Fan fault early warning and diagnosis method fusing multi-modal time series data

Abstract

The invention relates to the technical field of wind driven generator monitoring, in particular to a fan fault early warning and diagnosis method fusing multi-modal time sequence data, which comprises the following steps: respectively acquiring vibration time sequence data, current time sequence data and temperature time sequence data of a fan, and respectively calculating a first feature sequence, a second feature sequence and a third feature sequence; generating a fault early warning signal based on a three-state polarity synchronous counting mechanism; calculating a dispersion statistical value, and executing chronic fault diagnosis; when the fan is in a low-load steady-state working condition, active lubrication state detection and feedback are executed; when the first feature sequence is calculated, synchronously executing adjoint oil product pollution diagnosis; and the environment temperature is judged through the temperature time sequence data, and low-temperature working condition adaptability compensation is executed under the low-temperature working condition. The main purpose of the invention is to solve the problems in the prior art that early fault signs cannot be effectively captured due to information loss during fusion processing of multi-modal data, and real faults and external interference are difficult to distinguish.

Description

Technical Field

[0001] This invention relates to the field of wind turbine monitoring technology, and in particular to a method for early warning and diagnosis of wind turbine faults that integrates multimodal time-series data. Background Technology

[0002] In the field of wind power generation technology, in order to ensure the safe and stable operation of large wind turbine generators, monitoring systems including various types of sensors such as vibration, current and temperature are usually used to continuously track the health status of key components such as gearbox bearings. A common practice is to use data fusion technology to comprehensively analyze the sensor data from different sources and of different natures in order to obtain an assessment that can comprehensively reflect the overall health status of the unit.

[0003] However, when it comes to identifying early, latent faults that evolve slowly and may lead to downtime losses, this approach reveals a limitation inherent in its analytical methods. During wind turbine operation, vibration signal changes occur on the order of milliseconds, while changes in thermodynamic parameters such as temperature occur on the order of seconds or even minutes. There is a huge difference in the time scale between the two. In order to achieve data fusion, existing technologies have to use downsampling or time window averaging to time-align high-frequency vibration data with low-frequency temperature data. This data preprocessing comes at the cost of actively and irreversibly erasing the transient change information contained in the high-frequency signal that can precisely characterize early, minor damage.

[0004] Meanwhile, most existing early warning systems rely on static threshold judgments based on the absolute or statistical values ​​of various physical quantities. However, many early faults do not show significant changes in the absolute values ​​of various physical quantities during their nascent stages. Instead, they manifest as a coordinated disorder of the changing trends of multiple physical quantities over time. For example, a heating process caused by lubrication deterioration will result in an abnormally sudden increase in the disorder of the vibration signal, while a healthy heating process will show a relatively smooth increase in vibration intensity. Existing technologies lack the ability to measure the coordination of such cross-modal changing trends, thus failing to capture this early signal that marks the transition of the system from a healthy to an abnormal state. Specifically, the shortcomings of existing technologies are mainly: the time alignment processing of high-frequency signals for data fusion leads to the removal of transient information that can characterize early faults; and the static threshold alarm logic based on the absolute values ​​of physical quantities has a blind spot in perceiving the coordinated disorder of multimodal changing trends in nascent faults.

[0005] Therefore, how to effectively identify the coordinated disorder among the changing trends of multimodal physical quantities while preserving the original temporal resolution and complete information of each modal data, and reliably distinguish early warning methods that originate from mechanical faults inside the equipment and interference from the external environment. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for early warning and diagnosis of wind turbine faults by fusing multimodal time-series data. Based on the wind turbine's vibration time-series data, current time-series data, and temperature time-series data, this invention calculates a first feature sequence, a second feature sequence, and a third feature sequence, and generates a fault warning signal based on a three-state polarity synchronous counting mechanism. The main objective of this invention is to solve the problems of existing technologies failing to effectively capture early fault symptoms due to information loss when fusing multimodal data, and the difficulty in distinguishing between actual faults and external interference.

[0007] The technical means employed in this invention are as follows: A method for early warning and diagnosis of wind turbine faults integrating multimodal time-series data includes: acquiring vibration time-series data, current time-series data, and temperature time-series data of the wind turbine, and calculating a first characteristic sequence, a second characteristic sequence, and a third characteristic sequence respectively; generating a fault warning signal based on a three-state polarity synchronous counting mechanism; calculating the discreteness statistics and performing chronic fault diagnosis; performing active lubrication status detection and feedback when the wind turbine is under low-load steady-state operating conditions; simultaneously performing accompanying oil contamination diagnosis while calculating the first characteristic sequence; and using temperature time-series data to determine the ambient temperature and performing low-temperature adaptive compensation under low-temperature operating conditions.

[0008] Furthermore, the calculation of the first feature sequence, the second feature sequence, and the third feature sequence specifically includes: Based on the vibration time series data, a first feature sequence is calculated to characterize the disorder of the vibration time series data; based on the current time series data, a second feature sequence is calculated to characterize the deviation of the current time series data; based on the temperature time series data, a third feature sequence is calculated to characterize the changing trend of the temperature time series data.

