Wind turbine generator unit fault early warning method fusing multi-modal data

Abstract

The present application relates to the field of wind power generation technology, and particularly relates to a wind turbine unit fault early warning method fusing multi-modal data. The method comprises the following steps: for any moment, based on the cumulative temperature rise condition and load condition of the gearbox of the target wind turbine unit at the moment, the heat conduction delay time of the gearbox is determined, and then the diagnosis evaluation reference moment of the moment is determined, and the active power level of the target wind turbine unit before the diagnosis evaluation reference moment is calculated, and the vibration data discrete sequence of the target wind turbine unit before the diagnosis evaluation reference moment is constructed; the pure mechanical vibration level of the target wind turbine unit is calculated; based on the cumulative temperature rise condition of the gearbox of the target wind turbine unit at the moment, the pure mechanical vibration level of the target wind turbine unit and the active power level of the target wind turbine unit before the diagnosis evaluation reference moment, the comprehensive feature degradation index of the moment is calculated. The data internal cause alignment is realized.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and more specifically to a method for early warning of faults in wind turbine generator sets that integrates multimodal data. Background Technology

[0002] Because wind turbine generators operate in harsh natural environments for extended periods, components such as gearboxes in their main drivetrain are prone to failure. Current technology relies primarily on the generator's data acquisition and monitoring control system to obtain minute-level operating condition data (including gearbox temperature and active power) and an independent condition monitoring system to acquire high-frequency vibration data. The extracted data from the same sampling time are then compared with corresponding threshold values ​​to determine if a gearbox failure has occurred.

[0003] Due to the limitations of the gearbox structure, the gearbox temperature sensor cannot be installed inside the gearbox and must be mounted on the outer casing. Since it takes time for the heat generated by metal friction inside the gearbox to conduct to the temperature sensor on the outer casing, the currently observed temperature rise is actually a result of past work load. However, the current high-frequency vibration data and active power data do not require this transmission time; they are essentially the result of the current operation of the wind turbine generator. Therefore, the existing method of triggering early warnings using data from various points at the same time suffers from data misalignment, leading to low warning accuracy. Summary of the Invention

[0004] To address the technical problem of low early warning accuracy caused by data misalignment, the present invention aims to provide a method for early warning of wind turbine generator faults by fusing multimodal data. The specific technical solution adopted is as follows: This invention provides a method for early warning of faults in wind turbine generator sets by fusing multimodal data, comprising the following steps: Calculate the comprehensive characteristic degradation index of the target wind turbine at various times; calculate the comprehensive degradation characteristic development rate; determine the state of the target wind turbine based on the comprehensive degradation characteristic development rate, and provide fault warning; Methods for calculating the overall characteristic degradation index at various time points include: For any given moment, based on the cumulative temperature rise and load conditions of the gearbox of the target wind turbine generator set at that moment, the heat conduction delay time of the gearbox is determined; based on the heat conduction delay time of the gearbox, the diagnostic evaluation reference moment is determined, and the active power level of the target wind turbine generator set before the diagnostic evaluation reference moment is calculated, and a discrete sequence of vibration data of the target wind turbine generator set before the diagnostic evaluation reference moment is constructed. Based on the discrete sequence of vibration data of the target wind turbine generator set before the diagnostic evaluation reference time, the pure mechanical vibration level of the target wind turbine generator set is calculated. Based on the cumulative temperature rise of the gearbox of the target wind turbine at that moment, the pure mechanical vibration level of the target wind turbine, and the active power level of the target wind turbine before the diagnostic assessment reference time, the comprehensive characteristic degradation index at that moment is calculated.

[0005] Among some possible implementations, methods for calculating the pure mechanical vibration level of the target wind turbine generator include: Based on the geographical location and wind field data of all wind turbine units in the wind power plant, the upstream units of the target wind turbine unit and their time in the same wind field are determined. Based on the active power discrete sequence of the upstream units before their time in the same wind field, the normal wind vibration prediction value of each sampling point is iteratively calculated. The vibration data of each sampling point in the vibration data discrete sequence of the target wind turbine unit before the diagnostic evaluation reference time is subtracted from the normal wind vibration prediction value of that sampling point to obtain the pure mechanical vibration value of each sampling point. Based on the pure mechanical vibration values ​​at each sampling point, the pure mechanical vibration level of the target wind turbine generator is calculated.

