A monitoring and early warning method and system based on fan bolt vibration ultrasonic coupling

Abstract

A monitoring and early warning method and system based on fan bolt vibration ultrasonic coupling relate to the technical field of wind power equipment fatigue fault diagnosis. The vibration signal, ultrasonic signal and displacement signal of the bolt to be monitored are synchronously collected; the vibration signal, ultrasonic signal and displacement signal are compensated and extracted based on the real-time operating condition parameters of the wind turbine generator set, to obtain the vibration characteristic value, ultrasonic characteristic value and real-time displacement value; the vibration characteristic value, ultrasonic characteristic value and real-time displacement value are normalized and input into a preset fault recognition model for processing to obtain state information; the remaining service life is obtained through prediction calculation based on the material parameters of the bolt to be monitored, real-time operating condition parameters and state information; and the warning level and comprehensive maintenance scheme are generated based on the remaining service life, state information and wind resource prediction data. The method solves the problem that a single monitoring method cannot accurately extract early bolt faults in a strong interference environment.

Description

Technical Field

[0001] This application relates to the field of fatigue fault diagnosis technology for wind power equipment, specifically to a monitoring and early warning method and system based on ultrasonic coupling of wind turbine bolt vibration. Background Technology

[0002] As a core component of clean energy, wind turbines are subjected to harsh conditions such as alternating impacts from strong winds, high-frequency vibrations from rotating machinery, and fluctuating outdoor temperatures and humidity on their connecting bolts, which are constantly exposed to these factors. The condition of these bolt connections directly affects the operational safety of the entire turbine. Loosening, slippage, or fatigue fracture due to preload decay can easily lead to major safety accidents such as tower collapse and blade detachment, resulting in huge economic losses.

[0003] Existing wind turbine bolt monitoring technologies mainly rely on single-sensor monitoring methods. For example, vibration acceleration sensors are installed to collect low-frequency macroscopic vibration signals of the unit to analyze the vibration intensity of the overall structure. In addition, some monitoring systems use ultrasonic testing devices, which utilize the propagation characteristics of ultrasonic waves inside the bolts to assess the preload state of the bolts or detect the presence of internal defects by analyzing the changes in the acoustic time or amplitude of the echo signals.

[0004] However, the aforementioned existing technologies are accompanied by strong background vibrations from rotating machinery and instantaneous wind load impacts during wind turbine operation. These high-intensity background noises easily drown out the micro-amplitude signals generated by the micron-level dynamic slippage of bolts and the early characteristic signals of microcrack initiation. Single vibration or acoustic monitoring methods are insufficient to establish an effective anti-interference compensation mechanism in environments with strong interference, and cannot accurately distinguish between background vibrations and bolt fault characteristics. This makes it difficult to accurately capture the early dynamic slippage and microcrack initiation of bolts without shutting down the machine, and thus cannot meet the needs of online real-time monitoring and early warning. Summary of the Invention

[0005] This application provides a monitoring and early warning method and system based on ultrasonic coupling of wind turbine bolt vibration. This method solves the problem that the existing single sensor monitoring method is difficult to accurately extract early fault characteristics of bolts under strong interference environment.

[0006] Firstly, this application provides a monitoring and early warning method based on vibration and ultrasonic coupling of wind turbine bolts. The method includes: receiving vibration signals, ultrasonic signals, and displacement signals synchronously acquired from the bolts to be monitored; acquiring real-time operating parameters of the wind turbine; performing compensation processing on the vibration signals, ultrasonic signals, and displacement signals based on the real-time operating parameters to obtain compensated vibration signals, compensated ultrasonic signals, and compensated displacement signals; analyzing the compensated vibration signals and compensated ultrasonic signals to extract vibration feature values ​​and ultrasonic feature values, respectively; extracting real-time displacement values ​​from the compensated displacement signals; normalizing the vibration feature values, ultrasonic feature values, and real-time displacement values ​​to obtain a coupling feature vector; and inputting the coupling feature vector into a preset fault identification model for processing to obtain... The system obtains the status information of the bolt to be monitored, including no fault information, dynamic slip information, and crack information. When the status information is dynamic slip information or crack information, the system acquires the material parameters of the bolt to be monitored, and performs predictive calculations on the material parameters, real-time operating condition parameters, and dynamic slip information or crack information to obtain the remaining service life of the bolt to be monitored. The system acquires the instantaneous load data at the moment the status information is generated, and extracts the corresponding fault response amplitude from the status information. The system calculates the load sensitivity coefficient of the bolt to be monitored by combining the instantaneous load data and the fault response amplitude. The system acquires wind resource prediction data, and generates an early warning level and a comprehensive maintenance plan based on the remaining service life, load sensitivity coefficient, status information, and wind resource prediction data. The early warning level and comprehensive maintenance plan are then sent to the operation and maintenance terminal.

[0007] By adopting the above technical solution, vibration signals, ultrasonic signals, and displacement signals are collected simultaneously, and multi-level compensation processing is performed based on real-time operating condition parameters. This effectively solves the problem that existing single-sensor monitoring methods are difficult to accurately extract early fault characteristics of bolts under strong interference environments. By filtering out the fundamental frequency and harmonic interference of rotating machinery and wind load impact noise, the sensor sensitivity shift caused by environmental temperature drift and salt spray corrosion is eliminated, and the signal distortion caused by load stress concentration is corrected. This ensures that the compensated vibration signal, compensated ultrasonic signal, and compensated displacement signal can truly reflect the intrinsic state characteristics of the bolt. By extracting vibration feature values, ultrasonic feature values, and real-time displacement values ​​from the compensated signal, and constructing a coupled feature vector input to a preset fault identification model, accurate identification of bolt fault-free information, dynamic slippage information, and crack information is achieved, overcoming the high false alarm rate problem caused by simple threshold judgment in existing technologies. When dynamic slippage information or crack information is identified, prediction calculations are performed by combining material parameters and real-time operating condition parameters to obtain the remaining service life of the bolt to be monitored. By calculating the load sensitivity coefficient and combining it with wind resource prediction data, early warning levels and comprehensive maintenance plans are generated, realizing risk assessment and preventive maintenance decisions, avoiding major safety accidents that may be caused by traditional passive maintenance modes, and significantly improving the accuracy of wind turbine bolt monitoring.

[0008] Optionally, the vibration signal, ultrasonic signal, and displacement signal are compensated based on real-time operating parameters to obtain compensated vibration signal, compensated ultrasonic signal, and compensated displacement signal. Specifically, this includes: obtaining real-time rotational speed and real-time wind speed from the real-time operating parameters; determining the fundamental frequency and harmonic frequencies of the rotating machinery based on the real-time rotational speed; determining the frequency band range of wind load impact based on the real-time wind speed; locating the amplitude components coinciding with the fundamental frequency and harmonic frequencies, as well as the amplitude components within the wind load impact frequency band, in the frequency domain spectrum of the vibration signal to obtain the background vibration interference component; subtracting the background vibration interference component from the vibration signal to obtain the denoised vibration signal; acquiring real-time temperature data and real-time salt spray concentration data around the bolt to be monitored; calculating the temperature difference between the real-time temperature data and the preset calibration temperature; and multiplying the temperature difference by a temperature sensitivity factor. The sensitivity drift is obtained by using a degree offset coefficient. Based on the sensitivity drift, the amplitudes of the denoised vibration signal, ultrasonic signal, and displacement signal are reverse-corrected. The ultrasonic signal is then gain-compensated based on the signal transmission attenuation rate corresponding to the real-time salt spray concentration data, resulting in environmentally corrected vibration signal, environmentally corrected ultrasonic signal, and environmentally corrected displacement signal. Real-time load data is obtained from real-time operating condition parameters. The nominal stress is calculated based on the real-time load data and the cross-sectional area of ​​the bolt to be monitored. The dynamic stress concentration factor is determined based on the nominal stress, and a distortion correction factor is constructed based on the dynamic stress concentration factor. The environmentally corrected vibration signal, environmentally corrected ultrasonic signal, and environmentally corrected displacement signal are then divided by the distortion correction factor to obtain the compensated vibration signal, compensated ultrasonic signal, and compensated displacement signal.

[0009] By employing the above technical solution, real-time rotational speed and wind speed are obtained from real-time operating parameters. The fundamental and harmonic frequencies of the rotating machinery, as well as the frequency band of wind load impact, are determined respectively. Background vibration interference components are accurately located and removed from the frequency domain spectrum of the vibration signal to address the problem of high-frequency vibration of rotating machinery and the submersion of weak fault characteristic signals of bolts by instantaneous wind load impact. Furthermore, by acquiring real-time temperature data and real-time salt spray concentration data, the temperature difference is calculated, and the sensitivity drift is obtained by combining it with a temperature sensitivity offset coefficient. This is used to reverse-correct the denoised vibration signal, ultrasonic signal, and displacement signal. Simultaneously, the unique propagation attenuation characteristics of ultrasonic signals are addressed. Gain compensation is performed to overcome the measurement errors caused by sensitivity drift of sensors in marine salt spray environments in existing technologies. By obtaining real-time load data from real-time operating condition parameters, calculating the nominal stress and determining the dynamic stress concentration factor in combination with the cross-sectional area of ​​the bolt to be monitored, a distortion correction factor is constructed to normalize the environmentally corrected signal, eliminating the nonlinear signal distortion caused by stress concentration of the bolt under alternating loads. This allows the compensated vibration signal, compensated ultrasonic signal, and compensated displacement signal to accurately characterize the true health status of the bolt, improving the accuracy of fault diagnosis in monitoring systems under strong interference environments.

[0010] Optionally, the compensated vibration signal and the compensated ultrasonic signal are analyzed to extract vibration feature values ​​and ultrasonic feature values, respectively. Specifically, this includes: performing wavelet packet decomposition on the compensated vibration signal to obtain sub-signals of different frequency bands; selecting sub-signals within a preset frequency range from the sub-signals of different frequency bands as feature frequency band signals; performing time-domain statistical analysis on the feature frequency band signals to calculate the root mean square value, peak factor, and impulse factor of the feature frequency band signals; combining the root mean square value, peak factor, and impulse factor as vibration feature values; performing short-time energy analysis on the compensated ultrasonic signal to calculate the short-time energy change rate of the ultrasonic signal within a time window function to obtain the acoustic-ultrasonic energy rate; counting the number of times the ultrasonic signal exceeds the trigger threshold; dividing the number of times by the unit time to obtain the count rate; calculating the time interval from the first time the ultrasonic signal exceeds the trigger threshold to the signal peak to obtain the rise time; and combining the acoustic-ultrasonic energy rate, count rate, and rise time as ultrasonic feature values.

[0011] By adopting the above technical solution, for the compensated vibration signal, wavelet packet decomposition technology is used to decompose the signal into sub-signals of different frequency bands, and sub-signals within a preset frequency range are selected as characteristic frequency band signals, effectively focusing on the local high-frequency response range generated by the micron-level dynamic slip of the bolt; by performing time-domain statistical analysis on the characteristic frequency band signals, the extracted root mean square value can reflect the overall energy level of bolt vibration, the peak factor can characterize the abrupt change characteristics of impact failure, and the impulse factor is sensitive to the intermittent impulse signal of early fatigue cracks; for the compensated ultrasonic signal, the ultrasonic energy rate is obtained by calculating the short-time energy change rate within the time window function through short-time energy analysis; the number of times the ultrasonic signal exceeds the trigger threshold and the count rate are calculated, which can quantify the frequency of acoustic emission activity reflecting the initiation of microcracks inside the bolt; the rise time from the first exceeding of the trigger threshold to the signal peak is measured. By combining ultrasonic energy rate, count rate, and rise time as ultrasonic feature values, a multi-dimensional characterization of the internal defect state of bolts is achieved. The collaborative extraction of the two types of feature values ​​overcomes the limitation of the single feature parameter in the existing technology in recognizing complex fault modes, and improves the early recognition accuracy of dynamic slip information and crack information.

[0012] Optionally, real-time displacement values ​​are extracted from the compensated displacement signal, and the vibration feature values, ultrasonic feature values, and real-time displacement values ​​are normalized to obtain a coupled feature vector. Specifically, this includes: performing time-domain analysis on the compensated displacement signal, calculating the average amplitude of the displacement signal within a preset sampling period, and using the average amplitude as the real-time displacement value; constructing a feature dataset from the vibration feature values, ultrasonic feature values, and real-time displacement values; processing each feature data in the feature dataset using a max-min normalization algorithm to obtain multiple normalized feature data; and sorting the multiple normalized feature data according to a preset order to obtain a multi-dimensional coupled feature vector.

[0013] By adopting the above technical solution, time-domain analysis is performed on the compensated displacement signal, and the average amplitude within the preset sampling period is calculated as the real-time displacement value. This effectively filters out instantaneous impact noise and random fluctuation interference. Vibration characteristic values, ultrasonic characteristic values, and real-time displacement values ​​are uniformly constructed into a feature dataset, overcoming the limitation of the single physical quantity representation dimension in the existing technology. By using the maximum-minimum normalization algorithm to process each feature data in the feature dataset, the problem of feature weight imbalance caused by the difference in dimensions and numerical magnitude between different physical quantities is eliminated, so that the normalized feature data are all mapped to a unified numerical range, ensuring that various features have equivalent expressive power in the fault identification model.

[0014] Optionally, the remaining service life of the bolt to be monitored can be predicted and calculated based on material parameters, real-time operating condition parameters, and dynamic slip or crack information. Specifically, this includes: obtaining fracture toughness value, fatigue crack propagation threshold, material constant, and stress-life curve from material parameters; when the status information is crack information, extracting the current crack depth from the crack information, calculating the stress intensity factor amplitude of the bolt to be monitored based on the real-time load data in the real-time operating condition parameters; constructing a functional relationship between crack propagation rate and stress intensity factor amplitude, integrating the current crack depth based on the functional relationship until the crack depth reaches the preset adjacent fracture depth, obtaining the remaining number of cycles, and converting the remaining number of cycles into the remaining service life. When the status information is dynamic slip information, the current slip amount is obtained from the dynamic slip information, and the equivalent fretting wear stress of the contact surface of the bolt to be monitored is calculated based on the current slip value; the corresponding theoretical total fatigue life is found in the stress life curve according to the equivalent fretting wear stress; the historical load cycle statistics of the bolt to be monitored are retrieved from the preset fault feature database, which includes the number of cycles executed under different stress levels; the ratio of the number of cycles executed under each stress level to the corresponding limit cycle number is calculated, and all ratios are summed to obtain the current cumulative fatigue damage degree of the bolt to be monitored; the difference between the theoretical total fatigue life and the current cumulative fatigue damage degree is multiplied by 1 to obtain the remaining service life.

[0015] By adopting the above technical solution, when the status information is crack information, key parameters such as fracture toughness value and fatigue crack propagation threshold are obtained from material parameters. The stress intensity factor amplitude is calculated in conjunction with real-time load data, and the current crack depth is integrated based on a functional relationship to reach a preset critical fracture depth. The resulting remaining cycle count accurately reflects the propagation and evolution law of crack-type failures. When the status information is dynamic slip information, the current slip amount is extracted from the dynamic slip information, and the equivalent fretting wear stress of the bolt contact surface is calculated. The corresponding theoretical total fatigue life is found in the stress-life curve, establishing a quantitative mapping relationship between slip wear and fatigue life. Historical load cycle statistics containing the number of cycles executed under different stress levels are retrieved from the preset fault feature database. Based on the linear cumulative damage theory, the sum of the ratios of the number of cycles executed to the limit cycle count at each stress level is calculated to obtain the current cumulative fatigue damage degree. The consumed life is then subtracted from the theoretical total fatigue life count to obtain the true remaining service life.

[0016] Optionally, instantaneous load data at the moment the state information is generated is acquired, and the corresponding fault response amplitude is extracted from the state information. The instantaneous load data and fault response amplitude are calculated to obtain the load sensitivity coefficient of the bolt to be monitored. Specifically, this includes: acquiring the moment the state is generated; matching the corresponding real-time wind turbine torque data in the real-time operating condition parameters based on the moment of generation; using the real-time wind turbine torque data as the instantaneous load data; when the state information is dynamic slip information, extracting the peak-to-peak displacement value from the real-time displacement value associated with the dynamic slip information as the fault response amplitude; when the state information is crack information, extracting the root mean square value of vibration from the vibration characteristic value associated with the crack information as the fault response amplitude; calculating the ratio of the fault response amplitude to the instantaneous load data to obtain the current load response ratio; and dividing the difference between the current load response ratio and the reference load response ratio by the reference load response ratio to obtain the load sensitivity coefficient.

[0017] By adopting the above technical solution, the moment of state information generation is obtained, and the corresponding real-time wind turbine torque data is accurately matched with the real-time operating condition parameters as instantaneous load data, ensuring the time synchronization of load and fault response and avoiding the problem of transient feature loss caused by the use of average load in existing technologies. Differentiated response amplitude extraction strategies are adopted for different fault modes. When the state information is dynamic slip information, the peak-to-peak displacement value is extracted from the associated real-time displacement value as the fault response amplitude; when the state information is crack information, the root mean square value of vibration is extracted from the associated vibration characteristic value as the fault response amplitude. The ratio of the fault response amplitude to the instantaneous load data is calculated to obtain the current load response ratio. The current load response ratio is normalized and compared with the reference load response ratio. The calculated load sensitivity coefficient can quantify the sensitivity of the bolt's current health state to load changes. The load sensitivity coefficient overcomes the limitation of existing technologies that only rely on static life prediction values ​​for early warning, enabling the generation of early warning levels to comprehensively consider the coupling effect of the current fault severity and future load excitation intensity, significantly improving the timeliness and accuracy of early warning decisions.