[0009] Furthermore, the first feature sequence is obtained by calculating the Shannon entropy from the amplitude distribution of the vibration time series data; the second feature sequence is obtained by calculating the KL divergence between the current time series data and the healthy current template; and the third feature sequence is obtained by calculating the temperature gradient sequence by calculating the difference between the current temperature and the temperature at the previous time point.

[0010] Furthermore, the generation of fault warning signals based on the three-state polarity synchronous counting mechanism specifically includes: Within one calculation cycle, the incremental polarity of the first feature sequence, the second feature sequence, and the third feature sequence is determined; before accumulating a synchronous count value, a spectral symmetry check is performed, a spectral analysis is performed on the current vibration time series data, and the symmetry of its spectral energy distribution on both sides of the center frequency is calculated. When the incremental polarity of the first feature sequence, the second feature sequence, and the third feature sequence are all positive, and the symmetry of the spectral energy distribution satisfies the symmetry condition, the synchronization count value is accumulated; when the synchronization count value exceeds the warning threshold within a time window, a fault warning signal is generated.

[0011] Furthermore, the symmetry condition for the spectral symmetry check is specifically: the ratio of the total energy of the spectrum in the left half-band to the total energy in the right half-band of the center frequency falls within a predetermined interval containing the value 1.

[0012] Furthermore, the warning threshold includes a first warning threshold and a second warning threshold, wherein the second warning threshold is greater than the first warning threshold. When the synchronization count value exceeds the first warning threshold, a first-level warning signal is generated, and when the synchronization count value exceeds the second warning threshold, a second-level warning signal is generated.

[0013] Furthermore, the calculation of the dispersion statistics and the performance of chronic fault diagnosis specifically include: Within a long-period sliding window, a time series consisting of synchronization count values ​​is recorded, and the standard deviation of the time series is calculated as the dispersion statistic. When the dispersion statistic is lower than the chronic fault threshold, an independent chronic fault warning signal, distinct from the fault warning signal, is generated.

[0014] Furthermore, the active lubrication status detection and feedback when the wind turbine is under low-load steady-state operation specifically includes: When the wind turbine is in a low-load steady-state condition, a standardized micro-disturbance current signal is injected into the motor by controlling the converter. During the injection, the current signal is collected and the phase difference between the fundamental wave and the higher harmonics is calculated. When the instantaneous jump of the phase difference exceeds the phase difference threshold, the weight W of the synchronization count value in the accumulation operation is temporarily increased, where W>1.

[0015] Furthermore, the simultaneous execution of accompanying oil contamination diagnosis during the calculation of the first feature sequence specifically includes: In the process of processing vibration time series data to calculate the first feature sequence, high-frequency residual signals are separated; and based on the impact pulse events in the high-frequency residual signals, the time delay entropy of the time interval distribution between adjacent impact pulse events is calculated; when the time delay entropy exceeds the oil contamination threshold, a maintenance suggestion signal indicating abnormal gearbox lubrication status is generated.

[0016] Furthermore, the step of using temperature time-series data to determine the ambient temperature and performing low-temperature adaptive compensation under low-temperature conditions specifically includes: When the ambient temperature is detected to be lower than the low temperature threshold, the standard deviation of the fundamental phase of the current time series data within the first time period is calculated; when the standard deviation is less than the phase stability threshold, the zero reference used to determine that the polarity of the first characteristic sequence increment is positive is adjusted to a negative reference.

[0017] Compared with the prior art, the present invention has the following advantages: The wind turbine fault early warning and diagnosis method provided by this invention, which integrates multimodal time-series data, establishes an early warning mode that directly captures the collaborative disorder process of multimodal data. The system processes three non-time-aligned data streams of vibration, current and temperature in parallel, and extracts three independent dynamic features from each data stream: disorder degree, deviation degree and temperature change trend. By judging whether the incremental direction of these three features shows a synchronous positive correlation in time, the early physical process of the system evolving from a healthy state to an abnormal state is identified. The early warning mechanism thus shifts from judging the absolute threshold of a single physical quantity after the fact to identifying the dynamic process of system degradation.

[0018] This invention, by analyzing the symmetry of the vibration spectrum, endows the early warning logic with the ability to distinguish the root cause of the fault. Since there are essential differences in the spectral distribution of vibration energy between internal mechanical faults and external mass imbalances, the system immediately verifies the symmetry of the vibration spectrum while capturing the coordinated disorder signal, and identifies the symmetrical spectral energy distribution as a manifestation of external mass imbalance. Thus, without relying on any additional sensors, the early warning mechanism has inherent immunity to interference caused by external environmental factors such as blade icing.

[0019] This invention opens up a diagnostic dimension for fault evolution dynamics by performing secondary analysis on the dispersion of the synchronization count value output by the early warning mechanism over a long period. A system entering a deterministic degradation channel will see its risk count increase shift from a healthy, intermittent, and irregular state to an abnormally smooth, low-dispersion state. By capturing this change in process morphology, the system gains the ability to identify extremely slow-evolving, chronic, progressive faults, achieving coverage of sudden risks. Furthermore, it extends the system's monitoring capabilities from passive state monitoring to active state detection. Under specific operating conditions, the system injects micro-disturbance current into the motor and analyzes its harmonic phase response to inversely determine the latent state of bearing lubrication. When a pre-wear risk is detected, the system does not directly trigger an alarm but instead increases the cumulative weight of the main early warning mechanism, making it more sensitive to subsequent coordinated disorder events, enabling the early warning system to self-optimize based on internal detection results. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the wind turbine fault early warning and diagnosis method that integrates multimodal time series data in this invention.