[0006] Among some possible implementations, methods for iteratively calculating the predicted values ​​of normal wind-induced vibrations at each sampling point include: 1) Determine the transfer function weight vector and target sampling points for this iteration; 2) Based on any sampling point in the discrete sequence of active power of the upstream unit before the same wind field time, construct the sliding time window vector of the sampling point and determine the transfer function weight vector of this iteration; calculate the inner product of the sliding time window vector and the transfer function weight vector to obtain the normal wind vibration prediction value of the sampling point; 3) Based on the pure mechanical vibration value, sliding window vector and transfer function weight vector of the sampling point, obtain the transfer function weight vector for the next iteration, and take the next sampling point as the target sampling point, iteratively execute step 1) until the normal wind vibration prediction value of all sampling points is obtained.

[0007] Among some possible implementations, methods for calculating the transfer function weight vector for the next iteration include: The effective step size is obtained by the ratio of the magnitude of the sliding window vector at the sampling point to the pure mechanical vibration value at the sampling point. The sliding window vector of the sampling point is scaled using the effective step size, and then superimposed with the transfer function weight vector of the current iteration to obtain the transfer function weight vector of the next iteration.

[0008] Among some possible implementations, methods for determining the transfer function weight vector for this iteration include: If this iteration is the first iteration, a long sequence of historical data during the fault-free stable period of the target wind turbine generator is input into the adaptive filter. After the operation converges, the transfer function weight vector of this iteration is obtained. If this iteration is not the first iteration, then the transfer function weight vector generated in the previous iteration will be used as the transfer function weight vector for this iteration.

[0009] Among some possible implementations, methods for constructing a discrete sequence of vibration data for the target wind turbine prior to the diagnostic assessment reference time include: High-frequency vibration data of the target wind turbine generator set before the diagnostic evaluation reference time is obtained, and the data is divided into multiple segments. The mean value of the high-frequency vibration data of each segment is calculated to obtain a discrete sequence of vibration data.

[0010] Among some possible implementations, methods for calculating the overall characteristic degradation index at this moment include: The sum of the pure mechanical vibration level and the active power level of the target wind turbine generator set is combined with the cumulative temperature rise of the gearbox to obtain the comprehensive characteristic degradation index at that moment.

[0011] Among some possible implementations, methods for calculating the rate of evolution of integrated degradation characteristics include: The comprehensive characteristic degradation index of the target wind turbine at various times is fitted with a straight line, and the slope of the fitted line is calculated to obtain the comprehensive degradation characteristic development rate.

[0012] Some possible implementations also include the following steps: If the target wind turbine is in a state of continuous deterioration, the remaining operating time before the target wind turbine reaches its maximum deterioration level is calculated based on the development rate of the comprehensive deterioration characteristics. Then, the moment when the target wind turbine reaches its maximum deterioration level is determined. Based on the wind speed prediction data before the moment when the target wind turbine reaches its maximum deterioration level, the downtime for maintenance of the target wind turbine is determined.

[0013] Among some possible implementations, methods for calculating the remaining operating time of the target wind turbine to reach its maximum degradation level include: Calculate the difference between the maximum comprehensive feature degradation index threshold and the comprehensive feature degradation index at the current moment, and divide it by the comprehensive feature degradation development rate to obtain the theoretical remaining running time to reach the maximum degradation level. The remaining operating time of the target wind turbine generator to reach its maximum deterioration level is obtained by combining the theoretical remaining operating time to reach the maximum deterioration level and the preset reduction factor.

[0014] The present invention has the following beneficial effects: Based on the cumulative temperature rise and load conditions of the gearbox of the target wind turbine generator set at various times, the present invention determines the gearbox heat conduction delay time, and then determines the diagnostic evaluation benchmark time at each time based on the gearbox heat conduction delay time. For the gearbox temperature rise, data from each time moment is used, and for vibration data and active power data, data from the diagnostic evaluation benchmark time is used, instead of using multiple data at the same time moment. This achieves internal alignment of temperature rise, vibration data, and active power data (data changes are all caused by load transmission at the same time moment). Furthermore, the present invention uses multiple data at different times to calculate the comprehensive characteristic degradation index and calculate the comprehensive degradation characteristic development speed, thereby providing early warning, solving the technical problem of data misalignment in the prior art and improving the early warning accuracy. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for calculating the comprehensive feature degradation index at each time point before the current time, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the calculation of the pure mechanical vibration level in an embodiment of the present invention. Detailed Implementation

[0016] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0017] The specific process of the wind turbine generator fault early warning method integrating multimodal data of the present invention is as follows: Figure 1 As shown, it includes the following steps: Step 1: Data Acquisition.

