as the loss function, and the temperature calculation delay does not exceed 50 milliseconds.

9. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 8, characterized in that, The dynamic threshold warning and interaction module includes a threshold calculation unit, a warning judgment unit, a control linkage unit, and an information interaction unit. The threshold calculation unit determines the dynamic temperature threshold based on the spindle material parameters and real-time working condition data. The dynamic temperature threshold is obtained by superimposing a base temperature threshold and a correction coefficient. The base temperature threshold is the recommended upper limit of the temperature corresponding to the bearing material. The correction factor is calculated using the following formula: in, This represents the percentage change in load torque. This represents the change in wind speed. This refers to the deviation in lubricating oil pressure. , , These are correction factors for torque, wind speed, and lubricating oil pressure, with values ​​ranging from 0.05 to 0.25, -0.3 to 0.05, and -0.4 to 0.1, respectively. The threshold calculation unit updates the dynamic temperature threshold in real time during each sampling period.

10. The intelligent temperature measurement and early warning system for fan main shaft thermal imaging based on a dual-branch neural network according to claim 9, characterized in that, The early warning determination unit calculates the difference between the real-time temperature value from the temperature output layer and the dynamic temperature threshold, and triggers graded early warnings based on the temperature difference range. A temperature difference between 2°C and 3°C is designated as a Level 1 warning. A temperature difference between 3°C and 5°C is designated as a Level II warning. A temperature difference greater than 5℃ is designated as a Level III warning. The early warning determination unit encodes the early warning status at each level into an independent identification signal and outputs it to the control linkage unit; The control linkage unit is communicatively connected to the wind turbine control system, receives the warning indicator signal output by the warning judgment unit, and generates a load reduction control command in the third-level warning state to limit the wind turbine power output to within 30% of the rated power. The information interaction unit includes a communication interface and a data push terminal. The communication interface uses the MODBUS protocol to connect with the wind farm monitoring system and uploads temperature data and early warning levels to the operation and maintenance platform. At the same time, it sends early warning information to the mobile terminal through the wireless communication module, supporting text and graphical display, with an update cycle of no more than 1 second. The units of the dynamic threshold warning and interaction module are interconnected through a system bus to form a closed-loop signal path from temperature detection to warning response.

Images