[0022] Figure 2 This is a schematic diagram illustrating the state transition from healthy to multi-level early warning in this invention.

[0023] Figure 3 This is a deployment architecture diagram of the system at the wind turbine site and the remote center in an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0028] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore should not be construed as limiting the scope of protection of this invention.

[0029] like Figure 1 As shown, this invention provides a method for early warning and diagnosis of wind turbine faults by integrating multimodal time-series data, including: acquiring vibration time-series data, current time-series data, and temperature time-series data of the wind turbine, and calculating a first feature sequence, a second feature sequence, and a third feature sequence respectively; in a preferred embodiment of this invention, based on the vibration time-series data, a first feature sequence is calculated to characterize the disorder of the vibration time-series data; based on the current time-series data, a second feature sequence is calculated to characterize the deviation of the current time-series data; and based on the temperature time-series data, a third feature sequence is calculated to characterize the changing trend of the temperature time-series data.

[0030] In a preferred embodiment of the present invention, the first feature sequence is obtained by calculating the Shannon entropy from the amplitude distribution of vibration time-series data; the second feature sequence is obtained by calculating the KL divergence between current time-series data and a healthy current template; and the third feature sequence is obtained by calculating the temperature gradient sequence by calculating the difference between the current temperature and the temperature at a previous time point. The increment polarity of the third feature sequence is determined to be positive, specifically by determining that the value of the temperature gradient sequence is greater than zero.

[0031] A fault warning signal is generated based on a three-state polarity synchronous counting mechanism. In a preferred embodiment of the present invention, the incremental polarity of the first feature sequence, the second feature sequence, and the third feature sequence is determined within a calculation cycle. Before accumulating a synchronous count value, a spectral symmetry check is performed, the current vibration time series data is subjected to spectral analysis, and the symmetry of its spectral energy distribution on both sides of the center frequency is calculated.

[0032] When the incremental polarity of the first feature sequence, the second feature sequence, and the third feature sequence are all positive, and the symmetry of the spectral energy distribution satisfies the symmetry condition, the synchronization count value is accumulated; when the synchronization count value exceeds the warning threshold within a time window, a fault warning signal is generated.

[0033] In a specific implementation, as a preferred embodiment of the present invention, the symmetry condition for the spectral symmetry verification is as follows: the ratio of the total energy of the spectral energy in the left half-band of the center frequency to the total energy in the right half-band falls within a predetermined interval containing the value 1.

[0034] In a preferred embodiment of this invention, the warning threshold includes a first warning threshold and a second warning threshold, wherein the second warning threshold is greater than the first warning threshold. A first-level warning signal is generated when the synchronization count value exceeds the first warning threshold, and a second-level warning signal is generated when the synchronization count value exceeds the second warning threshold. In practice, 100 milliseconds is used; the time window is one minute.

[0035] The dispersion statistics are calculated, and chronic fault diagnosis is performed. In a preferred embodiment of the present invention, a time series consisting of synchronization counts is recorded within a long-period sliding window, and the standard deviation of the time series is calculated as the dispersion statistics. When the dispersion statistics are lower than the chronic fault threshold, an independent chronic fault warning signal, which is different from the fault warning signal, is generated.

[0036] When the wind turbine is under low load steady-state operation, active lubrication status detection and feedback are performed. In a preferred embodiment of the present invention, when the wind turbine is under low load steady-state operation, a standardized micro-disturbance current signal is injected into the motor by controlling the converter. During the injection, the current signal is collected and the phase difference between the fundamental wave and the higher harmonics is calculated. When the instantaneous jump of the phase difference exceeds the phase difference threshold, the weight W of the synchronization count value in the accumulation operation is temporarily increased, where W>1.

[0037] When calculating the first feature sequence, a concurrent oil contamination diagnosis is performed simultaneously. In a preferred embodiment of the present invention, during the processing of vibration time series data to calculate the first feature sequence, a high-frequency residual signal is separated. Based on the impact pulse events in the high-frequency residual signal, the time delay entropy of the time interval distribution between adjacent impact pulse events is calculated. When the time delay entropy exceeds the oil contamination threshold, a maintenance recommendation signal indicating abnormal gearbox lubrication status is generated.

[0038] The ambient temperature is determined using temperature time-series data, and low-temperature adaptive compensation is performed under low-temperature conditions. In a preferred embodiment of the invention, when the ambient temperature is detected to be below the low-temperature threshold, the standard deviation of the fundamental phase of the current time-series data within the first time period is calculated; when the standard deviation is less than the phase stability threshold, the zero reference used to determine that the polarity of the first characteristic sequence increment is positive is adjusted to a negative reference.

[0039] Example 1 like Figure 3 As shown, this invention provides a system for wind turbine fault early warning and diagnosis based on a method that integrates multimodal time-series data, deployed at both the wind turbine field and a remote control center. The edge computing unit is an industrial-grade microcontroller, and the entire three-state polarity synchronous counting mechanism is executed locally on the edge computing unit.