[0018] The system acquires low-frequency monitoring data from the Supervisory Control and Data Acquisition (SCADA) system of the target wind turbine generator set, as well as high-frequency vibration data from each wind turbine generator set collected by the condition monitoring system. The low-frequency monitoring data includes gearbox temperature data and active power data for each wind turbine generator set, with a sampling period of 1 minute. In this embodiment, vibration data specifically uses vibration acceleration, hereinafter referred to as vibration data.

[0019] It should be noted that the gearbox temperature data is collected by a temperature sensor located on the outer casing. Therefore, abnormal frictional heat generated by the unit's drive train during operation requires a certain amount of time to be conducted to the location of the temperature sensor on the unit's outer casing before it can be detected.

[0020] Step 2: Calculate the comprehensive characteristic degradation index for each time point before the current time.

[0021] All times before the current time (including the current time) are relative to the current time. Taking the current time and the four preceding times (a total of five times) as an example, a comprehensive characteristic degradation index is calculated for each time before the current time (the sampling time of the data acquisition and monitoring control system), and the calculation method for the comprehensive characteristic degradation index at each time is the same. In this implementation, the current time is used as an example. Taking this as an example, the calculation method of the comprehensive characteristic degradation index at the current moment is explained, specifically including the following steps: The specific process of calculating the comprehensive characteristic degradation index of each time point before the current time point in this invention is as follows: Figure 1 As shown, it includes the following steps: 2.1) Calculate the cumulative temperature rise and active power level of the gearbox of the target wind turbine generator set at the current moment, and construct a discrete sequence of vibration data of the target wind turbine generator set.

[0022] 2.1.1) Obtain the gearbox temperature data within the first set time period before the current time, and calculate the difference between the final temperature and the starting temperature within the first set time period to obtain the cumulative temperature rise of the gearbox.

[0023] Due to the abnormal frictional heat generated by the unit's drive chain during operation, there is an inevitable time delay in the temperature sensor on the outer surface being conducted to the temperature. The temperature change currently observed actually reflects the mechanical load impact on the equipment at a certain point in the past.

[0024] In this embodiment, the first set time period is taken as 10 minutes. The first set time period before the current time is obtained. , Gearbox temperature data within [the data].

[0025] Theoretically, the final temperature refers to the final moment. Temperature, initial temperature refers to the initial time. The temperature can be measured directly, but because temperature sensors may generate extreme spikes in data due to single-point electromagnetic interference, directly calculating the temperature at these two moments will result in a cumulative temperature rise in the gearbox that deviates from the actual situation, affecting the accuracy of diagnostic warnings. To solve this problem, this embodiment obtains the final temperature based on the gearbox temperature distribution over a second set time period before the final moment, and the initial temperature based on the gearbox temperature distribution over a third set time period after the initial moment. In this embodiment, the gearbox temperature distribution specifically uses the average gearbox temperature, and the second set time period equals the third set time period, which equals 1 minute. In other embodiments, the second set time period may not be equal to the third set time period, as long as the duration of both the second and third set time periods is less than the duration of the first set time period and greater than the maximum duration of a single electromagnetic interference event (usually less than 1 second), thus reducing electromagnetic interference.

[0026] Based on the difference between the final temperature and the initial temperature, the cumulative temperature rise of the gearbox during the first set time period before the current moment is obtained.

[0027] 2.1.2) Based on the load conditions and temperature level of the target wind turbine generator set in the first set time period before the current moment, determine the gearbox heat conduction delay time at the current moment, and then determine the diagnostic evaluation benchmark time at the current moment.

[0028] In this embodiment, the load condition is the load ratio of active power, and the temperature level is the average temperature of the target wind turbine generator set within a first set time period before the current moment.

[0029] Due to the abnormal frictional heat generated by the unit's drive train during operation, there is an inevitable time delay in the temperature sensor on the outer casing. The temperature rise currently measured by the sensor (effect) is caused by the internal gear friction work and abnormal impacts at a certain point in the past (cause). To align cause and effect, the current temperature must be used to match and diagnose the vibration and active power at the reference time (past).

[0030] Because the gearbox heat conduction delay time varies under different temperature and load conditions—for example, the gearbox heat conduction delay time is longer in extremely cold environments with light loads than in high-temperature, heavy-load environments—it is impossible to determine the true gearbox temperature at the current moment by using the same gearbox heat conduction delay time to trace back. To solve this problem, this embodiment conducts experimental calibration under different temperature and load conditions (the proportion of active power to load) to obtain a gearbox heat conduction delay time mapping table under different temperature and load conditions.

[0031] Based on the active power data of the target wind turbine generator set, the load ratio data of the active power is calculated, and then the average load ratio of the active power of the target wind turbine generator set in the first set time period before the current time is calculated.