[0018] Optionally, the early warning levels include a Level 1 early warning level and a Level 2 early warning level. Wind resource forecast data is acquired, and based on remaining service life, load sensitivity coefficient, status information, and wind resource forecast data, an early warning level and a comprehensive maintenance plan are generated. Specifically, this includes: extracting the maximum predicted wind speed within a preset future time window from the wind resource forecast data; determining the basic health status based on status information and remaining service life; determining the basic health status as high-risk when the status information is crack information or the remaining service life is less than or equal to the service life threshold; determining the basic health status as abnormal when the status information is dynamic slip information and the remaining service life is greater than the service life threshold; calculating the product of the maximum predicted wind speed and the load sensitivity coefficient to obtain the future risk index; generating a Level 1 early warning level when the basic health status is high-risk, or abnormal and the future risk index is greater than or equal to a preset risk threshold; generating a Level 2 early warning level when the basic health status is abnormal and the future risk index is less than the preset risk threshold; generating a maintenance plan including a shutdown command and bolt replacement strategy if the bolt to be monitored is at Level 1; and generating a maintenance plan including a bolt preload re-tightening strategy if the bolt to be monitored is at Level 2.

[0019] By employing the above technical solution, the maximum predicted wind speed within a preset time window is extracted from wind resource forecast data. The basic health status is determined based on status information and remaining service life. A high-risk state is identified when the status information indicates cracking or the remaining service life is less than or equal to a service life threshold. An abnormal state is identified when the status information indicates dynamic slippage and the remaining service life exceeds the service life threshold, thus achieving precise grading of bolt health. The future risk index is obtained by multiplying the maximum predicted wind speed by the load sensitivity coefficient. This future risk index comprehensively reflects the accelerating deterioration effect of future extreme wind conditions on the current fault state. A Level 1 warning is generated when the basic health status is high-risk or, although abnormal, the future risk index exceeds a preset risk threshold. A Level 2 warning is generated when the basic health status is abnormal and the future risk index is low. This approach avoids excessive intervention in low-risk faults while ensuring timely response to high-risk faults. Differentiated comprehensive maintenance solutions are matched for different warning levels. For the first-level warning level, a maintenance solution including shutdown instructions and bolt replacement strategy is generated. For the second-level warning level, a maintenance solution including bolt preload re-tightening strategy is generated. This achieves optimized allocation of maintenance resources and effective control of maintenance costs, maximizing equipment availability while ensuring the safe operation of wind turbine units.

[0020] The second aspect of this application provides a monitoring and early warning system based on vibration and ultrasonic coupling of wind turbine bolts. The system includes an acquisition unit, a compensation unit, an extraction unit, an identification unit, and an early warning unit. The acquisition unit receives vibration signals, ultrasonic signals, and displacement signals synchronously acquired from the bolts to be monitored. The compensation unit acquires real-time operating parameters of the wind turbine and performs compensation processing on the vibration, ultrasonic, and displacement signals based on these parameters to obtain compensated vibration signals, compensated ultrasonic signals, and compensated displacement signals. The extraction unit analyzes the compensated vibration and ultrasonic signals, extracting vibration and ultrasonic feature values ​​respectively. Real-time displacement values ​​are extracted from the compensated displacement signals, and the vibration, ultrasonic, and real-time displacement values ​​are normalized to obtain a coupling feature vector. The identification unit identifies the coupling feature vector. The eigenvector is input into a preset fault identification model for processing to obtain the status information of the bolt to be monitored. The status information includes no fault information, dynamic slip information, and crack information. The early warning unit, when the status information is dynamic slip information or crack information, obtains the material parameters of the bolt to be monitored, and performs prediction calculations on the material parameters, real-time operating condition parameters, and dynamic slip information or crack information to obtain the remaining service life of the bolt to be monitored. It obtains the instantaneous load data at the time of the status information generation, extracts the corresponding fault response amplitude from the status information, and calculates the load sensitivity coefficient of the bolt to be monitored by calculating the instantaneous load data and fault response amplitude. It obtains wind resource prediction data, and generates an early warning level and a comprehensive maintenance plan based on the remaining service life, load sensitivity coefficient, status information, and wind resource prediction data, and sends the early warning level and comprehensive maintenance plan to the operation and maintenance terminal.

[0021] In a third aspect, this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory, causing the electronic device to perform any of the methods described above in this application.

[0022] In a fourth aspect, this application provides a computer-readable storage medium storing instructions that, when executed, perform any of the methods described above in this application.

[0023] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Simultaneously acquire vibration signals, ultrasonic signals, and displacement signals, and perform multi-level compensation processing based on real-time operating parameters. This effectively solves the problem that existing single-sensor monitoring methods are difficult to accurately extract early fault characteristics of bolts under strong interference environments. By filtering out the fundamental frequency and harmonic interference of rotating machinery and wind load impact noise, the sensor sensitivity shift caused by ambient temperature drift and salt spray corrosion is eliminated, and the signal distortion caused by load stress concentration is corrected. This ensures that the compensated vibration signals, compensated ultrasonic signals, and compensated displacement signals can truly reflect the intrinsic state characteristics of the bolt. By extracting vibration feature values, ultrasonic feature values, and real-time displacement values ​​from the compensated signal, and constructing a coupled feature vector input to a preset fault identification model, accurate identification of bolt fault-free information, dynamic slippage information, and crack information is achieved, overcoming the high false alarm rate problem caused by simple threshold judgment in existing technologies. When dynamic slippage information or crack information is identified, prediction calculations are performed by combining material parameters and real-time operating condition parameters to obtain the remaining service life of the bolt to be monitored. By calculating the load sensitivity coefficient and combining it with wind resource prediction data, early warning levels and comprehensive maintenance plans are generated, realizing risk assessment and preventive maintenance decisions, avoiding major safety accidents that may be caused by traditional passive maintenance modes, and significantly improving the accuracy of wind turbine bolt monitoring. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the first process of a monitoring and early warning method based on ultrasonic coupling of wind turbine bolt vibration provided in an embodiment of this application; Figure 2 This is a schematic diagram of the second process of a monitoring and early warning method based on ultrasonic coupling of wind turbine bolt vibration provided in an embodiment of this application; Figure 3 This is a schematic diagram of the third process of a monitoring and early warning method based on ultrasonic coupling of wind turbine bolt vibration provided in an embodiment of this application; Figure 4 This is a schematic diagram of the fourth process of a monitoring and early warning method based on ultrasonic coupling of wind turbine bolt vibration provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

[0025] Explanation of reference numerals in the attached figures: 500, electronic device; 501, processor; 502, memory; 503, user interface; 504, network interface; 505, communication bus. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0028] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0029] Therefore, how to improve the existing single-sensor monitoring method's inability to accurately extract early bolt fault characteristics under strong interference environments is a pressing problem that needs to be solved. This application provides a monitoring and early warning method based on ultrasonic coupling of wind turbine bolt vibration, applied in a server. The server in this application serves as a platform for fault diagnosis of wind turbine bolts. Figure 1 This is a schematic diagram of the first process of a monitoring and early warning method based on ultrasonic coupling of wind turbine bolt vibration provided in an embodiment of this application. (Refer to...) Figure 1 The method includes the following steps S101-S108.

[0030] S101: Receives vibration signals, ultrasonic signals, and displacement signals from the bolt to be monitored, which are collected synchronously.

[0031] In S101 above, in order to achieve accurate online monitoring of wind turbine bolts under complex dynamic conditions such as high-frequency vibration of rotating machinery and alternating impact of strong wind loads, it is necessary to simultaneously collect multi-modal signals that can characterize different fault modes of bolts. This is because single vibration monitoring can only capture low-frequency macroscopic vibration characteristics and is not sensitive to micro-amplitude signals of micron-level dynamic slip. Traditional ultrasonic testing is an offline static method that cannot capture the real-time slip and crack propagation process of bolts under dynamic conditions. Therefore, this step constructs a dynamic synchronous acquisition system of vibration-ultrasound-displacement three-modal signals to achieve comprehensive online monitoring of dynamic slip caused by bolt preload attenuation, microcrack initiation in stress concentration areas, and relative displacement of connection surfaces.

[0032] To address the challenges of strong vibrations and outdoor salt spray corrosion in wind turbine applications, the sensors were optimized for vibration and corrosion resistance. For the low-frequency accelerometer, a dual-layer encapsulation system consisting of a silicone rubber buffer layer and a 316L stainless steel anti-corrosion shell was adopted. The silicone rubber buffer layer was 0.3mm thick to reduce the impact of high-frequency vibrations from rotating machinery on the sensor, while the stainless steel anti-corrosion shell was 0.6mm thick to resist salt spray corrosion. Simultaneously, the ultrasonic sensors operating in the 200kHz to 2MHz frequency band underwent high-frequency signal attenuation optimization to enhance their ability to capture micro-amplitude stress waves generated by micron-level slip. In the practical application of the flange connection bolts of a 2.5MW wind turbine tower in an onshore wind farm, M36 bolts made of 40Cr material were selected for monitoring. The encapsulated vibration sensor and ultrasonic sensor were integrated and installed in the middle area of ​​the bolt shank using a magnetic anti-loosening fixing bracket. This location is a stress concentration area where cracks are prone to initiation and slippage is likely to occur. Anti-loosening nuts were added to the magnetic fixing bracket to prevent the sensor from falling off due to strong vibration. At the same time, a micro-displacement auxiliary sensor was attached to the contact end face between the bolt and the flange to capture the relative slippage between the bolt and the connecting parts. Temperature and humidity sensors and salt spray concentration sensors were deployed 10cm around the sensors as environmental sensors. All sensors were connected to a multi-channel high-speed data acquisition terminal through a wind farm-specific anti-interference shielded cable for sensor calibration. After calibration, the signal acquisition error of the sensors was controlled within 2% in an environment with a vibration frequency of 10Hz and a salt spray concentration of 50mg / m³, ensuring that the sensors work stably in the rotational vibration range of 3Hz to 15Hz and under wind loads of level 0 to 12.

[0033] After the sensors are installed and calibrated, the wind power-dedicated multi-channel high-speed data acquisition terminal is started and the sampling frequency adaptive rule is set. When the real-time wind speed is less than or equal to 10 m / s, the sampling frequency of the vibration sensor is set to 5 kHz and the sampling frequency of the ultrasonic sensor is set to 1 MHz. When the real-time wind speed is greater than 10 m / s, the sampling frequency of the vibration sensor is adaptively increased to 10 kHz and the sampling frequency of the ultrasonic sensor is adaptively increased to 2 MHz. The sampling frequency of the micro-displacement sensor is fixed at 1 kHz and the sampling interval of the environmental sensor is fixed at 5 seconds. This adaptive sampling strategy can ensure data storage efficiency under low wind speed conditions while ensuring accurate capture of transient characteristics under high wind speed and strong wind load impact conditions.

[0034] A high-precision synchronous triggering module built into the multi-channel high-speed data acquisition terminal enables precise synchronous acquisition of signals from four channels: vibration sensor, ultrasonic sensor, micro-displacement sensor, and environmental sensor. The synchronization error is controlled within 1 microsecond. During the acquisition process, a power frequency notch filter is activated to remove 60Hz power frequency interference, and an adaptive threshold filtering algorithm is used to remove sudden noise caused by instantaneous strong wind loads, ensuring the validity of the original vibration signal, original ultrasonic signal, original displacement signal, and environmental parameter data. After the above implementation process, under the operating conditions of a wind turbine with a real-time speed of 15 r / min and a real-time wind speed of 8 m / s, the effective frequency band of the vibration signal of the monitored bolt was successfully acquired, covering 0 to 5 kHz; the effective frequency band of the ultrasonic signal covered 200 kHz to 2 MHz; and the displacement signal resolution reached the 0.1 micrometer level. At the same time, environmental parameters of 25℃ temperature, 60%RH humidity, and 30 mg / m³ salt spray concentration were acquired, providing a high-quality multimodal raw data foundation for subsequent dynamic anti-interference compensation and engineering feature extraction.

[0035] S102: Obtain the real-time operating condition parameters of the wind turbine, and perform compensation processing on the vibration signal, ultrasonic signal and displacement signal based on the real-time operating condition parameters to obtain the compensated vibration signal, compensated ultrasonic signal and compensated displacement signal.

[0036] In S102 above, to address the issue that the high-frequency background vibration of rotating machinery strongly interferes with monitoring signals under the combined dynamic conditions of high-frequency vibration of rotating machinery and alternating impact of strong wind loads, making it difficult for conventional monitoring technologies to distinguish between background vibration and bolt fault characteristic signals, and considering the sensor sensitivity drift and signal transmission attenuation caused by alternating outdoor temperature and humidity and salt spray corrosion, it is necessary to perform dynamic condition-specific anti-interference compensation processing on the collected vibration signals, ultrasonic signals, and displacement signals. This is because single vibration monitoring is easily overwhelmed by the strong background vibration of rotating machinery, making it impossible to identify the early characteristics of microcrack initiation, and environmental factors can cause signal distortion. Existing monitoring algorithms have not been optimized for the strong interference and multi-coupling characteristics of wind turbine dynamic conditions, resulting in poor engineering practicality. Therefore, by establishing wind load-speed-vibration dynamic anti-interference compensation information and adopting an engineered multi-factor step-by-step compensation algorithm, the superimposed interference under dynamic conditions is completely eliminated, achieving effective signal purification.

[0037] In addition, real-time operating parameters are first obtained through the SCADA system or dedicated data acquisition interface of the wind turbine. These real-time operating parameters include the wind turbine's real-time wind speed, real-time rotational speed, real-time load data, as well as temperature, humidity, and salt spray concentration data collected by environmental parameter sensors. In the actual application of a 2.5MW wind turbine in an onshore wind farm, the specific real-time operating parameters obtained are: real-time wind speed 8m / s, real-time rotational speed 15r / min, real-time load 800kN, temperature 25℃, humidity 60%RH, and salt spray concentration 30mg / m³.

[0038] Compensation processing is performed on vibration, ultrasonic, and displacement signals based on real-time operating parameters to obtain compensated vibration, ultrasonic, and displacement signals. Specifically, this includes: obtaining real-time rotational speed and wind speed from the real-time operating parameters; determining the fundamental and harmonic frequencies of the rotating machinery based on the real-time rotational speed; determining the frequency band of wind load impact based on the real-time wind speed; locating the amplitude components coinciding with the fundamental and harmonic frequencies, as well as the amplitude components within the wind load impact frequency band, in the frequency domain spectrum of the vibration signal to obtain the background vibration interference component; subtracting the background vibration interference component from the vibration signal to obtain the denoised vibration signal; acquiring real-time temperature data and real-time salt spray concentration data around the bolt to be monitored; calculating the temperature difference between the real-time temperature data and the preset calibration temperature; and multiplying the temperature difference by a temperature sensitivity bias. The sensitivity drift coefficient is used to obtain the sensitivity drift amount. Based on the sensitivity drift amount, the amplitudes of the denoised vibration signal, ultrasonic signal, and displacement signal are reversed. The gain of the ultrasonic signal is compensated based on the signal transmission attenuation rate corresponding to the real-time salt spray concentration data, resulting in the environmentally corrected vibration signal, environmentally corrected ultrasonic signal, and environmentally corrected displacement signal. Real-time load data is obtained from the real-time operating condition parameters. The nominal stress is calculated based on the real-time load data and the cross-sectional area of ​​the bolt to be monitored. The dynamic stress concentration factor is determined based on the nominal stress. The distortion correction factor is constructed based on the dynamic stress concentration factor. The environmentally corrected vibration signal, environmentally corrected ultrasonic signal, and environmentally corrected displacement signal are divided by the distortion correction factor to obtain the compensated vibration signal, compensated ultrasonic signal, and compensated displacement signal.

[0039] Specifically, real-time rotational speed and wind speed data are extracted from the real-time operating parameters obtained from the wind turbine's SCADA system. In a practical application of a 2.5MW wind turbine in an onshore wind farm, the obtained real-time rotational speed is 15 r / min (15 revolutions per minute), and the obtained real-time wind speed is 8 m / s. Based on the real-time rotational speed of 15 r / min, the fundamental frequency of the rotating machinery is determined using a frequency conversion calculation formula. Specifically, the rotational frequency is 0.25 Hz, obtained by dividing the real-time rotational speed by 60 seconds. Then, the harmonic frequencies are calculated based on the harmonic characteristics of the rotating machinery's vibration. The fundamental frequency of 0.25 Hz is multiplied by integer multiples of 2, 3, 4, and 5 to obtain a series of harmonic frequencies, such as 0.5 Hz (first harmonic), 0.75 Hz (second harmonic), 1.0 Hz (third harmonic), and 1.25 Hz (fourth harmonic). These fundamental and harmonic frequencies correspond to the periodic vibration characteristics generated by the rotation of the machinery itself, rather than the characteristics of bolt failure. Based on a real-time wind speed of 8 m / s, the frequency band range of wind load impact is determined by empirical formulas or lookup tables for wind load excitation frequency. According to the aerodynamic theory of wind turbines, the wind load excitation corresponding to a wind speed of 8 m / s is mainly concentrated in the low frequency band range of 0.5 Hz to 2 Hz. The vibration energy in this frequency band range mainly comes from the instantaneous impact of strong wind load and turbulent pulsation rather than the fault characteristics of the bolt itself.