[0040] The overall architecture is built on an edge computing unit deployed inside the wind turbine generator. This unit is directly coupled to the wind turbine's existing vibration, current, and temperature sensors. Its task is to process these three raw time-series data streams, which differ in time scale, in parallel, and execute a three-state polarity synchronization counting mechanism. This mechanism determines the synchronization of the incremental polarity of the dynamic characteristic sequences of the three data streams and, combined with the verification of the physical characteristics of the vibration source, ultimately generates an early warning signal characterizing potential early-stage faults. In the engineering context of wind turbine monitoring, a challenge stems from the significant differences in time resolution between multi-modal signals such as vibration, current, and temperature. Changes in vibration signals occur at millisecond intervals. While temperature changes occur on the order of seconds or even minutes, existing technologies using downsampling or time window alignment to achieve data fusion lose the transient change information contained in high-frequency signals that can characterize early minor damage. In view of this, the system disclosed in this invention has its edge computing unit configured to receive vibration time-series data, current time-series data, and temperature time-series data from vibration sensors, current sensors, and temperature sensors in parallel without time alignment. This parallel processing architecture avoids information compression caused by data alignment by setting up independent acquisition and processing channels for each signal, thereby preserving the original time resolution and information of each modal signal.

[0041] To transform the raw data stream into a sensitive indicator that can characterize changes in health status, the system is configured to perform parallel and unaligned feature extraction. For the vibration channel, it is necessary to quantify the transition of a vibration signal from stationary to irregular, caused by early faults such as bearing pitting or gearbox lubrication deterioration. To this end, the system calculates a first feature sequence based on the vibration time-series data. This sequence is obtained by calculating the Shannon entropy of the amplitude distribution of vibration time-series data within one calculation cycle, thus characterizing the disorder of the vibration signal. For the current channel, it is necessary to capture the deviation of the current waveform from the normal pattern caused by small irregular fluctuations in the motor load. To this end, the system calculates a second feature sequence based on the current time-series data. The sequence is obtained by calculating the KL divergence between the current time-series data and a healthy current template pre-stored in the edge computing unit, thus characterizing the deviation of the current signal. For the temperature channel, it is necessary to capture the temperature rise trend caused by the accumulation of long-term small abnormal friction. To this end, the system calculates a third feature sequence based on the temperature time-series data to characterize the trend of the temperature time-series data. The temperature gradient sequence is obtained by calculating the difference between the temperature at the current time point and the temperature at a previous time point. In this way, the system converts the three raw data streams into three independent low-dimensional feature sequences reflecting the transition direction of the system's health status. The core early warning logic of the system, namely the three-state polarity synchronous counting mechanism, is designed to address the situation where the absolute values ​​of various physical quantities do not change significantly in the nascent stage of a large number of early faults, but rather the trends of multiple physical quantities change in a temporally coordinated and disordered manner. To achieve this purpose, the mechanism is configured to first determine the first feature sequence within a preset calculation period, such as one hundred milliseconds. Second feature sequence and the third feature sequence The incremental polarity, where positive incremental polarity is defined as the eigenvalue of the current calculation cycle being greater than the eigenvalue of the previous calculation cycle, for example... The polarity of the increment of the third feature sequence is determined to be positive, specifically by determining that the value of the temperature gradient sequence is greater than zero; if and only if , and When the increment polarity is positive at the same time, a synchronization count value Only then is the accumulation operation allowed; this synchronous count accumulation process takes place within a preset time window, such as one minute, and when the time window ends, if... When the cumulative value exceeds a warning threshold, the system generates a corresponding fault warning signal. This logic, which accumulates risk counts only when three things happen simultaneously—vibration becoming more disordered, current becoming more deviated, and temperature rising—captures the cross-modal collaborative characteristics in fault evolution, thereby shifting the warning method from judging static absolute values ​​to identifying the dynamic process of system degradation.

[0042] To improve the accuracy of early warnings, the system further integrates the ability to physically distinguish the root cause of faults into the early warning logic. This is to address certain operating conditions caused by external environmental factors, such as asymmetric icing of wind turbine blades, which can trigger strong periodic vibrations and increased loads, leading to... , and When the incremental polarity of the three values ​​is simultaneously positive, a false alarm is generated; therefore, the system synchronizes the count values. Before accumulation, a spectral symmetry check step is configured to be performed. This step is activated immediately upon capturing a signal with positive tri-state polarity synchronization. A Fast Fourier Transform is performed on the current vibration time series data, followed by spectral analysis. The symmetry of the spectral energy distribution on both sides of the center frequency is then calculated. Specifically, the total energy in the left half-band of the center frequency is compared to the total energy in the right half-band. Only when the ratio falls within a predetermined interval including the value 1, such as between 0.9 and 1.1, is the vibration determined to originate from an internal mechanical fault related to symmetry, and the synchronization count is then allowed to be checked. The system performs an accumulation operation. Since vibrations caused by internal mechanical faults often exhibit symmetrical frequency distribution, while external mass imbalances show significant asymmetry, this verification step analyzes the spectral distribution of vibration energy to enable the early warning logic to distinguish the physical source of the fault, thereby suppressing interference from external environmental factors. Furthermore, to cover different fault evolution modes, the system also integrates diagnostic capabilities for slowly evolving chronic, progressive faults. The risk of such faults increases slowly at a stable, approximately linear rate, which may cause the main early warning mechanism's judgment logic based on cumulative thresholds to become sluggish. To address this issue, the edge computing unit is also configured to execute a chronic fault diagnosis step, which continuously records synchronous count values ​​within a long-period sliding window, such as one hour. The time series formed is calculated, and the standard deviation of the time series is calculated. As a statistic of dispersion; for a system entering a deterministic degradation path, the increase in its risk count will shift from a healthy, discontinuous, irregular state to a smooth, low-dispersion state. Therefore, when the dispersion statistic... When the fault value falls below a preset chronic fault threshold, the system generates a chronic fault warning signal that is distinct from and independent of the main fault warning signal. This diagnostic dimension enables the system to identify the dynamics of different fault evolutions.