[0032] Based on the average temperature and average load ratio of active power during the first set time period before the current moment, the gearbox heat conduction delay time is found in the gearbox heat conduction delay time mapping table. .

[0033] Rewind the gearbox heat conduction delay time from the current moment. To obtain the diagnostic assessment baseline time at the current moment. .

[0034] Active power data and vibration data within a first set time period prior to the diagnostic assessment baseline are acquired and aligned according to timestamps. In this embodiment, the alignment method uses the Network Time Protocol (NTP) to perform absolute timestamp alignment calibration, aligning data with the same timestamps to ensure that the temperature heating results and the causes of vibration load are correlated in time.

[0035] 2.1.3) Based on the active power data and high-frequency vibration data within the first set time period before the diagnostic assessment benchmark time, generate a discrete sequence of vibration data of the target wind turbine generator set (hereinafter referred to as the generator set) within the first set time period before the diagnostic assessment benchmark time, and calculate the active power level of the generator set.

[0036] In this embodiment, since temperature and active power data are low-frequency data, while vibration data is high-frequency data, it is necessary to down-frequency the vibration data. The methods for down-frequency processing include: For the high-frequency vibration data within the first predetermined time period before the diagnostic assessment baseline time, it is divided into multiple segments according to the fourth predetermined time period. The root mean square (RMS) of the high-frequency vibration data in each segment is calculated to obtain the discrete RMS sequence of the machine vibration data. The fourth time period is set to 1 minute, which is equal to the sampling period of the data acquisition and monitoring control system, thereby completing the frequency reduction of high-frequency vibration data, and the sampling period after frequency reduction is equal to the sampling period of the data acquisition and monitoring control system.

[0037] In this embodiment, the root mean square discrete sequence of vibration data obtained by frequency reduction processing is the discrete sequence of vibration data of the machine within the first set time period before the diagnostic evaluation reference time.

[0038] In this embodiment, the active power level is quantitatively assessed using the root mean square (RMS) of active power. The method for calculating the active power level of the machine includes: calculating the RMS of active power data within a first set time period before the diagnostic assessment reference time to obtain the RMS of the machine's active power.

[0039] 2.2) Using the active power data of the upstream unit, the predicted values ​​of normal wind vibration at each moment of the machine are obtained. By subtracting the predicted values ​​of normal wind vibration at each moment from the discretized sequence of the machine's vibration data, the pure mechanical vibration level of the machine is obtained.

[0040] 2.2.1) Based on the geographical location and wind field data of all wind turbine generators in the wind power plant, determine the co-wind field time of the upstream generators, then construct a discrete sequence of active power for the upstream generators, and obtain high-frequency vibration data of the upstream generators within a first predetermined time period before the co-wind field time. Specific methods include: Since the wind field requires a certain travel time from the upstream turbine to the wind farm in the wind power plant, the wind field impact on the wind farm at the diagnostic assessment baseline time should be equal to the wind field impact on the upstream turbine at a more distant historical time (diagnostic assessment baseline time minus travel time). There is a travel time delay between the two. To eliminate this travel time delay, this embodiment calculates the time required for the wind field to travel from the upstream turbine to the wind farm and then backtracks to determine the same wind field time at the upstream turbine.

[0041] Calculate the average wind direction and average wind speed based on the wind field data (including wind direction and wind speed) within the first set time period before the current time.

[0042] Based on the geographical location and average wind speed of all wind turbines in the wind farm, the upstream turbine is selected from the windward sector (e.g., within a range of ±15 degrees of the average wind direction) and has the shortest spatial distance from the current turbine in the average wind direction. Specifically, if no other wind turbine meets the criteria within the sector (e.g., the current turbine is located in the first row of windward turbines under a specific wind direction), step 2.2.2 is skipped, and the discrete sequence of vibration data within the first set time period before the diagnostic assessment benchmark time obtained in step 2.1.3) is directly used as the pure mechanical vibration sequence of the current turbine. Based on this sequence, the pure mechanical vibration level of the current turbine is obtained and substituted into step 2.3) for calculation.

[0043] The calculation involves determining the time required for the wind farm to travel from the upstream turbine to the local turbine, and then backtracking to determine the exact time of the upstream turbine's arrival at the same wind farm location. Specifically, the calculation method for the time required for the wind farm to travel from the upstream turbine to the local turbine includes: In the formula, This refers to the time required for the wind farm to reach this unit from the upstream turbine. This represents the distance between the upstream unit and this unit in the average wind direction. The average wind speed, This represents the ideal time required for the wind farm to reach this unit from the upstream turbine. To comprehensively correct the empirical coefficients.