[0040] After determining the fundamental frequency, harmonic frequencies, and wind load impact frequency band, the vibration signal needs to be processed by Fast Fourier Transform (FFT) to convert the time-domain vibration signal to the frequency domain spectrum. The frequency domain amplitude spectrum of the vibration signal is obtained through FFT, with the horizontal axis representing frequency and the vertical axis representing the vibration amplitude at the corresponding frequency. In the frequency domain spectrum of the vibration signal, amplitude components coinciding at the fundamental frequency of 0.25Hz and various harmonic frequencies of 0.5Hz, 0.75Hz, 1.0Hz, and 1.25Hz are identified. These amplitude components exhibit obvious narrow-band spikes in the frequency domain spectrum. By setting a frequency tolerance range of ±0.05Hz, the peak amplitudes and their corresponding phase information are extracted near the fundamental and harmonic frequencies to obtain the fundamental frequency interference component of the rotating machinery. Simultaneously, all amplitude components within the wind load impact frequency band (0.5Hz to 2Hz) are located in the frequency domain spectrum. These amplitude components exhibit a wideband energy distribution in the frequency domain spectrum. The total energy of the wind load impact interference component is obtained by integrating the amplitudes at all frequency points within the 0.5Hz to 2Hz band. The extracted rotating machinery fundamental frequency interference component is then superimposed and fused with the wind load impact interference component to obtain the complete background vibration interference component. This background vibration interference component is represented in the frequency domain as a set of complex vectors containing specific frequency amplitude and phase information.

[0041] To remove background vibration interference components from the vibration signal, an inverse fast Fourier transform (IFFT) is first performed on the background vibration interference components to convert them from the frequency domain back to the time domain, obtaining the time-domain background vibration interference signal. Then, in the time domain, the background vibration interference signal is subtracted from the original vibration signal point by point, completing the time-domain cancellation of background vibration interference and obtaining the denoised vibration signal. In the actual processing, the amplitude of the original vibration signal is 0.15V at the fundamental frequency of 0.25Hz, 0.10V at the first harmonic of 0.5Hz, and the average amplitude is 0.12V in the wind load impact band from 0.5Hz to 2Hz. After background vibration interference compensation processing, the amplitude of the denoised vibration signal in the above frequency band is reduced to 0.015V, meaning the interference component is reduced by 90%. Meanwhile, the amplitude of the denoised vibration signal remains unchanged in the bolt fault characteristic frequency band from 500Hz to 2000Hz, achieving accurate removal of background vibration interference without damaging the fault characteristic signal of the bolt itself.

[0042] After compensating for background vibration interference, real-time temperature and salt spray concentration data around the monitored bolt are obtained from data collected by environmental parameter sensors. In the above-mentioned onshore wind farm application example, the real-time temperature data is 25℃, and the real-time salt spray concentration data is 30mg / m³. Since the piezoelectric material or strain gauge material of the sensor is sensitive to temperature, temperature changes can cause the sensor's sensitivity to drift. To compensate for the impact of temperature changes on sensor sensitivity, the temperature difference between the real-time temperature data of 25℃ and the preset calibration temperature is first calculated. The preset calibration temperature is the ambient temperature during the sensor's factory calibration, typically set to 20℃. Therefore, the temperature difference is 25℃ minus 20℃, which equals 5℃. Multiplying the temperature difference of 5℃ by the temperature sensitivity offset coefficient yields the sensitivity drift. This temperature sensitivity offset coefficient characterizes the rate at which the sensor's sensitivity changes with temperature. According to the technical parameters provided by the sensor manufacturer or laboratory calibration results, the temperature sensitivity offset coefficient for vibration sensors is 0.002% / ℃, for ultrasonic sensors it is 0.0025% / ℃, and for displacement sensors it is 0.0018% / ℃.

[0043] For a vibration sensor, the sensitivity drift is 5℃ multiplied by 0.002% / ℃, which equals 0.01%. This means that the sensor's sensitivity at the current temperature of 25℃ is 0.01% higher than its sensitivity at the calibration temperature of 20℃. To inversely correct the amplitude of the denoised vibration signal and eliminate the influence of sensitivity drift, the amplitude of each sampling point of the denoised vibration signal is divided by a sensitivity drift factor. This sensitivity drift factor is 1 plus the sensitivity drift amount, i.e., 1 plus 0.01%, which equals 1.0001. By dividing the amplitude of the denoised vibration signal by 1.0001, the inverse correction is completed, resulting in the temperature-corrected vibration signal. Similarly, the amplitude of the ultrasonic signal is inversely corrected using a sensitivity drift factor of 1.0125, and the amplitude of the displacement signal is inversely corrected using a sensitivity drift factor of 1.0090, thus completing the temperature sensitivity drift compensation for the three signals. After temperature compensation, further compensation is needed to mitigate the attenuation effect of salt spray corrosion on signal transmission. Salt spray corrosion causes an oxide film to form at the connection interface between the sensor and the anti-interference shielded transmission cable, increasing contact resistance and thus causing signal transmission attenuation, especially for high-frequency ultrasonic signals. Based on the real-time salt spray concentration data of 30 mg / m³, a pre-established signal transmission attenuation rate data table was consulted. This table, obtained through accelerated salt spray corrosion experiments, records the transmission attenuation rates of various signals at different salt spray concentrations. The table shows that the ultrasonic signal transmission attenuation rate corresponding to a salt spray concentration of 30 mg / m³ is 5%, indicating that the ultrasonic signal loses 5% of its amplitude energy during transmission. To compensate for this transmission attenuation, gain compensation is performed on the temperature-corrected ultrasonic signal. This involves multiplying the amplitude of each sampling point of the ultrasonic signal by a gain compensation factor. This gain compensation factor is 1 divided by 1 minus the transmission attenuation rate, i.e., 1 divided by 1 minus 0.05 equals 1.0526. By multiplying the ultrasonic signal amplitude by 1.0526 to complete gain compensation, the salt spray-compensated ultrasonic signal is obtained.

[0044] Since the vibration and displacement signals have low frequencies and are less affected by salt spray corrosion, only temperature sensitivity drift compensation is performed on them, without salt spray attenuation compensation. After the above temperature and salt spray compensation processes, environmentally corrected vibration, ultrasonic, and displacement signals are obtained. In practical applications, the amplitude of the original ultrasonic signal is 0.76 mV·ms / s. After temperature compensation, the amplitude is corrected to 0.751 mV·ms / s. After salt spray gain compensation, the amplitude is restored to 0.790 mV·ms / s. The error with the standard value of 0.800 mV·ms / s in a non-corrosive environment is reduced from 5% to 1.25%, significantly improving the consistency and accuracy of the signal.

[0045] After environmental factor compensation is completed, real-time load data of the wind turbine is obtained from real-time operating condition parameters. This real-time load data includes the axial load and bending moment load of the rotor on the tower. In the above application example, the real-time load data is the axial load of 800kN at the tower flange connection. Because the load on the wind turbine changes continuously during operation, the stress distribution of the bolts differs under different load levels. At higher loads, stress concentration is more pronounced in the middle region of the bolt shank and at the root of the thread. This stress concentration leads to increased amplitude of vibration and ultrasonic signals in these areas. This increased amplitude is not caused by bolt failure but is a normal phenomenon resulting from load changes. Without compensation, this can lead to misdiagnosis of faults.

[0046] To compensate for signal distortion caused by uneven load stress, the nominal stress is first calculated based on the real-time load data of 800kN and the cross-sectional area of ​​the bolt to be monitored. The bolt to be monitored is an M36 bolt with a thread minor diameter of 31.67mm. The stress cross-sectional area of ​​the bolt is π multiplied by the square of the thread minor diameter and divided by 4, which is 3.14159 multiplied by the square of 31.67mm and divided by 4, equaling 787.7mm². The nominal stress is the real-time load of 800kN divided by the cross-sectional area of ​​787.7mm², which equals 1015MPa. The dynamic stress concentration factor is determined based on the nominal stress of 1015 MPa. The dynamic stress concentration factor characterizes the ratio of the actual stress to the nominal stress in the stress concentration area of ​​the bolt under the nominal stress. According to the finite element analysis results or by consulting the bolt stress concentration factor handbook, for an M36 40Cr bolt under the nominal stress of 1015 MPa, the dynamic stress concentration factor in the middle area of ​​the bolt shank is 1.2, and the dynamic stress concentration factor at the root of the thread is 1.35. Since the vibration sensor and the ultrasonic sensor are installed in the middle area of ​​the bolt shank, a dynamic stress concentration factor of 1.2 is used for compensation.

[0047] A distortion correction factor is constructed based on a dynamic stress concentration factor of 1.2. This factor is directly equal to the dynamic stress concentration factor, which represents the amplification factor of the signal amplitude caused by stress concentration. To eliminate this amplification effect, the signal amplitude needs to be normalized by dividing it by the distortion correction factor. The amplitude of each sampling point of the environmentally corrected vibration signal, ultrasonic signal, and displacement signal is divided by the distortion correction factor of 1.2 to complete the distortion correction for variable load stress, resulting in the compensated vibration signal, ultrasonic signal, and displacement signal. In actual processing, the root mean square (RMS) value of the environmentally corrected vibration signal in the 500Hz to 2000Hz frequency band was 0.30V. This amplitude is 20% larger than the standard value of 0.25V under no-load conditions. The increase is due to signal amplification caused by the stress concentration factor of 1.2 resulting from a load of 800kN. After division by a distortion correction factor of 1.2, the RMS value of the compensated vibration signal is restored to 0.30V divided by 1.2, which equals 0.25V, consistent with the standard value under no-load conditions, thus achieving accurate compensation for variable load stress distortion. Similarly, the energy rate of the environmentally corrected ultrasonic signal was 0.96mV·ms / s. After distortion correction, the energy rate of the compensated ultrasonic signal is 0.80mV·ms / s. The slip of the environmentally corrected displacement signal was 2.4 micrometers. After distortion correction, the slip of the compensated displacement signal is 2.0 micrometers. The compensated signal truly reflects the dynamic slip and microcrack initiation characteristics of the bolt under standardized stress conditions.

[0048] like Figure 2 As shown, after the above-mentioned preliminary noise reduction, background vibration interference compensation, environmental factor compensation, and variable load stress compensation, the original vibration signal, ultrasonic signal, and displacement signal are processed under the composite dynamic working conditions of real-time rotation speed of 15 r / min, real-time wind speed of 8 m / s, real-time temperature of 25℃, real-time humidity of 60%RH, real-time salt spray concentration of 30 mg / m³, and real-time load of 800 kN. The resulting compensated vibration signal, compensated ultrasonic signal, and compensated displacement signal can accurately characterize the dynamic slippage and microcrack initiation state of the bolt under monitoring.

[0049] S103: Analyze the compensated vibration signal and the compensated ultrasonic signal, and extract the vibration feature value and ultrasonic feature value respectively.

[0050] In the above S103, in order to achieve accurate identification of the dynamic slippage and microcrack initiation state of wind turbine bolts, it is necessary to extract characteristic parameters that can accurately characterize the bolt preload attenuation, dynamic slippage development, and microcrack initiation and propagation from the compensated vibration signal and compensated ultrasonic signal obtained after multi-factor step-by-step compensation processing. This is because although the compensated vibration signal has eliminated background vibration interference and corrected the influence of environmental factors and variable load stress, it is still essentially a complex signal containing rich frequency domain information and time domain statistical characteristics. The vibration energy distribution of different frequency bands reflects the dynamic response characteristics of the bolt under different working conditions, while the compensated ultrasonic signal carries the micro-amplitude stress wave information generated by the dynamic slippage of the bolt and the elastic stress wave energy released during the microcrack initiation and propagation process. If these signals are directly used for fault identification without engineering feature extraction processing, it will lead to information redundancy.

[0051] Furthermore, the compensated vibration signal and compensated ultrasonic signal are analyzed to extract vibration feature values ​​and ultrasonic feature values, respectively. Specifically, this includes: performing wavelet packet decomposition on the compensated vibration signal to obtain sub-signals of different frequency bands; selecting sub-signals within a preset frequency range from the sub-signals of different frequency bands as feature frequency band signals; performing time-domain statistical analysis on the feature frequency band signals to calculate the root mean square value, peak factor, and impulse factor of the feature frequency band signals; combining the root mean square value, peak factor, and impulse factor as vibration feature values; performing short-time energy analysis on the compensated ultrasonic signal to calculate the short-time energy change rate of the ultrasonic signal within a time window function to obtain the acoustic-ultrasonic energy rate; counting the number of times the ultrasonic signal exceeds the trigger threshold; dividing the number of times by the unit time to obtain the count rate; calculating the time interval from the first time the ultrasonic signal exceeds the trigger threshold to the signal peak to obtain the rise time; and combining the acoustic-ultrasonic energy rate, count rate, and rise time as ultrasonic feature values.

[0052] Specifically, the compensated vibration signal is processed by wavelet packet decomposition. The reason for using wavelet packet decomposition instead of the traditional Fourier transform or wavelet transform is that wavelet packet decomposition can provide good resolution in both the time and frequency domains. It has significant analytical advantages for non-stationary signals such as wind turbine bolt vibration signals, which contain both low-frequency preload attenuation characteristics and mid-to-high-frequency dynamic slip and microcrack initiation characteristics. Moreover, wavelet packet decomposition can uniformly subdivide the frequency band of the signal, avoiding the defect of insufficient resolution in the high-frequency band of wavelet transform.

[0053] In the practical application of the tower flange connection bolts of a 2.5MW wind turbine in an onshore wind farm, the Daubechies wavelet packet db4 was selected as the wavelet basis function. This wavelet basis function has good time-frequency localization characteristics and orthogonality, which can effectively avoid frequency aliasing. The wavelet packet decomposition level was set to 5 levels. Through 5 levels of wavelet packet decomposition, the compensated vibration signal was decomposed into 32 sub-signals with different frequency bands. The bandwidth of each sub-signal is the sampling frequency divided by 64. Since the vibration sampling frequency is determined to be 5kHz based on the real-time wind speed of 8m / s, the bandwidth of each sub-signal after 5-layer wavelet packet decomposition is approximately 78Hz (5000Hz divided by 64). The first layer decomposition yields two sub-signals corresponding to the 0-2500Hz and 2500Hz-5000Hz frequency bands, respectively. The second layer decomposition yields four sub-signals corresponding to the 0-1250Hz, 1250Hz-2500Hz, 2500Hz-3750Hz, and 3750Hz-5000Hz frequency bands, respectively. This process continues until the fifth layer decomposition yields 32 sub-signals with a bandwidth of approximately 78Hz, covering the complete 0-5000Hz frequency band. In the wavelet packet decomposition process, the Mallat fast algorithm is used to implement the recursive calculation of wavelet packet decomposition. The signal is filtered and extracted layer by layer by cascading low-pass and high-pass filters. In the first layer of decomposition, the compensated vibration signal is passed through the low-pass and high-pass filters simultaneously to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The low-frequency approximation coefficients and high-frequency detail coefficients are each decimated by 2 to obtain two sub-signals. In the second layer of decomposition, the two sub-signals obtained in the first layer are passed through the low-pass and high-pass filters again and decimated by 2 to obtain four sub-signals. This process is recursively calculated up to the fifth layer to complete the complete wavelet packet decomposition process.

[0054] After wavelet packet decomposition, the frequency band most relevant to bolt preload attenuation and dynamic slip characteristics needs to be selected from the 32 sub-signals of different frequency bands as the characteristic frequency band signal. According to the spectrum analysis results of a large amount of wind turbine bolt fault test data and field measured data, the vibration characteristics caused by bolt preload attenuation are mainly concentrated in the mid-frequency band of 500Hz to 2000Hz. This frequency band corresponds to the frictional vibration generated when dynamic slip occurs at the contact interface between the bolt and the flange, as well as the elastic vibration generated in the stress concentration area of ​​the bolt shank. The low-frequency vibration below 500Hz mainly comes from the overall tower swaying and the fundamental frequency vibration of the impeller rotation, which is not a fault characteristic of the bolt itself. The high-frequency vibration above 2000Hz mainly comes from the gear meshing and bearing rolling of the transmission system, which is not a fault characteristic of the bolt connection. Therefore, the preset frequency range is set to 500Hz to 2000Hz. Among the 32 sub-signals, based on the center frequency and bandwidth of each sub-signal, all sub-signals with center frequencies falling within the range of 500Hz to 2000Hz are selected as characteristic frequency band signals, i.e., 20 sub-signals. These sub-signals are then reconstructed in the time domain to obtain the complete characteristic frequency band signal. This characteristic frequency band signal retains all time-domain and frequency-domain characteristic information of the compensated vibration signal within the 500Hz to 2000Hz frequency band while filtering out irrelevant frequency band information below 500Hz and above 2000Hz.