[0043] To extend the system's monitoring capabilities from passive condition monitoring to active condition detection, especially in the pre-wear stage before physical wear occurs in the bearing, where implicit state variables cannot be directly observed through passive monitoring, the system is also configured to perform an active lubrication condition detection and feedback step. This step is triggered when the fan is in a preset low-load steady-state condition. A standardized, weak, non-harmonic micro-disturbance current signal is injected into the motor via the converter. During the injection, the current signal is acquired at high frequency, and the instantaneous phase difference between its fundamental frequency and a preset higher harmonic is calculated. Because the instantaneous rupture of the lubricating oil film causes nonlinear mechanical damping, it leads to a significant jump in the harmonic phase difference. When the system detects the phase difference... When the instantaneous jump exceeds a preset phase difference threshold, it does not immediately trigger an alarm, but instead temporarily increases the synchronization count value. Weight in the cumulative operation of the main early warning mechanism For example, weight Setting the value to be greater than 1, this closed-loop feedback enables the main early warning mechanism to self-optimize based on internal detection results, making it more sensitive to potential subsequent collaborative disorder events. To enrich diagnostic information, especially for tracing the root cause of faults, the system also utilizes byproduct information from signal processing, in order to calculate vibration entropy. When the original vibration signal is bandpass filtered, the resulting high-frequency residual signal is usually discarded as computational noise. However, this signal contains hidden information characterizing the contamination of the gearbox lubricating oil by tiny metal particles. Therefore, the system is also configured to perform an accompanying oil contamination diagnosis step, which processes the vibration time-series data to calculate the first feature sequence. During the process, the high-frequency residual signal is separated, and based on the impact pulse events in the signal, the time delay entropy of the time interval distribution between adjacent impact pulse events is calculated. The random intervention of tiny particles in contaminated oil can cause gear meshing impacts to become chaotic and disordered, leading to increased time delay entropy. Increase, at that time entropy When the oil contamination threshold is exceeded, the system generates an independent maintenance suggestion signal indicating abnormal gearbox lubrication. This function provides a possible root cause explanation for the alarm of the main early warning mechanism, extending the diagnostic capability from symptom judgment to root cause tracing. To ensure the system's reliability under extreme low-temperature conditions, the system also incorporates adaptive compensation capabilities for changes in sensor characteristics. This addresses the potential decrease in sensitivity or frequency response drift of piezoelectric vibration sensors caused by extreme low temperatures such as -40°C, which could affect... The issue of computational accuracy exists; therefore, the edge computing unit is also configured to perform a low-temperature adaptive compensation step, which is automatically activated when the ambient temperature is detected by a temperature sensor to be below a low-temperature threshold, such as -20°C. This step first checks the standard deviation of the fundamental phase of the current signal. After verification, if the current signal is confirmed to be reliable, the system recognizes that it is currently in a condition where the vibration sensor may be unreliable, and dynamically and temporarily uses the first characteristic sequence for judgment. The zero reference with positive incremental polarity is adjusted to a small negative reference, for example... This compensation logic acknowledges that the sensor may have negative drift at low temperatures, but as long as its trend is relatively upward, it is still considered to meet the trend condition of vibration deterioration, thus ensuring the effectiveness of the core early warning principle under different environmental conditions.

[0044] In a specific implementation of this invention, the values ​​of the calculation cycle, time window, and early warning threshold are all determined through a standardized pre-calibration procedure. This procedure begins with the analysis of the known physical parameters of the monitored wind turbine and its components, along with the collected health and fault status datasets. Specifically, the calculation cycle... The value range is determined by the highest characteristic fault frequency of the monitored component. Minimum processing time per cycle for edge computing units Common constraints, their relationship satisfies Time window The value of is related to the thermodynamic time constant of the monitored system. Related, As a physical quantity characterizing 63.2% of the time required for a system to reach a new steady state from a thermodynamic steady state, The range of values ​​is set to The first and second warning thresholds are determined by the health status. Dataset Compared with known early fault states Dataset The results were obtained through receiver operating characteristic (ROC) analysis. This analysis plotted the relationship between the true positive rate (TPR) and the false positive rate (FPR) at different thresholds, and determined the first warning threshold as meeting the criteria. The corresponding value with the minimum FPR under the condition is used to determine the second warning threshold. The highest corresponding value of TPR under the given conditions; to enable the early warning mechanism of this invention to adapt to special working conditions, the adjustment logic of its key dynamic parameters also follows a deterministic procedure; wherein, in the active lubrication state detection and feedback step, the value of weight W is a sudden jump of the phase difference with the detected harmonic. Piecewise function associated with amplitude This function is established through offline calibration. The mapping relationship between amplitude and PPM concentration of metal abrasive particles in lubricating oil is established, and the PPM concentration is further divided into multiple deterioration levels, with each level assigned a monotonically increasing deterministic weight value; in addition, in the low-temperature operating condition adaptability compensation step, a negative value reference is used. The value, through low temperature Under healthy operating conditions, a micro-perturbation is applied to the system by a standard exciter. This micro-perturbation can produce a known small increment of positive vibrational disorder at room temperature. , The value is directly determined as the vibration disorder increment measured in this low-temperature experiment. and The absolute value of the difference, i.e. .