[0044] Because real wind fields have wake effects and topographical undulations, airflow propagation is not an ideal uniform linear motion. Therefore, this embodiment introduces an empirical coefficient for combined correction of topography and wake. To make corrections.

[0045] In this embodiment, if the average wind speed is less than the grid-connected cut-in wind speed threshold, it indicates that there is no wind-induced normal vibration. The discrete sequence of vibration data within the first set time period before the diagnostic evaluation benchmark time obtained in step 2.1.3) is the pure mechanical vibration sequence. At this time, it is only necessary to calculate the root mean square of the discrete sequence of vibration data and substitute it into step 2.3) for calculation. There is no need to execute step 2.2.2.

[0046] The grid connection cut-in wind speed threshold is the minimum wind speed that allows the wind turbine generator to operate, such as 3 m / s.

[0047] Tracing back from the diagnostic assessment baseline of this unit, we find the time required for the wind farm to reach this unit from the upstream turbine. The calculated wind field time of the upstream unit is obtained. The calculated wind field time is rounded and aligned with the nearest minute-level timestamp, thereby obtaining the active power data of the upstream unit within the first set time period before that timestamp, and obtaining the discrete sequence of active power of the upstream unit.

[0048] Since the vibration of the discrete sequence of the local vibration data consists of two parts: mechanical vibration and wind-induced normal vibration, directly using the root mean square discrete sequence of vibration data may cause misjudgment due to wind-induced normal vibration. In order to solve this problem, the wind-induced normal vibration component is filtered out through the following steps in this embodiment.

[0049] 2.2.2) Based on the discrete sequence of active power of the upstream unit, the wind-induced normal following vibration component is separated from the mixed vibration signal of the unit using the normalized least mean square (NLMS) adaptive filter, and the pure mechanical vibration level is calculated.

[0050] After spatiotemporal alignment, the active power sequence of the upstream unit is strongly correlated with normal wind-induced vibration (the greater the wind power, the stronger the forced vibration of the unit), but completely unrelated to internal mechanical damage vibration (cracks and pitting of components are not directly affected by upstream wind power). Based on this premise, the adaptive filter uses the active power sequence of the upstream unit as a reference input to automatically fit a linear mapping relationship between the active power of the upstream unit and the wind-induced vibration of the unit, accurately estimates the amplitude of pure wind-induced vibration, and then subtracts it from the mixed vibration. The remaining residual is the pure mechanical damage characteristic unaffected by wind conditions.

[0051] Least Mean Square (LMS) is the fundamental algorithm for adaptive filtering. It continuously adjusts the weights using the steepest descent method to minimize the mean square value of the prediction error. However, ordinary LMS has a core drawback: a fixed step size cannot simultaneously ensure convergence speed and stability when the input signal power fluctuates drastically. In the wind power generation scenario of this invention, the wind speed fluctuation range is extremely large (3~25m / s), and the corresponding active power fluctuation can reach tens of times. If a fixed step size LMS is used: when the wind is strong and the power is high, an excessively large step size will cause the algorithm to oscillate and diverge; when the wind is weak and the power is low, an excessively small step size will cause the weights to converge too slowly, resulting in a lag in tracking wind condition changes. NLMS (Normalized LMS) introduces an input power normalization term on top of LMS, which is equivalent to dynamically adjusting the effective step size according to the current input signal power: when the input power is high, the step size is automatically reduced; when the input power is low, the step size is automatically increased. This perfectly adapts to the wind power variable operating condition scenario, ensuring that the algorithm is both stable and can quickly track wind condition changes across the entire wind speed range.

[0052] The specific process for calculating the level of pure mechanical vibration is as follows: Figure 2 As shown, it includes the following steps: Based on the active power discrete sequence of the upstream unit within the first set time period before its time in the same wind field, each sampling point (one active power value corresponds to one sampling point) is taken as the target sampling point in sequence according to the sampling point order. The normal wind vibration prediction value of each sampling point is calculated iteratively. The vibration data of each sampling point in the vibration data discrete sequence of the target wind turbine generator before the diagnostic evaluation reference time is subtracted from the normal wind vibration prediction value of that sampling point to obtain the pure mechanical vibration value of each sampling point. Then, based on the pure mechanical vibration value of each sampling point, the pure mechanical vibration level of the target wind turbine generator is calculated.

[0053] The methods for iteratively calculating the predicted values ​​of normal wind-induced vibration at each sampling point include: (1) Determine the transfer function weight vector for this iteration.