[0055] After extracting the characteristic frequency band signals, time-domain statistical analysis is required to calculate time-domain statistical characteristic parameters that can quantitatively characterize the bolt preload attenuation and dynamic slippage. The reason for choosing time-domain statistical analysis instead of frequency-domain analysis is that time-domain statistical characteristic parameters, such as root mean square value, peak factor, and impulse factor, can intuitively reflect the signal's energy level, waveform steepness, and impact characteristics. These characteristics have a clear physical correspondence with the increase in vibration energy caused by bolt preload attenuation, the abrupt change in vibration waveform caused by dynamic slippage, and the impact pulse caused by microcrack initiation. Moreover, the computational complexity of time-domain statistical features is low, making them very suitable for real-time processing at edge computing terminals to meet the engineering practicality requirements of online monitoring of wind turbine units. First, calculate the root mean square (RMS) value of the characteristic frequency band signal. The RMS value is a measure of the effective value of the signal and reflects the average energy level of the signal. For bolt vibration signals, an increase in the RMS value indicates that the bolt preload has decreased, leading to a decrease in the stiffness of the bolt-flange connection interface and an increase in vibration energy. The formula for calculating the RMS value is to sum the squares of the amplitude of each sampling point of the characteristic frequency band signal, divide by the total number of sampling points, and then take the square root. Specifically, in the above application example, the amplitudes of the 50,000 sampling points of the characteristic frequency band signal are denoted as x1, x2, x3 up to x50,000. Squaring these 50,000 amplitudes respectively yields x1², x2², x3² up to x50,000². Adding these 50,000 squared values ​​gives the sum of squares S. Dividing the sum of squares S by the total number of sampling points 50,000 gives the mean square S. Taking the square root of the mean square gives the root mean square (RMS) value, which is 0.25V.

[0056] Next, the peak factor of the characteristic frequency band signal is calculated. The peak factor is the ratio of the signal peak value to the root mean square value, which reflects the steepness and impact of the signal waveform. For bolt vibration signals, an increase in the peak factor indicates that a sharp impact vibration waveform is generated during the dynamic slippage of the bolt. The formula for calculating the peak factor is to divide the peak value of the characteristic frequency band signal by the root mean square value. First, it is necessary to determine the peak value of the characteristic frequency band signal, which is the maximum value among the amplitudes of 50,000 sampling points. By traversing the amplitudes of 50,000 sampling points, the maximum amplitude value is found and denoted as Peak. In the above application example, the peak value Peak of the characteristic frequency band signal is 0.80V. Dividing the peak value of 0.80V by the root mean square value of 0.25V, we get the peak factor CF, which is equal to 0.80V divided by 0.25V, which equals 3.2. Then, the impulse factor of the characteristic frequency band signal is calculated. The impulse factor, which is the ratio of the signal peak value to the absolute value of the average amplitude, more sensitively reflects the impact pulse characteristics of the signal. For bolt vibration signals, an increase in the impulse factor indicates that the elastic stress wave released during the initiation of bolt microcracks triggers high-intensity impulse vibration. The formula for calculating the impulse factor is to divide the peak value of the characteristic frequency band signal by the absolute value of the average amplitude, that is, to calculate the absolute value of the average amplitude of the characteristic frequency band signal, which is the sum of the absolute values ​​of the amplitudes of 50,000 sampling points, divided by the total number of sampling points. The amplitudes of the 50,000 sampling points are then multiplied by x1, x2, x3 up to x50,000 respectively. The absolute values ​​are taken to obtain |x1|, |x2|, |x3| up to |x50000|. These 50000 absolute values ​​are added together to obtain the absolute sum A, which is equal to |x1| plus |x2| plus |x3| up to |x50000|. The absolute sum A is divided by the total number of sampling points, 50000, to obtain the average amplitude absolute value AVG, which is equal to A divided by 50000. In the above application example, the average amplitude absolute value AVG of the characteristic frequency band signal is 0.286V. The peak value 0.80V is divided by the average amplitude absolute value 0.286V to obtain the pulse factor IF, which is approximately 2.8 (0.80V divided by 0.286V). The root mean square value of 0.25V, peak factor of 3.2, and impulse factor of 2.8 obtained from the above calculations are combined as vibration characteristic values. This vibration characteristic value is a three-dimensional feature vector denoted as [0.25V, 3.2, 2.8]. This three-dimensional feature vector comprehensively characterizes the energy level, waveform steepness, and impact pulse characteristics of the signal in the characteristic frequency band, and can quantitatively reflect the degree of development of bolt preload attenuation and dynamic slippage.

[0057] After extracting the vibration characteristic values, short-time energy analysis is required on the compensated ultrasonic signal to calculate the acoustic-ultrasonic energy rate, which characterizes the energy release rate of the micro-amplitude stress waves generated by the dynamic slippage of the bolt. Short-time energy analysis is used instead of overall energy statistics because bolt dynamic slippage is a dynamic process that evolves with changes in wind load and rotational speed. The energy release of frictional stress waves and elastic stress waves generated during slippage is time-varying and sudden. Short-time energy analysis can capture the energy change trend of the ultrasonic signal within a short time window, thus reflecting the real-time development state of dynamic slippage. In contrast, overall energy statistics only provide the average energy level of the signal and cannot reflect the time-varying characteristics of energy release. The core of short-time energy analysis is to divide the ultrasonic signal into multiple short time windows, calculate the signal energy within each time window, and then statistically analyze the energy change rate between adjacent time windows. This energy change rate is the acoustic-ultrasonic energy rate. In the above application example, the sampling frequency of the compensated ultrasound signal is 1MHz and the sampling duration is 10 seconds, so the total number of sampling points is 10 million. The length of the short time window is set to 1 millisecond, that is, each time window contains 1000 sampling points, dividing the 10-second ultrasound signal into 10000 short time windows. For the i-th time window, the starting sampling point is the 1000(i-1)+1 sampling point, and the ending sampling point is the 1000i sampling point. The amplitude of the 1000 sampling points in the i-th time window is denoted as yi. The formula for calculating the energy Ei of the i-th time window is to sum the squares of the amplitudes of all sampling points in the time window. After calculating the energy of all time windows, it is necessary to calculate the energy change rate between adjacent time windows. For the i-th time window and the i+1-th time window, the energy change rate ΔEi is calculated by subtracting the energy Ei of the i-th time window from the energy Ei+1 of the i+1-th time window and then dividing by the time window length of 1 millisecond. The unit of this energy change rate is energy units divided by time units, i.e., mV²·ms / s, or simplified to mV·ms / s. By traversing 9999 adjacent time window pairs, 9999 energy change rates ΔE1, ΔE2, ΔE3 up to ΔE9999 are calculated. The average of these 9999 energy change rates is taken as the acoustic-ultrasound energy rate. The formula for calculating the acoustic-ultrasound energy rate ER is the sum of the 9999 energy change rates and divided by 9999, i.e., ER equals the sum of ΔE1 plus ΔE2 plus ΔE3 up to ΔE9999 divided by 9999. In the above application example, the energy E1 of the compensated ultrasound signal in the first time window is 0.64 mV²·ms, the energy E2 in the second time window is 0.66 mV²·ms, the energy change rate ΔE1 between the first and second time windows is 0.02 mV·ms / s, and so on, the energy change rate of all adjacent time windows is calculated and averaged to obtain the ultrasound energy rate ER of 0.80 mV·ms / s.

[0058] After extracting the ultrasonic energy rate, the compensated ultrasonic signal needs to be subjected to ring counting statistics to calculate the count rate, which characterizes the frequency of stress wave events during bolt dynamic slippage and microcrack initiation. Ring counting statistics are used instead of simple amplitude statistics because the stress waves released by bolt dynamic slippage and microcrack initiation exhibit ringing characteristics; that is, after stress wave excitation, the signal will produce multiple decaying oscillations within a short period. Ring counting can capture each stress wave event without being affected by the magnitude of the stress wave amplitude, offering better robustness and sensitivity compared to amplitude statistics. The core of ring counting statistics is to set a trigger threshold. When the amplitude of the ultrasonic signal first exceeds this threshold, it is recorded as a stress wave event, i.e., one ring count. Then, when the signal amplitude falls back below the trigger threshold and exceeds it again, it is recorded as the next ring count. The count rate is obtained by counting the total number of ring counts per unit time. In the above application example, the count rate CR is 5 times / s. In actual processing, to avoid duplicate counting caused by a single stress wave event repeatedly crossing the trigger threshold, a refractory period is set after a rising edge trigger is detected. This means that all rising edge triggers are ignored for a period of time before the end of the stress wave event. The length of this refractory period is determined according to the typical duration of the stress wave event, and is usually set to 0.1 milliseconds, or 100 sampling points. By setting the refractory period, it can be ensured that each ring count corresponds to a real independent stress wave event, avoiding the falsely high count rate caused by duplicate counting.

[0059] After the count rate is extracted, the rise time of the compensated ultrasonic signal needs to be statistically analyzed to calculate the rise time that can characterize the stress wave excitation velocity during bolt dynamic slippage and microcrack initiation. The reason for statistically analyzing the rise time is that different types of stress wave events have different rise time characteristics. Friction stress waves generated by bolt dynamic slippage usually have a slow rise process and a long rise time, while elastic stress waves generated by bolt microcrack initiation usually have a sudden and rapid rise process and a short rise time. Statistical analysis of the rise time can help distinguish between the two different failure modes of dynamic slippage and microcrack initiation. Rise time is defined as the time interval from when the ultrasonic signal first exceeds the trigger threshold to when it reaches its peak value. For each stress wave event corresponding to each ring count, the rise time of the stress wave event is first determined, i.e., the sampling point when the amplitude first exceeds the trigger threshold of 0.25mV, denoted as t1. Then, the sampling points are traversed backward to find the peak point of the stress wave event, i.e., the sampling point corresponding to the maximum amplitude of all sampling points during the stress wave event, denoted as t2. The rise time RT of the stress wave event is obtained by subtracting the trigger time t1 from the peak time t2, which is equal to t2 minus t1. The rise time RT is 0.10 milliseconds, or 0.1ms. The ultrasonic energy rate of 0.80 mV·ms / s, count rate of 5 times / s, and rise time of 0.1 ms obtained from the above calculations are combined as ultrasonic characteristic values. These ultrasonic characteristic values ​​are a three-dimensional feature vector denoted as [0.80 mV·ms / s, 5 times / s, 0.1 ms]. This three-dimensional feature vector comprehensively characterizes the energy release rate of the compensated ultrasonic signal, the frequency of stress wave events, and the stress wave excitation velocity, and can quantitatively reflect the state characteristics of bolt dynamic slippage and microcrack initiation.

[0060] S104: Extract the real-time displacement value from the compensated displacement signal, and normalize the vibration characteristic value, ultrasonic characteristic value and real-time displacement value to obtain the coupled feature vector.

[0061] In S104 above, real-time displacement values ​​are extracted from the compensated displacement signal, and the vibration feature values, ultrasonic feature values, and real-time displacement values ​​are normalized to obtain a coupled feature vector. Specifically, this includes: performing time-domain analysis on the compensated displacement signal, calculating the average amplitude of the displacement signal within a preset sampling period, and using the average amplitude as the real-time displacement value; constructing a feature dataset from the vibration feature values, ultrasonic feature values, and real-time displacement values; processing each feature data in the feature dataset using a max-min normalization algorithm to obtain multiple normalized feature data; and sorting the multiple normalized feature data according to a preset order to obtain a multi-dimensional coupled feature vector.

[0062] Specifically, time-domain analysis is performed on the compensated displacement signal to extract the real-time displacement value that can accurately characterize the current slip state of the bolt. The reason for using time-domain analysis instead of frequency-domain analysis is that the dynamic slip of the bolt is a quasi-static process that gradually accumulates with changes in wind load and rotational speed. The change in slip is mainly reflected in the change of signal amplitude rather than the change of frequency characteristics. Time-domain analysis can directly reflect the amplitude statistical characteristics of the displacement signal and thus accurately extract the real-time slip, while frequency-domain analysis can reveal the frequency components of the displacement signal but cannot directly give the numerical value of the slip. In the above application example, the compensated displacement signal is obtained by the micro-displacement auxiliary sensing module by continuously sampling at a sampling frequency of 1kHz for 10 seconds. Therefore, the compensated displacement signal contains 10,000 sampling points, denoted as d1, d2, d3 up to d10,000. Since the micro-displacement auxiliary sensing module works based on the principle of capacitive or inductive displacement sensing, the output signal will be affected by sensor circuit noise and mechanical vibration on the microsecond time scale, resulting in high-frequency jitter. If the amplitude of a certain sampling point is directly taken as the real-time displacement value, the measurement result will be unstable and easily affected by instantaneous noise. Therefore, it is necessary to average the amplitude of multiple sampling points within a preset sampling period to filter out high-frequency jitter noise and obtain a stable real-time displacement value. The setting of the preset sampling period needs to comprehensively consider the signal-to-noise ratio improvement requirements and real-time requirements of the displacement signal. A preset sampling period that is too short will result in too few sampling points in the average processing, failing to effectively filter out high-frequency jitter noise. A preset sampling period that is too long will result in an excessively wide time window for the average processing, failing to reflect the real-time changes in bolt slippage status in a timely manner. Based on the statistical analysis results of a large amount of wind turbine bolt dynamic slippage monitoring test data and field measurement data, the duration of a single slippage event of a bolt under strong wind load impact is typically between 0.5 seconds and 2 seconds. Therefore, setting the preset sampling period to 1 second can effectively filter out high-frequency jitter noise while ensuring real-time performance. In the above application example, a preset sampling period of 1 second corresponds to 1000 sampling points. After determining the preset sampling period, the average amplitude of the compensated displacement signal within the preset sampling period needs to be calculated as the real-time displacement value. In practical applications, the average amplitude of the most recent preset sampling period is usually taken as the real-time displacement value at the current moment to reflect the latest slippage state of the bolt. In the above application example, the average amplitude D10 of the 10th preset sampling period is taken as the real-time displacement value, such as 2 micrometers, or 2μm.It is important to note that during the actual operation of wind turbines, displacement signals may exhibit low-frequency trend drift. This can be caused by factors such as thermal expansion of bolt materials due to temperature changes or creep relaxation of bolts due to long-term loads. These low-frequency trend drifts can be superimposed on the actual dynamic slip signal, resulting in real-time displacement values ​​that are either too large or too small. To eliminate the influence of low-frequency trend drift, a high-pass filter can be applied to the compensated displacement signal before calculating the average amplitude to remove low-frequency components below 0.1Hz. However, in the application example mentioned above, since the dynamic anti-interference compensation already includes environmental factor compensation to correct for temperature drift, the low-frequency trend drift in the compensated displacement signal has been effectively suppressed and no further high-pass filtering is required.

[0063] After extracting the real-time displacement values, the previously extracted vibration feature values, ultrasonic feature values, and real-time displacement values ​​need to be constructed into a unified feature dataset. The reason for constructing a feature dataset instead of directly combining them into a feature vector is that all the feature datasets to be processed need to be organized before normalization to allow for unified parameter configuration and batch processing of the normalization algorithm. In the above application example, the vibration feature value is a three-dimensional feature vector containing three feature data: root mean square value 0.25V, peak factor 3.2, and impulse factor 2.8. The ultrasonic feature value is a three-dimensional feature vector containing three feature data: acoustic-ultrasonic energy rate 0.80mV·ms / s, count rate 5 times / s, and rise time 0.1ms. The real-time displacement value is a scalar containing one feature data: displacement amount 2μm. These seven feature data are arranged in the order of extraction to construct a feature dataset denoted as {0.25V, 3.2, 2.8, 0.80mV·ms / s, 5 times / s, 0.1ms, 2μm}. This feature dataset contains seven elements corresponding to seven different physical measurements. Each feature data has a clear physical meaning and dimension, but these feature data have different dimensions and huge differences in numerical range. The root mean square value has the dimension of voltage V and the numerical range is usually between 0.1V and 1V. The peak factor and impulse factor are dimensionless numbers and the numerical range is usually between 2 and 5. The dimension of acoustic and ultrasonic energy rate is mV·ms / s and the numerical range is usually between 0.5mV·ms / s and 2mV·ms / s. The dimension of count rate is times / s and the numerical range is usually between 1 times / s and 10 times / s. The dimension of rise time is ms and the numerical range is usually between 0.05ms and 0.2ms. The dimension of real-time displacement value is μm and the numerical range is usually between 0.5μm and 10μm.