[0045] Example 2 In a continuously operating large wind turbine generator, a microscale pitting corrosion initially forms on the surface of a main bearing in its drivetrain. The incremental vibration energy generated by this pitting in its initial stage, along with the resulting motor load fluctuations and heat accumulation, are all below the static alarm threshold set by conventional monitoring systems, causing the generator to appear healthy in the conventional monitoring system interface. The fault early warning and diagnosis system of this invention deployed in the generator uses its edge computing unit to receive and process unaligned data streams from vibration, current, and temperature sensors in parallel. In the early stages of fault evolution, the microscale pitting corrosion on the bearing surface slightly disrupts the regularity of the vibration signal, causing the first characteristic sequence characterizing the disorder of the vibration signal to... During certain calculation cycles, the incremental polarity is positive. Correspondingly, this mechanical irregularity is transmitted to the motor, causing slight fluctuations in the load, which in turn affects the second characteristic sequence characterizing the deviation of the current signal. Also in terms of time The fluctuations are related, and the incremental polarity is also positive in some calculation cycles; as the process continues, the long-term additional friction caused by microscopic damage makes the third characteristic sequence characterizing the temperature change trend... It also began to show a value greater than zero, that is, the incremental polarity was positive.

[0046] During this process, the system's three-state polarity synchronization counting mechanism is executed; this mechanism does not make judgments. , or Instead of determining their absolute values, it judges whether their incremental polarities are simultaneously positive; within a one-minute time window, in each 100-millisecond calculation cycle, , and When all three conditions are met, the synchronization count value An accumulation is then performed; although the frequency of such three-state synchronization events is not high in the early stages of a fault, its frequency is higher than that of occasional synchronization caused by random noise in a fault-free state. As time progresses and micro-damage accumulates, the frequency of three-state synchronization events gradually increases, leading to... When the cumulative value steadily increased within a time window and eventually exceeded the first warning threshold set through offline calibration experiments, the system immediately generated a Level 1 warning signal. After receiving the Level 1 warning signal, the maintenance team inspected the wind turbine and confirmed that there was pitting damage in the early stage on the main bearing raceway. At this time, the wind turbine's total vibration energy level and gearbox temperature, among other key absolute values, had not reached the industry-standard alarm limits. Subsequently, a planned and low-cost maintenance operation was carried out, replacing a high-cost, unplanned repair that might have occurred months later, accompanied by greater component damage and longer downtime.

[0047] Example 3 To objectively verify the effectiveness of the three-state polarity synchronous counting mechanism of this invention in identifying early latent faults, this embodiment constructed and executed a controlled bench test. The purpose was to verify, in a reproducible laboratory environment, whether the mechanism could distinguish a system with implanted microscopic early damage from a known healthy system. The test used a wind turbine gearbox transmission system test platform, which consisted of a drive motor, a speed-increasing gearbox, and a load system, and was equipped with vibration sensors, a three-phase current sensor, and a gearbox lubricating oil temperature sensor. All sensor signals were connected to a device equipped with the pre-... The edge computing unit of the alarm and diagnostic system was used. The test set up two parallel sample groups: a control group equipped with brand new and intact bearings, and an experimental group with bearings that had been pre-treated with micro-sized rolling element pitting by electrical discharge machining at the same location. Except for the bearing condition, all mechanical parts, lubricant type, operating load and speed curves of the two sample groups were kept consistent. In the test, the calculation cycle was set to 100 milliseconds. This setting was intended to balance the ability to capture high-frequency dynamics in the monitored signal with the data processing load of the edge computing unit, so as to ensure effective monitoring of the evolution process of fault characteristics without data omission.