[0054] Determine whether this iteration is the first iteration globally. If it is, extract a long sequence of historical data from the machine's fault-free stable period as test data and input it into the adaptive filter for warm-up and initialization. After the operation converges, generate the initial transfer function weight vector. If it is not the first iteration globally, extract the transfer function weight vector obtained at the end of the previous iteration and use it as the transfer function weight vector for this iteration.

[0055] It should be noted that the dimension of the transfer function weight vector is equal to the order of the adaptive filter. The order of the adaptive filter is greater than 1 and less than the number of sampling points in the discrete sequence of active power within the first set time period. In this embodiment, the first set time period is 10 minutes, and the sampling interval of active power is 1 minute. Therefore, the number of sampling points in the discrete sequence of active power of the upstream unit within the first set time period before its wind field time and the number of sampling points in the discrete sequence of vibration data of the local unit within the first set time period before the diagnostic evaluation reference time are both 10. Therefore, the order of the adaptive filter is greater than 1 and less than or equal to 10.

[0056] Since there are 10 sampling points, 10 iterations are required to calculate the comprehensive feature degradation index at each time point. For each time point before the current time point (a total of 5 time points in this implementation), a total of 50 iterations are required. The first iteration of these 50 iterations is called the global first iteration.

[0057] (2) Calculate the predicted value of normal wind vibration and the value of pure mechanical vibration at the target sampling point.

[0058] The k-th sampling point is used as the target sampling point for explanation. The specific calculation method includes: Based on the discrete sequence of active power of upstream units, a sliding time window vector of the target sampling point is constructed. The inner product operation of the sliding time window vector and the transfer function weight vector is performed to obtain the normal wind vibration prediction value of the target sampling point. The vibration data of the target sampling point (also the kth sampling point) in the discrete sequence of local vibration data is subtracted from the normal wind vibration prediction value of each sampling point to obtain the pure mechanical vibration value of the sampling point.

[0059] (3) The transfer function weight vector is updated based on the sliding window vector, transfer function weight vector, and pure mechanical vibration value of the target sampling point (kth sampling point). Specific update methods include: In the formula, This is the updated transfer function weight vector for the k-th sampling point (i.e., the transfer function weight vector for the next iteration). Let the transfer function weight vector be the k-th sampling point. Let k be the sliding window vector for the k-th sampling point. For L2 norm operations, The pure mechanical vibration value at sampling point k. This is a preset fixed iteration compensation factor, with a classic value of 0.01. The preset zero-factor is a minimum value, such as 10 to the power of -6. This is the effective step size. The update method uses... The pure mechanical vibration values ​​at sampling point k are normalized, hence the name Normalized Least Mean Square Adaptive Filter. This filter is used in high wind and high power conditions. Larger values ​​can reduce the weight update step size, automatically shrinking the effective step size and preventing oscillations caused by excessive weight adjustments, thus ensuring algorithm stability; this is especially important in low wind and low power conditions. When the numerical value is small, the effective step size is automatically increased to accelerate the convergence speed of the weights and track changes in wind conditions in a timely manner.

[0060] In this embodiment, the method for constructing the sliding time window vector includes: for any sampling point, constructing a sliding time window vector for that sampling point based on a predetermined number of discrete active power data from upstream units preceding (including) that sampling point. The predetermined number is equal to the order of the adaptive filter. If the number of discrete active power data from upstream units preceding a certain sampling point is less than the predetermined number, then all data are padded with zeros so that the number of discrete active power data from upstream units preceding that sampling point equals the predetermined number after zero-padding.

[0061] (4) Repeat steps (1) to (3) in the order of sampling points to obtain the pure mechanical vibration values ​​of each sampling point.

[0062] Pure mechanical vibration values ​​at each sampling point Construct pure mechanical vibration sequences for each sampling point Then, the root mean square of the sequence is calculated to obtain the root mean square of the pure mechanical vibration. It is used to quantify the level of pure mechanical vibration.

[0063] 2.3) Based on the cumulative temperature rise of the gearbox in the first set time period before the current time, the active power level of the machine, and the root mean square of the pure mechanical vibration (pure mechanical vibration level), calculate the comprehensive characteristic deterioration index of the machine.

[0064] In this embodiment, the cumulative temperature rise of the gearbox within the first set time period before the current time is measured by the difference between the final temperature and the initial temperature. The calculation method for the comprehensive characteristic deterioration index of this machine includes: In the formula, The overall characteristic degradation index of this machine, This refers to the active power level of the machine. The difference between the final temperature and the initial temperature. The preset temperature fluctuation threshold, To perform the maximum value operation, This indicates normalization processing. In this embodiment, maximum and minimum value normalization processing is used. During the maximum and minimum value normalization process, the maximum and minimum values ​​are respectively adopted from the target wind turbine generator's historical fault-free healthy operation period. The maximum and minimum values, for During the maximum-minimum normalization process, the maximum value is taken as the maximum allowable temperature rise of the target wind turbine generator set in the gearbox design and the maximum value of the minimum-minimum-maxim ... The maximum value of the difference (its value is greater than 0) and the minimum value is 0.