[0064] After constructing the feature dataset, the min-max normalization algorithm is used to normalize each feature data in the dataset to eliminate differences in dimensions and numerical ranges. The min-max normalization algorithm is chosen over Z-score normalization or other normalization methods because it can linearly map feature data of any numerical range to a unified interval of 0 to 1. The normalized feature data maintains the relative magnitudes of the original data and is unaffected by the distribution characteristics of the original data. Furthermore, the min-max normalization algorithm has low computational complexity, making it ideal for real-time processing on edge computing terminals to meet the practical engineering requirements of online monitoring of wind turbine units. The core idea of ​​the min-max normalization algorithm is to query a pre-established feature value calibration database for each feature data in the dataset to obtain the maximum and minimum values ​​of that feature data under all historical monitoring conditions. Then, the current feature data is subtracted from the historical minimum value, and then divided by the historical maximum value minus the historical minimum value. This yields the normalized feature data, whose numerical range is a dimensionless number between 0 and 1. The normalization formula is: the normalized feature data equals the current feature data minus the historical minimum value divided by the historical maximum value minus the historical minimum value. In the above application example, the eigenvalue calibration database was established before the system was put into operation through a large number of fault tests and field measurements on bolts of the same type of wind turbine. The database stores the historical maximum and minimum values ​​of each type of eigenvalue under different conditions such as no fault, dynamic slip, and microcrack initiation. For the root mean square value of 0.25V in the vibration eigenvalue, the eigenvalue calibration database is queried to obtain the historical maximum value Vmax of the eigenvalue of this type of bolt under all working conditions. Vmax is equal to 1.2V, which corresponds to the vibration energy level when the bolt has a serious microcrack propagation. Vmin is equal to 0.1V, which corresponds to the vibration energy level when the bolt is in a fault-free and well-tightened state. The normalized root mean square value RMS_norm is obtained, which is 0.15V divided by 1.1V, which is approximately equal to 0.136. The normalized value of 0.136 is in the range of 0 to 1, which means that the vibration energy level of the current bolt is about 13.6% between the historical minimum and historical maximum values.

[0065] The same method was used to perform max-min normalization on the other six feature data in the feature dataset. The normalized peak factor CF_norm is equal to 0.400, the normalized impulse factor IF_norm is equal to 0.320, the normalized acoustic-ultrasonic energy rate ER_norm is equal to 0.294, the normalized count rate CR_norm is equal to 0.444, the normalized rise time RT_norm is equal to 0.333, and the normalized real-time displacement value D_norm is equal to 0.158. Through the above-mentioned max-min normalization algorithm, the seven feature data in the feature dataset are normalized into seven normalized feature data, denoted as {0.136, 0.400, 0.320, 0.294, 0.444, 0.333, 0.158}. The dimensions of all normalized feature data are unified to dimensionless numbers and the numerical range is unified between 0 and 1, eliminating the differences in dimensions and numerical ranges between the original feature data. At the same time, the relative position information of each feature data with respect to historical data is preserved. The closer the normalized feature data is to 0, the closer the current bolt's feature is to a fault-free state. The closer the normalized feature data is to 1, the closer the current bolt's feature is to a severe fault state. It is important to note that the historical maximum and minimum values ​​in the feature value calibration database need to be dynamically updated based on the actual monitoring data of different wind fields and different bolt models. In the initial stage of system operation, the historical maximum and minimum values ​​are set based on laboratory fault test data. During system operation, as field monitoring data is continuously accumulated, the feature value calibration database will continuously supplement new data samples and recalculate the historical maximum and minimum values ​​to ensure that the reference range of normalization processing always matches the actual working conditions of the current wind field and bolt model, thereby improving the accuracy and reliability of the normalized feature data.

[0066] After normalizing all feature data in the feature dataset, the normalized feature data needs to be sorted according to a preset order to construct a multidimensional coupled feature vector with a unified format. The reason for sorting the normalized feature data instead of random combination is that the subsequent hierarchical recognition model based on the improved BP neural network (the preset fault recognition model) has already established the weight connection relationship between the input layer and the hidden layer according to a specific feature order during the training phase. If the feature order of the input feature vectors in the inference phase is inconsistent with that in the training phase, it will cause misalignment between the model's input data and the weight connection relationship, severely affecting the recognition accuracy. Therefore, it is essential to strictly sort the normalized feature data according to the preset order to ensure the consistency of the feature vector format. The preset order follows the principle of arranging physically similar feature data adjacently. Feature data from the same signal type are grouped and arranged so that the model can better learn the coupling relationship between features of different signal types. Simultaneously, feature data representing the same fault mode are arranged adjacently so that the model can more sensitively capture feature combinations of specific fault modes. In the example above, the preset order is defined as follows: first, arrange the three normalized feature data of vibration feature values, including the normalized root mean square value, the normalized peak factor, and the normalized impulse factor; then, arrange the three normalized feature data of ultrasonic feature values, including the normalized acoustic-ultrasonic energy rate, the normalized count rate, and the normalized rise time; and finally, arrange the normalized real-time displacement value. The preset order follows the grouping principle of vibration feature - ultrasonic feature - displacement feature, and also conforms to the progressive relationship of physical meaning of energy feature - waveform feature - time domain feature - displacement feature.

[0067] like Figure 3 As shown, after obtaining the compensated vibration signal, compensated ultrasonic signal, and compensated displacement signal, wavelet decomposition is performed on the compensated vibration signal to extract the root mean square value, peak factor, and impulse factor. Short-time energy analysis is performed on the compensated ultrasonic signal to extract the acoustic-ultrasonic energy rate, count rate, and rise time. Real-time slip is obtained from the compensated displacement signal. The obtained root mean square value, peak factor, impulse factor, acoustic-ultrasonic energy rate, count rate, rise time, and real-time slip are normalized to obtain the coupled feature vector. S105: Input the coupled feature vector into the preset fault identification model for processing to obtain the status information of the bolt to be monitored. The status information includes no fault information, dynamic slip information and crack information.

[0068] In S105 above, a pre-defined fault identification model that can accurately identify three states of bolts: no fault, dynamic slippage, and microcrack initiation needs to be constructed first. The reason for constructing this model in advance instead of directly using a general fault identification algorithm is that the fault characteristics of wind turbine bolts under dynamic operating conditions are significantly different from those of other mechanical equipment. Wind turbine bolts are subjected to the combined effects of alternating impacts from strong wind loads and high-frequency vibrations from rotating machinery over a long period of time. The evolutionary patterns of dynamic slippage and microcrack initiation have obvious characteristics of the wind power industry. General fault identification algorithms that are not optimized for these characteristics will result in low identification accuracy and high false alarm rate. Therefore, it is necessary to construct a dedicated pre-defined fault identification model based on historical monitoring data of wind turbine bolts.

[0069] Furthermore, before processing the coupled feature vector into the preset fault identification model to obtain the state information of the bolt to be monitored, the preset fault identification model needs to be constructed. The construction process specifically includes: acquiring historical monitoring data of bolts in different wind turbine locations under different operating conditions; performing feature extraction and normalization on the historical monitoring data to obtain training sample feature vectors, the dimension of which is consistent with the coupled feature vector; labeling the training sample feature vectors, including state category labels, slip value labels, and crack depth value labels; determining the dimension of the training sample feature vectors corresponding to the number of input layer nodes, determining the dimension of the labels corresponding to the number of output layer nodes, setting the number of hidden layer nodes, and randomly initializing the connection weights between each layer. The process involves: inputting the sample feature vector into the input layer, calculating the hidden layer output and the output layer output through the activation function to obtain the predicted state vector; calculating the mean squared error between the predicted state vector and the label, and determining whether the mean squared error is less than a preset convergence threshold; if the mean squared error is greater than or equal to the preset convergence threshold, calculating the error gradient between the output layer and the hidden layer; based on the error gradient, combined with a preset learning rate and a preset momentum factor, calculating the weight adjustment amount, and using the weight adjustment amount to update the connection weights and the threshold, wherein the weight adjustment amount includes the gradient influence of the current iteration and the weight change amount of the previous iteration; repeating the calculation to the weight update step until the mean squared error is less than the preset convergence threshold or the maximum number of iterations is reached to obtain the preset fault identification model.

[0070] Specifically, historical monitoring data of bolts from different parts of the wind turbine under various operating conditions were obtained as training samples. This historical monitoring data originated from long-term monitoring records of tower flange connection bolts, hub fixing bolts, and blade connection bolts under normal operating conditions, strong wind load impact conditions, and extreme temperature conditions, and also included laboratory simulated fault test data. For example, the historical monitoring data covered bolt monitoring records of 20 2.5MW wind turbines in an onshore wind farm over a two-year operating cycle, as well as laboratory data from preload relaxation tests and fatigue crack propagation tests on 40Cr M36 bolts. A total of 1500 sets of historical bolt monitoring data were obtained, including 800 sets of tower flange connection bolt data, 450 sets of hub fixing bolt data, and 250 sets of blade connection bolt data. Each set of data included vibration signals, ultrasonic signals, displacement signals, and environmental parameters at the time of acquisition. After acquiring historical monitoring data, feature extraction and normalization are performed to obtain training sample feature vectors. The feature extraction process is completely consistent with the feature extraction method in the above steps. Wavelet packet decomposition and time-domain statistical analysis are performed on the vibration signals in each group of historical monitoring data to extract three vibration feature values: root mean square, peak factor, and impulse factor. Short-time energy analysis is performed on the ultrasonic signals to extract three ultrasonic feature values: ultrasonic energy rate, count rate, and rise time. Time-domain averaging is performed on the displacement signals to extract real-time displacement values. Then, these seven feature values ​​are used to construct a feature dataset and normalized using the max-min normalization algorithm to obtain a seven-dimensional training sample feature vector with dimensions completely consistent with the above coupled feature vector. In the above application example, feature extraction and normalization are performed on 1500 groups of historical monitoring data to obtain 1500 seven-dimensional training sample feature vectors.

[0071] After obtaining the feature vector of the training samples, each training sample needs to be labeled to provide the target output for supervised learning. The label includes three types of information: the state category label is used to identify whether the bolt is in a fault-free, dynamic slip, or microcrack initiation state; the slip amount value label is used to quantify the specific slip amount of dynamic slip; and the crack depth value label is used to quantify the specific depth of microcracks. The state category label is represented as a three-dimensional vector using one-hot encoding. The fault-free state corresponds to [1, 0, 0], the dynamic slip state corresponds to [0, 1, 0], and the microcrack initiation state corresponds to [0, 0, 1]. The slip amount value label and the crack depth value label are both normalized scalar values. In the above example, among the 1500 training samples, 500 are in a fault-free state labeled [1, 0, 0] with both slip and crack depth being 0; 550 are in a dynamic slip state labeled [0, 1, 0] with slip ranging from 1 μm to 8 μm, which is normalized to 0.1 to 0.8; and 350 are in a microcrack initiation state labeled [0, 0, 1] with crack depth ranging from 0.1 mm to 1.5 mm, which is normalized to 0.067 to 1.0. The complete label for each training sample is a five-dimensional vector containing a three-dimensional one-hot encoding of the state category, a normalized slip value, and a normalized crack depth value.

[0072] After preparing the training sample feature vectors and labels, it is necessary to determine the structural parameters of the improved BP neural network. The number of input layer nodes must be consistent with the dimension of the training sample feature vectors to receive all feature inputs. Therefore, the number of input layer nodes is set to 7 to correspond to a seven-dimensional coupled feature vector. The number of output layer nodes must be consistent with the dimension of the labels to output complete prediction results. Therefore, the number of output layer nodes is set to 5 to correspond to a five-dimensional label vector. The setting of the number of hidden layer nodes needs to comprehensively consider the model's fitting ability and generalization ability. Too few nodes will lead to insufficient model fitting ability and inability to learn complex feature mapping relationships. Too many nodes will lead to model overfitting, reduce generalization ability, and increase computational complexity. According to the empirical formula, the number of hidden layer nodes is usually set to the square root of the sum of the number of input layer nodes and the number of output layer nodes, multiplied by an adjustment coefficient. In the above application example, setting the number of hidden layer nodes to 10 can achieve lightweight deployment while ensuring recognition accuracy. After determining the network structure, it is necessary to randomly initialize the connection weight matrix W1 between the input layer and the hidden layer with a dimension of 7 x 10 and the connection weight matrix W2 between the hidden layer and the output layer with a dimension of 10 x 5. At the same time, the hidden layer threshold vector b1 with a dimension of 10 and the output layer threshold vector b2 with a dimension of 5 are initialized. The initial values ​​of all weights and thresholds are randomly generated between -0.5 and +0.5.

[0073] The model training process uses batch gradient descent combined with momentum factor for weight update. The 1500 training sample feature vectors are divided into 1200 training set samples and 300 validation set samples in an 8:2 ratio. In each iteration, 32 samples are randomly selected from the training set as a batch for forward and backward propagation. During forward propagation, the training sample feature vectors in the batch are input into the input layer. The output of the hidden layer is calculated by the activation function f(x), which is equal to 1 divided by 1 plus e to the power of negative x. The hidden layer output H is equal to f(W1 multiplied by X plus b1), where X is the input training sample feature vector. The output Y of the output layer is equal to f(W2 multiplied by H plus b2). The resulting five-dimensional output vector is the predicted state vector. The mean squared error (MSE) between the predicted state vector and the true label is calculated as the sum of the squares of the differences between the predicted state vector and the true label, divided by the number of output layer nodes (5). The MSE is then checked against a preset convergence threshold of 0.001. If the MSE is greater than or equal to the threshold, backpropagation is performed to calculate the error gradient. The output layer's error gradient delta2 is equal to the predicted state vector minus the true label multiplied by the derivative of the Sigmoid function. The hidden layer's error gradient delta1 is equal to the transpose of W2 multiplied by delta2 and then by Si. The derivative of the gmoid function is used to calculate the weight adjustment based on the error gradient, a preset learning rate of 0.01, and a preset momentum factor of 0.9. The adjustment ΔW2 for W2 in this iteration is equal to the learning rate multiplied by the transpose of the hidden layer output H, multiplied by the output layer error gradient delta2, plus the momentum factor multiplied by the ΔW2 from the previous iteration. This weight adjustment includes the gradient influence of this iteration and the weight change from the previous iteration, thus accelerating convergence and reducing oscillations. The adjustments for W1, b1, and b2 are calculated in the same way, and all weights and thresholds are updated. The iterative process of forward propagation, error calculation, backpropagation, and weight update is repeated. After each round of training set traversal, the model performance is evaluated on the validation set. When the validation model meets the convergence threshold or reaches the maximum number of iterations, the model is output as the preset fault identification model.

[0074] After constructing the preset fault identification model, the coupled feature vector calculated above is input into the trained preset fault identification model for inference, outputting the state information of the bolt to be monitored. The state information includes no fault information, dynamic slip information, and crack information. For example, the coupled feature vector [0.136, 0.400, 0.320, 0.294, 0.444, 0.333, 0.158] is input into the trained preset fault identification model for inference. The preset fault identification model outputs a five-dimensional predicted state vector of [0.05, 0.92, 0.03, 0.16, 0.02]. The maximum value of the state category dimension, 0.92, corresponds to the dynamic slip state, so the identification result is dynamic slip information. The output of the slip amount numerical dimension, 0.16, after inverse normalization, yields an actual slip amount of 1.6 μm. The output of the crack depth numerical dimension, 0.02, which is close to 0, indicates that no crack has started. The preset fault identification model adopts an improved BP neural network as the core algorithm framework. The improvement lies in the introduction of a momentum factor to accelerate the convergence speed and enhance the generalization ability, enabling the model to quickly complete inference on the edge computing terminal to meet the real-time requirements of online monitoring. At the same time, through training with a large number of real working condition samples, the model has the ability to accurately identify micron-level dynamic slip and 0.1mm-level microcracks.

[0075] Those skilled in the art will understand that, in addition to using an improved BP neural network, the preset fault identification model can also be implemented using other machine learning or artificial intelligence models well known in the art, such as support vector machines (SVM), random forests, and deep learning networks (such as CNN and LSTM). The choice of these models should not be regarded as a limitation of the present invention.

[0076] S106: When the status information is dynamic slip information or crack information, obtain the material parameters of the bolt to be monitored, and perform prediction calculations on the material parameters, real-time operating condition parameters, and dynamic slip information or crack information to obtain the remaining service life of the bolt to be monitored.

[0077] In S106 above, when the state information output by the preset fault identification model is dynamic slip information or crack information, it is necessary to accurately predict the remaining service life of the bolt to be monitored. The reason for the need to predict the remaining service life rather than just staying at the fault identification stage is that the failure of wind turbine bolts under dynamic operating conditions is a gradual evolution process. There is a clear time window from micron-level dynamic slip to microcrack initiation, then to rapid crack propagation, and finally to bolt fracture. If the remaining service life of the bolt can be accurately predicted based on fault identification, it can provide wind farm operation and maintenance personnel with a forward-looking maintenance plan basis, realize the transformation from passive emergency maintenance to proactive preventive maintenance, and minimize the wind turbine shutdown and safety accidents caused by sudden bolt failure.

[0078] Traditional remaining life prediction methods are mostly based on unified empirical formulas or simplified models, failing to distinguish the evolution of two different failure mechanisms: dynamic slip and microcracks, resulting in insufficient prediction accuracy. This application constructs dedicated life prediction algorithms for these two failure mechanisms. For the microcrack initiation state, the Paris crack propagation formula in fracture mechanics is used for integral calculation to obtain the number of cycles required for the crack to propagate to the critical depth. For the dynamic slip state, the equivalent fatigue stress is calculated based on fretting wear theory, and the remaining fatigue life is predicted by combining Miner's cumulative damage theory. Both algorithms integrate the real-time operating parameters of the wind turbine and the bolt material parameters to achieve a high-precision prediction effect with a remaining life prediction error of less than 5%.