[0048] After the experiment started, the control group and the experimental group ran continuously for 8 hours under the same operating conditions. The edge computing unit synchronously recorded the multimodal data of the two systems and calculated the maximum vibration amplitude within minutes, the average temperature within minutes, and the average temperature within minutes, respectively. Cumulative values; Throughout the entire test period, there were no discernible differences in vibration amplitude and absolute temperature between the control and experimental groups. For example, after 8 hours of testing, the maximum vibration amplitude per minute for the control group was 0.16g, and for the experimental group it was 0.19g, while the average temperatures per minute for both were 46.6°C and 48.5°C, respectively, both within the normal safety threshold. However, at the same time point, the synchronous count values, which characterize the synergistic trend of multimodal data changes, showed no significant differences. The control group showed a significant difference. The cumulative value was only 2, while the experimental group's value had accumulated to 215; the experimental group The continuous upward trend in the cumulative value is due to the fact that the pre-fabricated micro-damage points, under continuous mechanical stress, induce three incremental polarities in vibration disorder, current deviation, and temperature gradient. These non-random synchronization events are positive events, and the three-state polarity synchronization counting mechanism accumulates them, thus translating a latent fault that is difficult to detect in absolute terms into a quantitative indicator with a significant changing trend. In contrast, the control group... The cumulative value fluctuated randomly at an extremely low level throughout the entire experiment, reflecting the non-correlation of the changing trends of various parameters in the healthy system. Experimental results show that, under the same operating conditions and monitoring cycle, monitoring indicators based on the absolute values ​​of physical quantities cannot distinguish between a healthy system and a system with early microscopic damage. However, the early warning system of this invention, based on the incremental polarity synchronization judgment of multimodal time-series data, outputs a synchronized count value... This can serve as an effective distinguishing indicator, thus confirming the effectiveness of the technical solution in identifying latent faults.

[0049] Example 4 This embodiment combines Figures 1 to 3 This document describes a wind turbine fault early warning and diagnosis system that integrates multimodal time-series data. Figure 1As shown, the central processing core of this process is a three-state polarity synchronous counting mechanism. Its inputs are vibration time-series data, current time-series data, and temperature time-series data acquired in parallel. The system first calculates, based on these three independent, non-time-aligned data streams, the disorder degree (characterizing vibration irregularity), the deviation degree (characterizing the difference between the current waveform and the healthy template), and the temperature change trend (capturing the heating process). Subsequently, the system performs a three-state incremental polarity synchronization judgment, determining whether the incremental polarity of the disorder degree, deviation degree, and temperature trend are simultaneously positive. After confirming positive synchronization, it further performs a spectral symmetry check to confirm that the vibration originates from an internal mechanical fault rather than external interference. Only when the spectral symmetry condition is met does the system update the synchronization count value. The system performs an accumulation operation, whereby the count value accumulates within a time window and is compared with a preset warning threshold. If the threshold is exceeded, a level one or level two fault warning signal is generated. Furthermore, this core logic is linked to two advanced diagnostic modules. One is an active lubrication detection and feedback module, which detects pre-wear risks under low-load conditions by injecting micro-disturbance current and analyzing harmonic phases. When a risk is detected, the accumulation weight of the synchronous count value is temporarily increased. ( The second is the chronic fault diagnosis module, which identifies low dispersion states caused by the smooth growth of risk by analyzing the standard deviation of count values ​​over a long period, and generates independent chronic fault early warning signals accordingly.

[0050] like Figure 2 The system is in a healthy operating state during normal operation. Starting from this state, if the synchronization count value exceeds the first-level threshold, the system will migrate to the first-level warning state. If it continues to deteriorate and exceeds the second-level threshold, it will enter the second-level warning state. In the first-level warning state, after maintenance intervention or state recovery, the system can return to the healthy operating state. In addition, starting from the healthy operating state, if the dispersion of the count value is lower than the preset chronic fault threshold, the system will directly migrate to the independent chronic fault state. At the same time, when the system enters a low-load condition, it can enter the active lubrication detection state from the healthy operating state. After the detection is completed, it will return to the healthy operating state. This detection process will have an impact, namely, improving the warning sensitivity.

[0051] like Figure 3The system is deployed at the field end of the wind turbine generator set, monitoring key components such as gearboxes and bearings. The system's hardware is based on an industrial-grade microcontroller, which acts as an edge computing unit. This unit receives real-time data from vibration, current, and temperature sensors via wired connections and executes an internal fault warning and diagnosis system. This core system consists of four main software modules: a data acquisition and processing module, a three-state polarity synchronous counting module, an advanced diagnostic module, and a warning generation and communication module. The system output is divided into two paths: first, when a warning or suggestion is generated, it is sent to the remote monitoring center via network communication for use by the target user, i.e., the maintenance team, for decision-making; second, when performing active detection functions, control signals are applied to the executing / controlled objects, i.e., the motor / converter.

[0052] Example 5 Before a wind turbine generator set that has undergone installation or major maintenance is put into regular operation, a systematic online self-calibration and parameter calibration procedure needs to be performed to ensure that the internal model and judgment threshold settings of the early warning and diagnosis system of this invention are matched with the electromechanical characteristics of the specific unit. This procedure aims to provide a benchmark based on the unit's own health characteristics for subsequent online monitoring. After the procedure is initiated, the edge computing unit enters a preset learning mode. In this mode, the wind turbine, under the monitoring of maintenance personnel, operates stably for a preset period of 50 hours under a series of known healthy operating conditions covering its main operating power range. During this period, the system collects vibration current and temperature data at high frequency for the calibration of the following parameters: First, to construct a second feature sequence for calculation. The system collects all three-phase current time-series data under the aforementioned healthy operating conditions, and based on this dataset, constructs a statistical model describing the probability distribution of the current vector under healthy conditions. This model is then stored in the non-volatile memory of the edge computing unit as a benchmark for subsequent calculations of the KL divergence. Secondly, to determine the symmetry conditions used in the spectral symmetry verification step, the system continuously performs spectral analysis on the collected vibration time-series data during the same learning mode operation, and calculates the symmetry ratio of the spectral energy distribution on both sides of the center frequency. Since the unit is in a confirmed healthy state, the internal mechanical vibrations generated The value will fluctuate slightly around 1.0, and the system will record all values ​​during this period. Values ​​and calculate their statistical standard deviation. Ultimately, the predetermined interval for the symmetry condition was determined as follows: This provides a quantitative basis for the subsequent judgment logic that distinguishes between internal mechanical failures and external quality imbalances, based on the machine's health characteristics.