[0065] The formula shows that when the unit is under high power heavy load output, even if there are only minor early mechanical abnormalities and minor abnormal temperature rises inside ( (The value is relatively small), but due to the level of active power With its extremely high weight, the system will significantly amplify the danger level of the potential hazard, effectively solving the problem of missed detection in the early stages of heavy load using the threshold method. Conversely, when the unit is stopped and idling or under light load, the multiplicative weight is extremely small, thereby greatly reducing any abnormal fluctuations caused by the electrical noise of the sensor and eliminating false alarms.

[0066] For each moment within the set time period before the current moment, repeat step two to obtain the comprehensive characteristic degradation index for each moment before the current moment.

[0067] Step 3: Based on the comprehensive characteristic degradation index of each time point before the current time, calculate the comprehensive degradation characteristic development speed, determine whether the machine is in a state of continuous deterioration, and if it is in a state of continuous deterioration, calculate the remaining running time to reach the maximum degree of degradation, and then determine the downtime for maintenance.

[0068] Based on the comprehensive feature degradation index at each time point before the current time, the least squares method is used to fit a straight line to obtain the change line of the comprehensive feature degradation index before the current time. The slope of the line is then calculated to obtain the development speed of the comprehensive degradation feature.

[0069] If the rate of development of the overall degradation characteristics is greater than 0, it indicates that the overall degradation index of the machine is continuously increasing. In this case, the machine is determined to be in a state of continuous deterioration and a fault warning will be issued. Otherwise, the machine is determined not to be in a state of continuous deterioration and no fault warning is required.

[0070] In this embodiment, the method for calculating the remaining runtime to reach the maximum degradation level includes: In the formula, To achieve the maximum remaining runtime degradation, The threshold for the maximum comprehensive feature degradation index. This represents the overall characteristic degradation index at the current moment. To comprehensively assess the rate of development of degradation characteristics, The theoretical remaining runtime to achieve the maximum degradation level, This is a preset reduction factor, with a value less than 1, used to reduce the remaining running time to allow for sufficient safety margin. The smaller the value, the larger the safety margin. In this embodiment... .

[0071] In this embodiment, the method for determining the downtime for maintenance includes: Based on the current moment and the remaining operating time before reaching maximum degradation, the time when the wind turbine reaches its maximum degradation is determined. Wind speed forecast data from the current moment to this point is obtained, and a time window with an average wind speed lower than the cut-in wind speed threshold is identified. This time window is used as the shutdown and maintenance time. If no time window with wind speed lower than the cut-in wind speed threshold exists before the wind turbine reaches its maximum degradation, then the time window with the minimum average wind speed is identified as the shutdown and maintenance time. This avoids shutdown and maintenance during periods of high wind speed and high power generation, which would negatively impact the economic efficiency of the wind turbine.

Claims

1. A method for early warning of wind turbine generator faults by integrating multimodal data, characterized in that, Includes the following steps: Calculate the comprehensive characteristic degradation index of the target wind turbine at various times; Calculate the rate of development of overall degradation characteristics; The state of the target wind turbine is determined based on the rate of development of comprehensive deterioration characteristics, and fault warnings are issued. Methods for calculating the overall characteristic degradation index at various time points include: For any given moment, based on the cumulative temperature rise and load conditions of the gearbox of the target wind turbine generator set at that moment, the heat conduction delay time of the gearbox is determined; based on the heat conduction delay time of the gearbox, the diagnostic evaluation reference moment is determined, and the active power level of the target wind turbine generator set before the diagnostic evaluation reference moment is calculated, and a discrete sequence of vibration data of the target wind turbine generator set before the diagnostic evaluation reference moment is constructed. Based on the discrete sequence of vibration data of the target wind turbine generator set before the diagnostic evaluation reference time, the pure mechanical vibration level of the target wind turbine generator set is calculated. Based on the cumulative temperature rise of the gearbox of the target wind turbine at that moment, the pure mechanical vibration level of the target wind turbine, and the active power level of the target wind turbine before the diagnostic assessment reference time, the comprehensive characteristic degradation index at that moment is calculated.