[0079] Furthermore, the remaining service life of the bolt to be monitored is obtained by predicting and calculating material parameters, real-time operating condition parameters, and dynamic slip or crack information. Specifically, this includes: obtaining fracture toughness value, fatigue crack propagation threshold, material constant, and stress-life curve from material parameters; when the status information is crack information, extracting the current crack depth from the crack information, calculating the stress intensity factor amplitude of the bolt to be monitored based on the real-time load data in the real-time operating condition parameters; constructing a functional relationship between crack propagation rate and stress intensity factor amplitude, integrating the current crack depth based on the functional relationship until the crack depth reaches the preset adjacent fracture depth, obtaining the remaining number of cycles, and converting the remaining number of cycles into the remaining service life. When the status information is dynamic slip information, the current slip amount is obtained from the dynamic slip information, and the equivalent fretting wear stress of the contact surface of the bolt to be monitored is calculated based on the current slip value; the corresponding theoretical total fatigue life is found in the stress life curve according to the equivalent fretting wear stress; the historical load cycle statistics of the bolt to be monitored are retrieved from the preset fault feature database, which includes the number of cycles executed under different stress levels; the ratio of the number of cycles executed under each stress level to the corresponding limit cycle number is calculated, and all ratios are summed to obtain the current cumulative fatigue damage degree of the bolt to be monitored; the difference between the theoretical total fatigue life and the current cumulative fatigue damage degree is multiplied by 1 to obtain the remaining service life.

[0080] Specifically, the material parameters of the bolt to be monitored are obtained. These material parameters are the basic data for fracture mechanics and fatigue analysis. Specifically, the fracture toughness value is used to determine whether the crack has reached the critical condition for unstable propagation; the fatigue crack propagation threshold is used to determine whether the crack will continue to propagate; the material constants, including the material coefficients and exponents in the Paris formula, are used to describe the relationship between the crack propagation rate and the stress intensity factor; and the stress-life curve, i.e., the SN curve, is used to find the corresponding number of fatigue life cycles based on the stress level. For example, if the bolt to be monitored is made of 40Cr material and is M36 specification, the standard material parameters for this material are retrieved from the material database of the wind power dedicated cloud platform. The fracture toughness value is 45 MPa √m, the fatigue crack propagation threshold is 6 MPa √m, the Paris formula material constant C is 2.8 x 10⁻¹² (in meters per cycle per MPa √m), the exponent m is 3.2, and the stress-life curve is the standard SN curve of 40Cr steel under alternating loads. This curve describes the number of cycles required for the bolt to reach fatigue failure under different stress amplitudes; for example, the fatigue life corresponding to a stress amplitude of 400 MPa is 1 million cycles.

[0081] When the status information is crack information, it indicates that the bolt under monitoring has developed a microcrack. It is necessary to predict the time required for the crack to propagate to the critical fracture depth, i.e., the remaining service life. The current crack depth is extracted from the crack information. Assuming that the crack depth output by the preset fault identification model is 0.15 mm, i.e. 0.00015 m, the stress intensity factor amplitude of the bolt under monitoring is calculated based on the real-time load data in the real-time operating condition parameters. The stress intensity factor is a key parameter in fracture mechanics that describes the stress field intensity at the crack tip. The calculation requires first determining the stress amplitude and then combining the crack depth and shape correction factor for calculation. The stress amplitude is obtained by dividing the real-time load by the effective cross-sectional area of ​​the bolt. For example, with a real-time load of 800 kN and an effective bolt cross-sectional area of ​​817 mm², the calculated stress amplitude is 979 MPa. The shape correction factor depends on the crack shape and bolt geometry. For a surface semi-elliptical crack, it is usually taken as 1.12. Multiplying the stress amplitude of 979 MPa by pi and the current crack depth of 0.00015 m, taking the square root of the product, and then multiplying by the shape correction factor of 1.12, we get a stress intensity factor amplitude of 24.5 MPa √m.

[0082] The Paris crack propagation formula is established to construct a functional relationship between crack propagation rate and stress intensity factor amplitude. This formula describes the incremental crack propagation in each load cycle as equal to the material constant C multiplied by the m-th power of the stress intensity factor amplitude. Substituting the material constant C (2.8 x 10^-12), the exponent m (3.2), and the stress intensity factor amplitude (24.5 MPa √m), the current crack propagation rate is calculated to be 1.1 x 10^-8 meters per cycle. Since the stress intensity factor increases with crack depth, the crack propagation rate also accelerates. Therefore, the Paris formula needs to be integrated to obtain the total number of cycles required for the crack to propagate from the current depth to the critical fracture depth. The integration process involves dividing the crack depth from the current value to the critical value into several small segments. Within each segment, the propagation rate is approximated as constant. The number of cycles required for propagation in each segment is calculated, and then the number of cycles for all segments is summed to obtain the total number of cycles. The critical fracture depth is determined by the fracture toughness criterion, i.e., when the stress intensity factor reaches the fracture toughness value of 45 MPa √m. When the crack becomes unstable and propagates, the critical fracture depth for the above application example is calculated to be 1.2 mm, or 0.0012 m. The remaining number of cycles is calculated to be 380,000 using the numerical integration method. Finally, the remaining number of cycles needs to be converted into the remaining service life, i.e., the actual operating time. Based on the speed data in the real-time operating parameters of the wind turbine, the wind turbine speed in the above application example is 15 revolutions per minute, or 900 load cycles per hour. The remaining service life is calculated to be 380,000 cycles divided by 900 cycles per hour and then divided by 24 hours, which gives approximately 17.5 days.

[0083] When the status information is dynamic slip information, it means that the bolt under monitoring is in the stage of fretting wear of the contact surface caused by preload decay and has not yet developed macroscopic cracks. It is necessary to predict the time required for the dynamic slip state to continue to develop and lead to bolt fatigue failure, i.e. the remaining service life, and obtain the current slip amount from the dynamic slip information.

[0084] For example, the default slip value output by the fault identification model is 5.2 μm, or 5.2 × 10⁻⁶. -6 Then, based on the current slip value, the equivalent fretting wear stress σeq of the bolt contact surface to be monitored is calculated. Fretting wear will generate additional shear stress and tensile stress on the contact surface, accelerating the accumulation of fatigue damage. The formula for calculating the equivalent fretting wear stress is σeq = base stress σbase multiplied by 1 plus k multiplied by δ to the power of n, where the base stress σbase is the nominal stress of the bolt under the current load, 979 MPa, and k is the fretting wear stress amplification factor, which is taken as 2.5 × 10 based on the roughness of the wind turbine bolt contact surface and material matching. 8 n is the nonlinear exponent, taken as 0.5, and the current slip is 5.2 × 10⁻⁶. -6The calculated equivalent fretting wear stress σeq is 1135 MPa. Based on the equivalent fretting wear stress, the corresponding theoretical total fatigue life Ntotal is found in the stress-life curve. For 40Cr steel, the fatigue life Ntotal corresponding to a stress amplitude of 1135 MPa in the SN curve is 2.3 × 10⁻⁶. 5 The loop continues.

[0085] Historical load cycle statistics of the bolt to be monitored are retrieved from a preset fault characteristic database. This data records the number of cycles performed by the bolt at different stress levels since installation. In the above application example, the bolt has already experienced 5 × 10 cycles at a stress level of 800 MPa. 4 Subsequent cycles, stress level 1000 MPa, 1.2 × 10 4 Sub-cycle, stress level 1200MPa, 3×10 3 In the next cycle, the ratio ni / Ni of the number of cycles performed to the corresponding limit cycle number at each stress level is calculated according to Miner's linear cumulative damage theory, and the sum is used to obtain the current cumulative fatigue damage degree D. For a stress level of 800 MPa, the limit cycle number is found to be 8 × 10 from the SN curve. 5 The damage ratio is 5×10 4 Divide by 8×10 5 The value equals 0.0625. Similarly, the damage ratios for other stress levels are calculated to be 0.0480 and 0.0150, respectively, and the cumulative fatigue damage degree D is 0.1255. The remaining service life, Nremain, is obtained by multiplying the theoretical total fatigue life Ntotal by 1 and subtracting the current cumulative fatigue damage degree D. This is equal to 2.3 × 10^5 multiplied by 1 minus 0.1255, which equals 2.01 × 10^5. 5 Based on a rotation speed of 15 r / min, or 900 cycles per hour, the remaining service life is approximately 9.3 days.

[0086] S107: Obtain the instantaneous load data at the moment the status information is generated, extract the corresponding fault response amplitude from the status information, calculate the load sensitivity coefficient of the bolt to be monitored by combining the instantaneous load data and the fault response amplitude.

[0087] In S107 above, after determining that the status information is dynamic slip or crack information, and predicting the remaining service life, it is also necessary to calculate the load sensitivity coefficient of the bolt to be monitored. This is because the load strength borne by the bolts of wind turbines varies significantly under different operating conditions. The same degree of dynamic slip or micro-crack is far more dangerous under high load conditions than under low load conditions. If warnings are issued solely based on the absolute value of slip or crack depth, it will lead to the risk of missed warnings under high load conditions and false alarms under low load conditions. The load sensitivity coefficient is a dimensionless parameter describing the sensitivity of the bolt fault response amplitude to load changes. The larger the load sensitivity coefficient, the more significant the bolt fault response under the current load, i.e., the more severe the bolt condition deterioration. By introducing the load sensitivity coefficient, the bolt fault characteristics under different load conditions can be normalized and compared, enabling bolt condition assessment across operating conditions. This provides a more accurate criterion for graded warnings and avoids warning failures caused by differences in operating conditions.

[0088] Furthermore, instantaneous load data at the moment the state information is generated is acquired, and the corresponding fault response amplitude is extracted from the state information. The instantaneous load data and fault response amplitude are calculated to obtain the load sensitivity coefficient of the bolt to be monitored. Specifically, this includes: acquiring the moment the state is generated; matching the corresponding real-time wind turbine torque data in the real-time operating condition parameters based on the moment of generation; using the real-time wind turbine torque data as the instantaneous load data; when the state information is dynamic slip information, extracting the peak-to-peak displacement value from the real-time displacement value associated with the dynamic slip information as the fault response amplitude; when the state information is crack information, extracting the root mean square value of vibration from the vibration characteristic value associated with the crack information as the fault response amplitude; calculating the ratio of the fault response amplitude to the instantaneous load data to obtain the current load response ratio; and dividing the difference between the current load response ratio and the reference load response ratio by the reference load response ratio to obtain the load sensitivity coefficient.

[0089] Specifically, the instantaneous load data at the moment the state information is generated is acquired. This load data reflects the actual stress level borne by the bolt when the fault characteristics are identified, and is a fundamental parameter for calculating the load sensitivity coefficient. The moment of state information generation is automatically output and timestamped by the preset fault identification model after completing state classification identification. In the above application example, assuming that the preset fault identification model identifies dynamic slip information with a slip of 5.2 micrometers, the recorded timestamp is 14:32:18 on March 7, 202x. This timestamp is accurate to the second to ensure the time matching accuracy with real-time operating condition parameters. Then, based on this generation moment, the corresponding real-time wind turbine torque data is matched in the real-time operating condition parameter database. Wind turbine torque is the parameter that most directly reflects the bolt's load state during wind turbine operation. Its magnitude is related to factors such as wind speed, blade pitch angle, and generator output power, and can comprehensively reflect the dynamic load intensity borne by the bolt at that moment. In the above application example, the wind turbine torque data at 14:32:18 on March 7, 202x was retrieved from the real-time operating condition parameter database and found to be 1850 kNm. This real-time wind turbine torque data was used as the instantaneous load data for subsequent calculations. Since the real-time operating condition parameters are collected once per second, the instantaneous load data and the time of generation of the status information can be accurately matched, ensuring the timeliness and accuracy of the load sensitivity coefficient calculation.

[0090] The fault response amplitude is extracted from the status information. The fault response amplitude is a key indicator characterizing the severity of the bolt's fault characteristics in its current state. Different types of status information require different fault response amplitudes to accurately reflect the actual severity of the fault. When the status information is dynamic slip information, the peak-to-peak displacement value associated with the dynamic slip information is extracted as the fault response amplitude. The peak-to-peak displacement value is chosen over the average displacement value because the harmfulness of dynamic slip is mainly reflected in the maximum amplitude fluctuation during the slip process. The peak-to-peak value, i.e., the difference between the maximum and minimum values ​​of the displacement signal within a monitoring cycle, can fully reflect the intensity of the dynamic fluctuation of the slip. For example, the peak value of the real-time displacement value collected by the micro-displacement auxiliary sensing module at the moment the status information is generated within the past 10-second monitoring cycle is 7.3 micrometers, and the valley value is 3.1 micrometers. The calculated peak-to-peak displacement value is 4.2 micrometers, which is then used as the fault response amplitude under dynamic slip conditions. When the state information is crack information, the root mean square (RMS) value of vibration needs to be extracted from the vibration characteristic values ​​associated with the crack information as the fault response amplitude. The RMS value of vibration is an effective parameter describing the intensity of vibration energy. The initiation and propagation of microcracks will lead to a decrease in the local stiffness of the bolt, thereby causing an abnormal increase in vibration energy. The RMS value of vibration can sensitively reflect this change. For example, the vibration signal collected by the vibration-resistant low-frequency accelerometer at the moment the state information is generated, after wavelet packet decomposition and time-domain statistical analysis, yields a RMS value of 0.38 volts in the 500 to 2000 Hz frequency band. This RMS value of vibration is used as the fault response amplitude under the crack state.

[0091] The current load response ratio is calculated by combining instantaneous load data and fault response amplitude. This ratio, which is the ratio of the fault response amplitude to the instantaneous load data, reflects the intensity of the fault response generated by the bolt under a unit load and serves as an intermediate parameter in calculating the load sensitivity coefficient. For example, in the dynamic slip condition, dividing the fault response amplitude of 4.2 micrometers by the instantaneous load data of 1850 kNm yields a current load response ratio of 0.00227 micrometers per kNm. This ratio indicates that for every 1 kNm increase in rotor torque, the average peak-to-peak value of the bolt's dynamic slip displacement increases by 0.00227 micrometers. In the crack condition, dividing the fault response amplitude of 0.38 volts by the instantaneous load data of 1850 kNm yields a current load response ratio of 0.000205 volts per kNm. This ratio indicates that for every 1 kNm increase in rotor torque, the average root mean square value of the bolt vibration increases by 0.000205 volts.

[0092] The load sensitivity coefficient is obtained by comparing the current load response ratio with the reference load response ratio. The reference load response ratio is the theoretical value of the load response ratio of the bolt under the same load condition in a healthy state. This reference value is stored in the bolt fault characteristic database of the wind power dedicated cloud platform, and is classified and stored according to different bolt models, different installation locations, and different materials. For example, for 40Cr material M36 specification tower flange connection bolts, the reference load response ratio of this type of bolt under the condition of a wind turbine torque of 1850 kNm in a healthy state is retrieved from the database. The reference load response ratio corresponding to the dynamic slip state is 0.00150 micrometers per kNm, and the reference load response ratio corresponding to the crack state is 0.000135 volts per kNm. The load sensitivity coefficient is obtained by dividing the difference between the current load response ratio and the reference load response ratio by the reference load response ratio. For the dynamic slip state, the calculation process is 0.00227 minus 0.00150 and then divided by 0.00150, resulting in a load sensitivity coefficient of 0.513. This coefficient indicates that the dynamic slip response of the bolt under the current load condition is 51.3% stronger than that under the healthy state, indicating that the decrease in bolt preload has led to a significant abnormal load response. For the crack state, the calculation process is 0.000205 minus 0.000135 and then divided by 0.000135, resulting in a load sensitivity coefficient of 0.519. This coefficient indicates that the vibration response of the bolt under the current load condition is 51.9% stronger than that under the healthy state, indicating that the initiation of microcracks has led to a significant decrease in the local stiffness of the bolt. The calculated load sensitivity coefficients are uploaded to the wind power dedicated cloud platform and used together with the remaining service life prediction results as the basis for graded early warning.

[0093] S108: Obtain wind resource forecast data, generate early warning levels and comprehensive maintenance plans based on remaining service life, load sensitivity coefficient, status information and wind resource forecast data, and send the early warning levels and comprehensive maintenance plans to the operation and maintenance terminal.

[0094] In S108 above, after completing the remaining service life prediction and load sensitivity coefficient calculation, it is necessary to further integrate wind resource prediction data to generate accurate early warning levels and formulate targeted comprehensive maintenance plans. Introducing wind resource prediction data, rather than relying solely on the current bolt status for early warning, is crucial because the failure risk of wind turbine bolts depends not only on their current health status but also significantly on future operating conditions, especially extreme wind loads. Bolts in abnormal conditions may not fail in the short term under low wind speeds, but the risk of failure will increase dramatically if strong winds occur in the coming days. Traditional monitoring methods only set fixed early warning levels based on the current state, without considering the accelerating effect of future operating condition changes on the failure evolution rate. This leads to a failure to provide timely warnings before extreme weather arrives, or excessive warnings under stable weather conditions, causing unnecessary downtime losses.