2. The wind turbine generator fault early warning method according to claim 1, characterized in that, Methods for calculating the pure mechanical vibration level of a target wind turbine include: Based on the geographical location and wind field data of all wind turbine units in the wind power plant, the upstream units of the target wind turbine unit and their time in the same wind field are determined. Based on the active power discrete sequence of the upstream units before their time in the same wind field, the normal wind vibration prediction value of each sampling point is iteratively calculated. The vibration data of each sampling point in the vibration data discrete sequence of the target wind turbine unit before the diagnostic evaluation reference time is subtracted from the normal wind vibration prediction value of that sampling point to obtain the pure mechanical vibration value of each sampling point. Based on the pure mechanical vibration values ​​at each sampling point, the pure mechanical vibration level of the target wind turbine generator is calculated.

3. The wind turbine generator fault early warning method based on multimodal data according to claim 2, characterized in that, The methods for iteratively calculating the predicted values ​​of normal wind-induced vibrations at each sampling point include: 1) Determine the transfer function weight vector and target sampling points for this iteration; 2) Based on any sampling point in the discrete sequence of active power of the upstream unit before the same wind field time, construct the sliding time window vector of the sampling point and determine the transfer function weight vector of this iteration; calculate the inner product of the sliding time window vector and the transfer function weight vector to obtain the normal wind vibration prediction value of the sampling point; 3) Based on the pure mechanical vibration value, sliding window vector and transfer function weight vector of the sampling point, obtain the transfer function weight vector for the next iteration, and take the next sampling point as the target sampling point, iteratively execute step 1) until the normal wind vibration prediction value of all sampling points is obtained.

4. The wind turbine generator fault early warning method based on multimodal data according to claim 3, characterized in that, Methods for calculating the transfer function weight vector for the next iteration include: The effective step size is obtained by the ratio of the magnitude of the sliding window vector at the sampling point to the pure mechanical vibration value at the sampling point. The sliding window vector of the sampling point is scaled using the effective step size, and then superimposed with the transfer function weight vector of the current iteration to obtain the transfer function weight vector of the next iteration.

5. The wind turbine generator fault early warning method according to claim 3, characterized in that, Methods for determining the transfer function weight vector for this iteration include: If this iteration is the first iteration, a long sequence of historical data during the fault-free stable period of the target wind turbine generator is input into the adaptive filter. After the operation converges, the transfer function weight vector of this iteration is obtained. If this iteration is not the first iteration, then the transfer function weight vector generated in the previous iteration will be used as the transfer function weight vector for this iteration.

6. The method for early warning of wind turbine generator faults by fusing multimodal data according to claim 1, characterized in that, Methods for constructing a discrete sequence of vibration data for a target wind turbine generator prior to the diagnostic assessment reference time include: High-frequency vibration data of the target wind turbine generator set before the diagnostic evaluation reference time is obtained, and the data is divided into multiple segments. The mean value of the high-frequency vibration data of each segment is calculated to obtain a discrete sequence of vibration data.

7. The method for early warning of wind turbine generator faults by fusing multimodal data according to claim 1, characterized in that, Methods for calculating the overall characteristic degradation index at this moment include: The sum of the pure mechanical vibration level and the active power level of the target wind turbine generator set is combined with the cumulative temperature rise of the gearbox to obtain the comprehensive characteristic degradation index at that moment.

8. The method for early warning of wind turbine generator faults by fusing multimodal data according to claim 1, characterized in that, Methods for calculating the rate of development of overall degradation characteristics include: The comprehensive characteristic degradation index of the target wind turbine at various times is fitted with a straight line, and the slope of the fitted line is calculated to obtain the comprehensive degradation characteristic development rate.

9. The method for early warning of wind turbine generator faults by fusing multimodal data according to claim 1, characterized in that, It also includes the following steps: If the target wind turbine is in a state of continuous deterioration, the remaining operating time before the target wind turbine reaches its maximum deterioration level is calculated based on the development rate of the comprehensive deterioration characteristics. Then, the moment when the target wind turbine reaches its maximum deterioration level is determined. Based on the wind speed prediction data before the moment when the target wind turbine reaches its maximum deterioration level, the downtime for maintenance of the target wind turbine is determined.

10. The method for early warning of wind turbine generator faults by fusing multimodal data according to claim 9, characterized in that, Methods for calculating the remaining operating time of a target wind turbine to reach its maximum degradation level include: Calculate the difference between the maximum comprehensive feature degradation index threshold and the comprehensive feature degradation index at the current moment, and divide it by the comprehensive feature degradation development rate to obtain the theoretical remaining running time to reach the maximum degradation level. The remaining operating time of the target wind turbine generator to reach its maximum deterioration level is obtained by combining the theoretical remaining operating time to reach the maximum deterioration level and the preset reduction factor.

Images