[0095] Furthermore, wind resource forecast data is acquired, and early warning levels and comprehensive maintenance plans are generated based on remaining service life, load sensitivity coefficient, status information, and wind resource forecast data. Specifically, this includes: extracting the maximum predicted wind speed within a preset future time window from the wind resource forecast data; determining the basic health status based on status information and remaining service life; determining the basic health status as high-risk when the status information is crack information or the remaining service life is less than or equal to the service life threshold; determining the basic health status as abnormal when the status information is dynamic slip information and the remaining service life is greater than the service life threshold; calculating the product of the maximum predicted wind speed and the load sensitivity coefficient to obtain the future risk index; generating a Level 1 early warning level when the basic health status is high-risk, or when the basic health status is abnormal and the future risk index is greater than or equal to the preset risk threshold; generating a Level 2 early warning level when the basic health status is abnormal and the future risk index is less than the preset risk threshold; generating a maintenance plan including shutdown instructions and bolt replacement strategies if the bolt to be monitored is at Level 1 early warning level; and generating a maintenance plan including bolt preload re-tightening strategies if the bolt to be monitored is at Level 2 early warning level.

[0096] Specifically, wind resource forecast data is acquired and the maximum predicted wind speed within a preset future time window is extracted. The wind resource forecast data is obtained in real time by the wind power-dedicated cloud platform from the short-term wind speed forecast module of the National Meteorological Administration's numerical weather prediction system or the on-site meteorological station of the wind farm. This data includes hourly wind speed forecasts and prediction confidence levels for the next seven to ten days. Considering that the time scale for bolt remaining life prediction is mostly several days to several weeks, this application sets the next seven days as the preset time window to match the prediction cycle of the remaining life. For example, the cloud platform acquired wind speed forecast data for the wind farm for the next seven days on March 7, 202x. The data shows that from March 8 to March 10, the weather will be sunny and light with a maximum wind speed not exceeding 8 meters per second. However, from March 11 to March 13, strong winds will occur due to the influence of cold air. Among them, the predicted wind speed reaches a maximum of 22 meters per second at 2:00 AM on March 12, corresponding to a wind load level of gale-force wind. This maximum predicted wind speed of 22 meters per second is extracted as the extreme value parameter within the preset future time window.

[0097] The basic health status is determined based on the status information and remaining service life. The basic health status is the current danger level of the bolt without considering future changes in operating conditions, and is divided into two levels: high-risk and abnormal. The criterion logic is that when the status information is crack information, it indicates that the bolt has developed a macroscopic crack. Under alternating loads, the crack propagates extremely rapidly and may become unstable and fracture at any time; regardless of the remaining service life, it is judged as a high-risk state. For example, if the preset fault identification model outputs crack information with a crack depth of 0.15 mm, the basic health status is directly determined to be a high-risk state. Another criterion is that when the remaining service life is less than or equal to the service life threshold, even if the status information is dynamic slip information and no macroscopic cracks have appeared, it is still judged as a high-risk state due to the extremely short remaining time. The service life threshold is set based on the wind turbine maintenance response time. Generally, it takes 2 to 3 days for a wind farm to complete bolt replacement from receiving an early warning; considering safety redundancy, the service life threshold is set to 5 days, or 120 hours. For example, if the remaining service life is 17.5 days, which is greater than the service life threshold of 120 hours, and the status information is dynamic slip information with a slip amount of 5.2 micrometers, then the basic health status is determined to be an abnormal state. The classification of basic health status provides an initial judgment benchmark for subsequent dynamic early warning level adjustments. High-risk states should, in principle, require immediate maintenance, while abnormal states require assessment of the urgency level based on future operating conditions.

[0098] The index is obtained by multiplying the maximum predicted wind speed by the load sensitivity coefficient. The physical meaning of this calculation method is that the load sensitivity coefficient describes the sensitivity of the bolt's failure response to load, and the maximum predicted wind speed is converted into the extreme load intensity the bolt will bear in the future. Multiplying the two gives the bolt's failure risk amplification factor under future extreme conditions. For example, with a load sensitivity coefficient of 0.513 and a maximum predicted wind speed of 22 meters per second, the calculated future risk index is 11.29. This value indicates that under strong wind conditions on March 12th, the bolt's dynamic slip response or crack propagation rate will be amplified by 11.29 times compared to the current state, significantly increasing the failure risk. To determine whether this risk index reaches the level requiring an upgrade in the warning level, a preset risk threshold of 8.0 is set. This threshold is determined through statistical analysis of historical failure cases. When the future risk index exceeds 8.0, the probability of bolt failure under extreme conditions exceeds 30%, requiring emergency measures to be taken in advance. For example, a future risk index of 11.29 is greater than the preset risk threshold of 8.0, indicating that future strong wind conditions pose a serious threat to the bolt.

[0099] The final warning level is generated based on a combination of basic health status and future risk index. When the basic health status is high-risk, a Level 1 warning (red warning) is generated directly regardless of the future risk index, because the bolts are in a dangerous state of imminent failure and must be shut down for maintenance immediately. For example, if the bolts are cracked or have less than 5 days of remaining life, a Level 1 warning is triggered. When the basic health status is abnormal, the future risk index needs to be further assessed. If the future risk index is greater than or equal to the preset risk threshold of 8.0, it indicates that extreme operating conditions in the future will accelerate bolt failure. Although the current status is abnormal, a Level 1 warning is still generated considering the future risk. For example, a bolt slippage of 5.2 micrometers and a remaining life of 17.5 days is an abnormal status, but because the future risk index of 11.29 exceeds the threshold of 8.0, the warning level is upgraded from Level 2 (yellow warning) to Level 1 (red warning). The cloud platform triggers a red warning from the on-site audible and visual alarm and pushes an emergency warning message to the maintenance PC and mobile APP. At the same time, a triple warning notification is sent to the wind farm duty personnel via SMS, reminding them that bolt maintenance must be completed before the arrival of strong winds on March 12. If the future risk index is less than the preset risk threshold of 8.0, it indicates that the future operating conditions are stable and the risk of bolt failure in the short term is controllable, thus generating a Level 2 warning, i.e., a yellow warning. For example, if the maximum predicted wind speed is only 12 meters per second, the future risk index is 6.16, which is less than the threshold of 8.0, generating a Level 2 warning to guide maintenance personnel to complete the pre-tightening re-tightening operation during routine inspections without the need for emergency shutdown.

[0100] The comprehensive maintenance plan is formulated by matching differentiated operation and maintenance strategies according to the warning level. If the monitored bolt is at the first warning level, a maintenance plan is generated that includes a shutdown order and a bolt replacement strategy. The shutdown order clearly requires the wind turbine to be safely shut down before 18:00 on March 11, 12 hours before the arrival of strong winds, to avoid major accidents such as bolt breakage, blade ejection, or tower collapse caused by operation under extreme wind loads of 22 meters per second. The bolt replacement strategy details the bolt number to be replaced, which is the tower flange connection bolt, made of 40Cr material and M36 specification. The required number of spare parts is 12, including 10% redundancy. The recommended replacement process is to unload first, then replace, and then tighten in three stages according to the standard pre-tightening torque of 280 Nm. This plan is pushed to the operation and maintenance PC and mobile APP through the cloud platform. After the operation and maintenance personnel confirm receipt, the system automatically generates a work order and dispatches spare parts inventory and maintenance personnel to ensure that the bolt replacement is completed within the specified time and the wind turbine's safe operation capability is restored.

[0101] If the bolt to be monitored is at the Level 2 warning level, a maintenance plan including a bolt preload re-tightening strategy is generated. This plan does not require shutdown. Maintenance personnel can use a special torque wrench to compensate for the preload of the bolt under low wind speed operating conditions of the wind turbine. The re-tightening torque is set to 80% of the initial preload torque, i.e., 224 Nm. After re-tightening, the slippage is verified by the micro-displacement auxiliary sensing module to see if it has dropped below 2 micrometers. This plan is pushed to maintenance personnel on the cloud platform in the form of a yellow alert and is included in the daily inspection plan. This avoids frequent shutdowns and waste of maintenance resources caused by excessive warnings, and achieves accurate matching between warnings and maintenance, significantly improving the economy and practicality of the wind turbine bolt monitoring system.

[0102] In one possible implementation, in addition to generating early warning levels based on remaining service life, load sensitivity coefficient, and wind resource prediction data, the crack depth and slippage output by a preset fault identification model can be compared in real time with their respective fixed thresholds. Based on the comparison results, the corresponding early warning level and maintenance plan can be directly generated. For example... Figure 4 As shown, the coupled feature vector is input into the preset fault identification model for identification to obtain status information. When the status information is no fault information, daily monitoring of the bolt to be monitored continues according to the sampling frequency. When the status information is microcrack initiation information, the crack depth output by identification is compared with the crack depth threshold. If the crack depth is greater than or equal to the crack depth threshold of 0.1 mm, it indicates that the bolt has developed a macro crack and reached the detectable scale. At this time, a first-level red warning level is directly generated. The cloud platform triggers the red warning of the on-site audible and visual alarm and pushes emergency warning information to the operation and maintenance PC and mobile APP at the same time. At the same time, a triple warning notification is sent to the wind farm duty personnel via remote SMS. The generated maintenance plan includes an immediate shutdown instruction and an emergency bolt replacement strategy, which clearly requires the wind turbine to be safely shut down and the faulty bolt to be replaced within 12 hours to avoid the rapid expansion of cracks under alternating loads, which may lead to bolt breakage and tower collapse or other major safety accidents. When the crack depth is less than the crack depth threshold of 0.1 mm, the monitoring data is recorded. When the status information is dynamic slip information, the slip amount output by identification is compared with the slip amount threshold. If the slip amount is greater than or equal to the slip amount threshold of 5 micrometers, it indicates that the bolt preload has significantly decreased and a measurable relative displacement has occurred between the bolt and the connector. At this time, a level two yellow warning is generated, the cloud platform triggers the yellow warning of the on-site audible and visual alarm and pushes a preload re-tightening prompt to the maintenance personnel. The generated maintenance plan includes a bolt preload re-tightening strategy, which specifies that the re-tightening torque is set to 80% of the initial preload torque. The maintenance personnel can use a special torque wrench to complete the re-tightening operation without stopping the machine under the low wind speed operation condition of the wind turbine. After re-tightening, the micro-displacement auxiliary sensing module is used to verify whether the slip amount has dropped to below 2 micrometers to ensure that the bolt returns to the normal preload state. When the slip amount is less than the slip amount threshold of 5 micrometers, the monitoring data of this time is recorded.

[0103] like Figure 4As shown, after determining that the state information output by the preset fault identification model is dynamic slip information or crack information, in addition to determining the remaining service life through the above-mentioned scheme, the remaining service life can also be predicted by training an LSTM variable parameter dynamic life evolution model by fusing bolt material parameters, real-time wind turbine operation data and bolt state identification data. The remaining service life can be accurately calculated by a hybrid driving modeling method based on the Paris formula of fracture mechanics and long short-term memory neural network. When the state information is dynamic slip information, the historical slip evolution curve of the bolt is extracted from the bolt fault feature database of the cloud platform. The slip time series data of the most recent 168 hours (7 days) is extracted using the sliding time window method. This time series data is input into the input layer of the pre-trained LSTM variable parameter dynamic life evolution model. The input layer of the LSTM model contains 6 neurons corresponding to 6 time-varying parameters, namely the current slip amount, slip change rate, wind turbine average wind speed, average rotational speed, cumulative load cycle count, and ambient temperature. The hidden layer adopts a 2-layer LSTM unit structure, with each layer containing 128 memory cells to capture the long-term dependency of slip evolution. The output layer outputs the remaining cycle count of the bolt due to the continuous decay of preload through a fully connected layer. Then, the remaining cycle count is converted into the remaining service life in hours by combining the daily average load cycle frequency in the real-time operation data of the wind turbine. For example, when the bolt slip is 3 micrometers, the LSTM model predicts based on the historical slip evolution trend that under the current operating conditions of wind speed of 8 meters per second and rotation speed of 15 revolutions per minute, the bolt preload will decay to the critical failure value after 420 hours, corresponding to a remaining service life of 75% of the design service life of 560 hours. When the status information is crack information, the crack depth output by identification is extracted as the initial crack size. The critical depth for crack instability propagation is determined based on the fatigue crack propagation threshold and fracture toughness in the bolt material parameters. A quantitative relationship between the crack propagation rate and the stress intensity factor amplitude is established using the Paris formula. The stress intensity factor amplitude is calculated from real-time load data of the wind turbine. Considering the randomness and time-varying nature of wind turbine loads, the material constants and exponents in the Paris formula are set as learnable parameters for the LSTM model. The LSTM model is trained under supervision using measured crack propagation data of replaced bolts from the cloud platform's fault feature database. After training, the LSTM variable parameter model can dynamically adjust the Paris formula parameters based on the wind speed forecast data and rotational speed plan data for the next 7 days of wind turbine operation. Iteratively, it calculates the number of load cycles required for the crack to propagate from the current depth to the critical depth, and then outputs the remaining service life in hours. For example, if the bolt crack depth is 0.15 mm, the LSTM variable parameter model, combined with future strong wind weather forecast data, calculates that the crack will propagate to the critical depth of 3 mm within 72 hours, triggering instability and fracture.The calculated remaining service life is compared with a remaining service life threshold, which can be set to 20%. When the calculated remaining service life is less than or equal to the remaining service life threshold, a level-two red alert is triggered. When the calculated remaining service life is greater than the remaining service life threshold, the monitoring data for this session is recorded.

[0104] This early warning method, which directly compares crack depth or slip amount with a fixed threshold, is simpler and more intuitive than dynamic early warning methods that integrate multiple parameters. It does not need to consider the impact of load sensitivity coefficient and future wind resource forecast data, making it suitable for application scenarios that require rapid judgment of the current state of bolts. Especially when edge computing resources are limited or network connections are unstable, making it impossible to obtain wind resource forecast data in real time, this method can serve as a backup early warning mechanism to ensure that critical faults are not missed. It complements the dynamic early warning method, jointly ensuring the reliability and robustness of the wind turbine bolt monitoring system and providing multiple technical guarantees for the safe and stable operation of wind turbines.

[0105] This application also provides a monitoring and early warning system based on ultrasonic coupling of wind turbine bolt vibration. The system includes a data acquisition unit, a compensation unit, an extraction unit, an identification unit, and an early warning unit. The acquisition unit receives vibration signals, ultrasonic signals, and displacement signals that are synchronously acquired from the bolt to be monitored. The compensation unit acquires the real-time operating condition parameters of the wind turbine and performs compensation processing on the vibration signal, ultrasonic signal and displacement signal based on the real-time operating condition parameters to obtain the compensated vibration signal, compensated ultrasonic signal and compensated displacement signal. The extraction unit analyzes the compensated vibration signal and the compensated ultrasonic signal, and extracts vibration feature values ​​and ultrasonic feature values ​​respectively; it extracts real-time displacement values ​​from the compensated displacement signal, and normalizes the vibration feature values, ultrasonic feature values ​​and real-time displacement values ​​to obtain a coupled feature vector. The identification unit inputs the coupled feature vector into the preset fault identification model for processing to obtain the state information of the bolt to be monitored. The state information includes fault-free information, dynamic slippage information, and crack information. The early warning unit, when the status information is dynamic slip or crack information, acquires the material parameters of the bolt to be monitored, performs predictive calculations on the material parameters, real-time operating condition parameters, and dynamic slip or crack information to obtain the remaining service life of the bolt to be monitored; acquires the instantaneous load data at the moment the status information is generated, extracts the corresponding fault response amplitude from the status information, calculates the load sensitivity coefficient of the bolt to be monitored based on the instantaneous load data and fault response amplitude; acquires wind resource prediction data, and generates an early warning level and a comprehensive maintenance plan based on the remaining service life, load sensitivity coefficient, status information, and wind resource prediction data, and sends the early warning level and comprehensive maintenance plan to the operation and maintenance terminal.

[0106] In one possible implementation, the acquisition unit is used to acquire real-time rotational speed and real-time wind speed from real-time operating condition parameters, determine the fundamental frequency and harmonic frequencies of the rotating machinery based on the real-time rotational speed, and determine the frequency band range of wind load impact based on the real-time wind speed; the compensation unit is used to locate the amplitude components that coincide with the fundamental frequency and harmonic frequencies, as well as the amplitude components located within the wind load impact frequency band range, in the frequency domain spectrum of the vibration signal to obtain the background vibration interference component; subtract the background vibration interference component from the vibration signal to obtain the denoised vibration signal; acquire real-time temperature data and real-time salt spray concentration data around the bolt to be monitored, calculate the temperature difference between the real-time temperature data and the preset calibration temperature; multiply the temperature difference by the temperature sensitivity offset coefficient to obtain the sensitivity drift, and adjust the sensitivity drift accordingly. The amplitudes of the denoised vibration, ultrasonic, and displacement signals are reverse-corrected, and the ultrasonic signal is gain-compensated based on the signal transmission attenuation rate corresponding to the real-time salt spray concentration data to obtain the environmentally corrected vibration, ultrasonic, and displacement signals. Real-time load data is obtained from the real-time operating condition parameters, and the nominal stress is calculated based on the real-time load data and the cross-sectional area of ​​the bolt to be monitored. The dynamic stress concentration factor is determined based on the nominal stress, and a distortion correction factor is constructed based on the dynamic stress concentration factor. The environmentally corrected vibration, ultrasonic, and displacement signals are divided by the distortion correction factor to obtain the compensated vibration, ultrasonic, and displacement signals.

[0107] In one possible implementation, the extraction unit performs wavelet packet decomposition on the compensated vibration signal to obtain sub-signals in different frequency bands, selects sub-signals within a preset frequency range from the sub-signals in different frequency bands as characteristic frequency band signals; performs time-domain statistical analysis on the characteristic frequency band signals to calculate the root mean square value, peak factor, and impulse factor of the characteristic frequency band signals, and combines the root mean square value, peak factor, and impulse factor as vibration characteristic values; performs short-time energy analysis on the compensated ultrasonic signal to calculate the short-time energy change rate of the ultrasonic signal within a time window function to obtain the acoustic ultrasonic energy rate; counts the number of times the ultrasonic signal exceeds the trigger threshold, divides the target number by the unit time to obtain the count rate; calculates the time interval from the first time the ultrasonic signal exceeds the trigger threshold to the signal peak to obtain the rise time; and combines the acoustic ultrasonic energy rate, count rate, and rise time as ultrasonic characteristic values.

[0108] In one possible implementation, the extraction unit is used to perform time-domain analysis on the compensated displacement signal, calculate the average amplitude of the displacement signal within a preset sampling period, and use the average amplitude as the real-time displacement value; construct a feature dataset by combining the vibration feature value, ultrasonic feature value, and real-time displacement value; process each feature data in the feature dataset using the max-min normalization algorithm to obtain multiple normalized feature data; and sort the multiple normalized feature data in a preset order to obtain a multidimensional coupled feature vector.

[0109] In one possible implementation, the acquisition unit is used to obtain fracture toughness values, fatigue crack propagation thresholds, material constants, and stress-life curves from material parameters; the early warning unit is used to extract the current crack depth from the crack information when the status information is crack information, calculate the stress intensity factor amplitude of the bolt to be monitored based on the real-time load data in the real-time operating condition parameters; construct a functional relationship between crack propagation rate and stress intensity factor amplitude, perform integral calculation on the current crack depth based on the functional relationship until the crack depth reaches the preset neighbor fracture depth, obtain the remaining cycle count, and convert the remaining cycle count into the remaining service life; when the status information is dynamic slip information, it is used to extract the current crack depth from the crack information when the status information is dynamic slip information. The current slip amount is obtained from the slip information, and the equivalent fretting wear stress of the contact surface of the bolt to be monitored is calculated based on the current slip value. The corresponding theoretical total fatigue life is found in the stress life curve according to the equivalent fretting wear stress. The historical load cycle statistics of the bolt to be monitored are retrieved from the preset fault feature database. The historical load cycle statistics include the number of cycles executed under different stress levels. The ratio of the number of cycles executed under each stress level to the corresponding limit cycle number is calculated, and all ratios are summed to obtain the current cumulative fatigue damage of the bolt to be monitored. The difference between the theoretical total fatigue life and the current cumulative fatigue damage is multiplied by 1 to obtain the remaining service life.

[0110] In one possible implementation, the early warning unit is used to acquire the time of state generation, match the corresponding real-time wind turbine torque data in the real-time operating condition parameters based on the time of generation, and use the real-time wind turbine torque data as instantaneous load data; when the state information is dynamic slip information, the peak-to-peak value of the displacement is extracted from the real-time displacement value associated with the dynamic slip information as the fault response amplitude; when the state information is crack information, the root mean square value of vibration is extracted from the vibration characteristic value associated with the crack information as the fault response amplitude; the ratio of the fault response amplitude to the instantaneous load data is calculated to obtain the current load response ratio; the difference between the current load response ratio and the reference load response ratio is divided by the reference load response ratio to obtain the load sensitivity coefficient.

[0111] In one possible implementation, the early warning unit is used to extract the maximum predicted wind speed within a preset future time window from wind resource forecast data; determine the basic health status based on status information and remaining service life; when the status information is crack information or the remaining service life is less than or equal to the service life threshold, the basic health status is determined to be a high-risk status; when the status information is dynamic slip information and the remaining service life is greater than the service life threshold, the basic health status is determined to be an abnormal status; calculate the product of the maximum predicted wind speed and the load sensitivity coefficient to obtain the future risk index; when the basic health status is a high-risk status, or the basic health status is an abnormal status and the future risk index is greater than or equal to the preset risk threshold, a first-level early warning level is generated; when the basic health status is an abnormal status and the future risk index is less than the preset risk threshold, a second-level early warning level is generated; if the bolt to be monitored is at the first-level early warning level, a maintenance plan including a shutdown command and a bolt replacement strategy is generated; if the bolt to be monitored is at the second-level early warning level, a maintenance plan including a bolt preload re-tightening strategy is generated.

[0112] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0113] This application also discloses an electronic device. (See reference...) Figure 5 , Figure 5 This application provides a schematic diagram of the structure of an electronic device. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 502, and at least one communication bus 505.

[0114] The communication bus 505 is used to enable communication between these components.

[0115] The user interface 503 may include a display screen and a camera. Optionally, the user interface 503 may also include a standard wired interface and a wireless interface.

[0116] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0117] The processor 501 may include one or more processing cores. The processor 501 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 502, and by calling data stored in memory 502. Optionally, the processor 501 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 501 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and application requests; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 501 and may be implemented as a separate chip.

[0118] The memory 502 may include random access memory (RAM) or read-only memory. Optionally, the memory 502 may include a non-transitory computer-readable storage medium. The memory 502 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 502 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory 502 may also be at least one storage device located remotely from the aforementioned processor 501.

[0119] like Figure 5 As shown, the memory 502, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application for monitoring and early warning based on ultrasonic coupling of wind turbine bolt vibration.

[0120] exist Figure 5 In the electronic device 500 shown, the user interface 503 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 501 can be used to call the application program stored in the memory 502 for monitoring and early warning based on the ultrasonic coupling of wind turbine bolt vibration. When executed by one or more processors, the electronic device performs one or more of the methods described in the above embodiments.

[0121] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0122] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0123] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some service interfaces; indirect couplings or communication connections between devices or units may be electrical or other forms.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0126] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0127] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and practical application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure.

Claims

1. A monitoring and early warning method based on ultrasonic coupling of wind turbine bolt vibration, characterized in that, The method includes: It receives vibration signals, ultrasonic signals, and displacement signals that are synchronously acquired from the bolts to be monitored. The real-time operating condition parameters of the wind turbine are obtained, and the vibration signal, the ultrasonic signal and the displacement signal are compensated based on the real-time operating condition parameters to obtain the compensated vibration signal, the compensated ultrasonic signal and the compensated displacement signal. The compensated vibration signal and the compensated ultrasonic signal are analyzed to extract vibration feature values ​​and ultrasonic feature values, respectively. Real-time displacement values ​​are extracted from the compensated displacement signal, and the vibration feature values, ultrasonic feature values, and real-time displacement values ​​are normalized to obtain a coupled feature vector. The coupled feature vector is input into a preset fault identification model for processing to obtain the state information of the bolt to be monitored. The state information includes fault-free information, dynamic slippage information, and crack information. When the status information is the dynamic slip information or the crack information, the material parameters of the bolt to be monitored are obtained, and the material parameters, the real-time operating condition parameters, and the dynamic slip information or the crack information are predicted and calculated to obtain the remaining service life of the bolt to be monitored. The instantaneous load data at the moment the state information is generated is obtained, and the corresponding fault response amplitude is extracted from the state information. The load sensitivity coefficient of the bolt to be monitored is calculated by calculating the instantaneous load data and the fault response amplitude. Obtain wind resource forecast data, generate early warning levels and comprehensive maintenance plans based on the remaining service life, the load sensitivity coefficient, the status information, and the wind resource forecast data, and send the early warning levels and comprehensive maintenance plans to the operation and maintenance terminal.

2. The method according to claim 1, characterized in that, The compensation processing of the vibration signal, the ultrasonic signal, and the displacement signal based on the real-time operating condition parameters to obtain the compensated vibration signal, the compensated ultrasonic signal, and the compensated displacement signal specifically includes: The real-time rotational speed and real-time wind speed are obtained from the real-time operating condition parameters. The fundamental frequency and harmonic frequency of the rotating machinery are determined based on the real-time rotational speed, and the frequency band range of wind load impact is determined based on the real-time wind speed. In the frequency domain spectrum of the vibration signal, the amplitude components that coincide with the fundamental frequency and the harmonic frequency, as well as the amplitude components located within the wind load impact frequency band, are used to obtain the background vibration interference components. Subtract the background vibration interference component from the vibration signal to obtain the denoised vibration signal; Acquire real-time temperature data and real-time salt spray concentration data around the bolt to be monitored, and calculate the temperature difference between the real-time temperature data and the preset calibration temperature; Multiply the temperature difference by the temperature sensitivity offset coefficient to obtain the sensitivity drift. Based on the sensitivity drift, the amplitudes of the denoised vibration signal, the ultrasonic signal, and the displacement signal are reverse-corrected. Based on the signal transmission attenuation rate corresponding to the real-time salt spray concentration data, the ultrasonic signal is gain-compensated to obtain the environmentally corrected vibration signal, the environmentally corrected ultrasonic signal, and the environmentally corrected displacement signal. Real-time load data is obtained from the real-time operating condition parameters. The nominal stress is calculated based on the real-time load data and the cross-sectional area of ​​the bolt to be monitored. The dynamic stress concentration factor is determined based on the nominal stress. The distortion correction factor is constructed based on the dynamic stress concentration factor. The environmentally corrected vibration signal, the environmentally corrected ultrasonic signal, and the environmentally corrected displacement signal are each divided by the distortion correction factor to obtain the compensated vibration signal, the compensated ultrasonic signal, and the compensated displacement signal.

3. The method according to claim 2, characterized in that, The analysis of the compensated vibration signal and the compensated ultrasonic signal, and the extraction of vibration feature values ​​and ultrasonic feature values ​​respectively, specifically includes: The compensated vibration signal is decomposed by wavelet packet to obtain sub-signals of different frequency bands. Sub-signals within a preset frequency range are selected from the sub-signals of different frequency bands as feature frequency band signals. Time-domain statistical analysis is performed on the characteristic frequency band signal to calculate the root mean square value, peak factor, and impulse factor of the characteristic frequency band signal, and the root mean square value, peak factor, and impulse factor are combined as the vibration characteristic value. Short-time energy analysis is performed on the compensated ultrasonic signal to calculate the short-time energy change rate of the ultrasonic signal within a time window function, thereby obtaining the ultrasonic energy rate. The number of times the ultrasound signal exceeds the trigger threshold is counted, and the count rate is obtained by dividing the number of times by the unit time. The rise time is obtained by calculating the time interval from when the ultrasound signal first exceeds the trigger threshold to when it reaches the signal peak. The ultrasonic energy rate, the count rate, and the rise time are combined as the ultrasonic characteristic value.

4. The method according to claim 2, characterized in that, The step of extracting real-time displacement values ​​from the compensated displacement signal and normalizing the vibration feature values, ultrasonic feature values, and real-time displacement values ​​to obtain a coupled feature vector specifically includes: The compensated displacement signal is subjected to time-domain analysis to calculate the average amplitude of the displacement signal within a preset sampling period, and the average amplitude is used as the real-time displacement value. The vibration feature values, the ultrasonic feature values, and the real-time displacement values ​​are used to construct a feature dataset. The maximum-minimum normalization algorithm is used to process each feature data in the feature dataset to obtain multiple normalized feature data. The normalized feature data are sorted in a preset order to obtain the multidimensional coupled feature vector.

5. The method according to claim 1, characterized in that, The step of predicting and calculating the remaining service life of the bolt to be monitored by analyzing the material parameters, the real-time operating condition parameters, and the dynamic slip information or the crack information specifically includes: The fracture toughness value, fatigue crack propagation threshold, material constant, and stress-life curve are obtained from the material parameters. When the status information is the crack information, the current crack depth is extracted from the crack information, and the stress intensity factor amplitude of the bolt to be monitored is calculated based on the real-time load data in the real-time operating condition parameters. A functional relationship between the crack propagation rate and the stress intensity factor amplitude is established. Based on the functional relationship, the current crack depth is integrally calculated until the crack depth reaches the preset adjacent fracture depth. The remaining number of cycles is then obtained, and the remaining number of cycles is converted into the remaining service life. When the status information is the dynamic slip information, the current slip amount is obtained from the dynamic slip information, and the equivalent fretting wear stress of the contact surface of the bolt to be monitored is calculated based on the current slip value; Based on the equivalent fretting wear stress, find the corresponding theoretical total fatigue life number in the stress life curve; Retrieve historical load cycle statistics of the bolt to be monitored from a preset fault feature database. The historical load cycle statistics include the number of cycles executed under different stress levels. Calculate the ratio of the number of cycles executed to the corresponding limit number of cycles under each stress level, sum all ratios, and obtain the current cumulative fatigue damage degree of the bolt to be monitored; The remaining service life is obtained by multiplying the theoretical total fatigue life by 1 and subtracting the current cumulative fatigue damage.

6. The method according to claim 1, characterized in that, The process of acquiring instantaneous load data at the moment the state information is generated, extracting the corresponding fault response amplitude from the state information, and calculating the load sensitivity coefficient of the bolt to be monitored based on the instantaneous load data and the fault response amplitude specifically includes: The moment when the state is generated is obtained, and the corresponding real-time wind turbine torque data is matched in the real-time operating condition parameters based on the moment of generation. The real-time wind turbine torque data is used as the instantaneous load data. When the status information is the dynamic slip information, the peak-to-peak value of the displacement is extracted from the real-time displacement value associated with the dynamic slip information and used as the fault response amplitude; When the state information is the crack information, the root mean square value of vibration is extracted from the vibration characteristic value associated with the crack information and used as the fault response amplitude. Calculate the ratio of the fault response amplitude to the instantaneous load data to obtain the current load response ratio; The load sensitivity coefficient is obtained by dividing the difference between the current load response ratio and the reference load response ratio by the reference load response ratio.

7. The method according to claim 6, characterized in that, The early warning levels include a Level 1 early warning level and a Level 2 early warning level. The acquisition of wind resource forecast data, and the generation of early warning levels and comprehensive maintenance plans based on the remaining service life, the load sensitivity coefficient, the status information, and the wind resource forecast data, specifically include: Extract the maximum predicted wind speed within a future preset time window from the wind resource prediction data; The basic health status is determined based on the status information and the remaining service life. When the status information is the crack information or the remaining service life is less than or equal to the service life threshold, the basic health status is determined to be a high-risk status. When the status information is the dynamic slip information and the remaining service life is greater than the service life threshold, the basic health status is determined to be an abnormal state. The future risk index is obtained by multiplying the maximum predicted wind speed by the load sensitivity coefficient. When the basic health status is the high-risk status, or the basic health status is the abnormal status and the future risk index is greater than or equal to the preset risk threshold, the first-level early warning level is generated. When the basic health status is in the abnormal state and the future risk index is less than the preset risk threshold, the level 2 early warning level is generated; If the bolt to be monitored is at the first-level warning level, a maintenance plan including a shutdown command and a bolt replacement strategy is generated; if the bolt to be monitored is at the second-level warning level, a maintenance plan including a bolt preload re-tightening strategy is generated.

8. A monitoring and early warning system based on ultrasonic coupling of wind turbine bolt vibration, characterized in that, The system includes a data acquisition unit, a compensation unit, an extraction unit, an identification unit, and an early warning unit. The acquisition unit receives vibration signals, ultrasonic signals, and displacement signals that are synchronously acquired from the bolt to be monitored. The compensation unit acquires the real-time operating condition parameters of the wind turbine, and performs compensation processing on the vibration signal, the ultrasonic signal and the displacement signal based on the real-time operating condition parameters to obtain the compensated vibration signal, the compensated ultrasonic signal and the compensated displacement signal. The extraction unit analyzes the compensated vibration signal and the compensated ultrasonic signal, and extracts vibration feature values ​​and ultrasonic feature values ​​respectively. Real-time displacement values ​​are extracted from the compensated displacement signal, and the vibration feature values, ultrasonic feature values, and real-time displacement values ​​are normalized to obtain a coupled feature vector. The identification unit inputs the coupled feature vector into a preset fault identification model for processing to obtain the state information of the bolt to be monitored. The state information includes no fault information, dynamic slippage information, and crack information. The early warning unit, when the status information is the dynamic slip information or the crack information, obtains the material parameters of the bolt to be monitored, and performs prediction calculations on the material parameters, the real-time operating condition parameters, and the dynamic slip information or the crack information to obtain the remaining service life of the bolt to be monitored. The instantaneous load data at the moment the status information is generated is obtained, and the corresponding fault response amplitude is extracted from the status information. The instantaneous load data and the fault response amplitude are calculated to obtain the load sensitivity coefficient of the bolt to be monitored. Wind resource prediction data is obtained, and an early warning level and a comprehensive maintenance plan are generated based on the remaining service life, the load sensitivity coefficient, the status information, and the wind resource prediction data. The early warning level and the comprehensive maintenance plan are sent to the operation and maintenance terminal.

9. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images