[0053] Third, to calibrate the threshold for the dispersion statistics used in the chronic fault diagnosis steps, the system also executes a three-state polarity synchronization counting mechanism during the learning mode and records the generated synchronization count values. The time series, since the unit is in a healthy state, The growth of the sequence is random and discontinuous, and the system calculates the health status based on this. Standard deviation of the sequence Subsequently, a chronic fault threshold is used to determine whether the system has entered a deterministic degradation path. By setting it as a predetermined proportion of the standard deviation of that health state. This is determined based on the premise that a chronic deterministic failure will lead to... The growth process tends to be smooth, and its volatility will be significantly lower than the random fluctuations in a healthy state. After completing the calibration of all the above parameters and the construction of the model, the edge computing unit exits the learning mode and switches to the regular online monitoring mode. At this point, all the core judgment logic inside the system has been individually configured based on the inherent health status baseline of the specific unit, providing support for the reliability of its subsequent early warning and diagnosis.

[0054] Example 6 When a wind turbine generator deployed in a high-altitude, cold region experiences an ambient temperature below the preset -20°C low-temperature threshold measured by a temperature sensor, this condition may cause a drift in the response sensitivity of the vibration sensor, thereby affecting the first characteristic sequence. The accuracy of the calculation; under this condition, the system automatically executes the low-temperature operating condition adaptive compensation step, first calculating the standard deviation of the fundamental phase of the current time series data in the first time period. When judged When the current signal is less than the preset phase stability threshold, indicating that the power grid frequency is stable and the current signal is reliable, the system will use it to determine the first feature sequence. The zero reference with positive incremental polarity is adjusted to a negative reference. .

[0055] Under this adaptive compensation logic, even a vibration deterioration process caused by a real mechanical failure will have its calculated increment affected by the sensor's low-temperature drift. For a small negative value, as long as the value is still greater than the negative baseline. The system still determines that its incremental polarity is positive; this adaptive adjustment enables the three-state polarity synchronous counting mechanism to effectively capture the coordinated disorder process that characterizes real mechanical faults under the boundary conditions of limited sensor performance, thereby maintaining the reliability of the system's early warning function in specific environments.

[0056] Example 7 In this embodiment, after the system self-calibration and parameter calibration procedures are completed, a targeted module parameter optimization process needs to be performed to enable and optimize specific advanced diagnostic functions within the system. The purpose is to determine the internal operating parameters that maximize the diagnostic discrimination of the two functional modules: active lubrication condition detection and accompanying oil contamination diagnosis. To determine the parameters of the micro-perturbation current signal in the active lubrication condition detection and feedback steps, an offline optimization experiment is conducted on a test platform that has completed basic calibration. The objective function of this experiment is to maximize the phase difference response under both healthy lubricating oil and early-contaminated lubricating oil conditions. The difference; In the experiment, a series of non-harmonic current signals with different frequencies and amplitudes were injected into the motor, where the frequency scanned within the non-harmonic band, and the amplitude was controlled within 1% of the motor's rated current. Simultaneously, a high-precision torque sensor was used to monitor and confirm that the injection operation did not have a measurable effect on the motor's output torque; Finally, a method that could achieve both healthy and polluted states was selected. The disturbance signal parameters with the greatest response difference, i.e., the specific combination of frequency and amplitude, are used as the standard micro-disturbance current signal parameters for this module, and the parameters under pollution conditions are also considered. The statistical lower bound of the response jump amplitude is set as the phase difference threshold.

[0057] To determine the internal parameters in the accompanying oil contamination diagnostic procedure, another optimization experiment was performed on the same test platform, with the goal of maximizing the time delay entropy. The indicator assesses the signal-to-noise ratio of trace contaminants in oil. First, the high-frequency residual signal is defined as the high-frequency component remaining after filtering out low- and mid-frequency components from the original vibration signal using a bandpass filter. The passband range of this filter is determined by the characteristic fault frequency, which is determined by the geometric parameters and rotational speed of the monitored bearing. Second, the threshold for identifying impact pulse events is varied under clean lubricating oil conditions, and the results are compared after adding trace standard contaminants. The relative change in value, finding a value that can make The optimal impact pulse recognition threshold is most sensitive to trace contaminants; ultimately, by adding standard contaminants of different concentrations to the lubricating oil in batches, a [system / mechanism] is established. The correlation between numerical values ​​and oil contamination levels provides experimental basis for setting oil contamination thresholds. After completing the above parameter optimization process, the two advanced functional modules, active lubrication condition detection and accompanying oil contamination diagnosis, are activated. Their internal operating parameters have been optimized for the current tested system, making the entire early warning and diagnosis system functionally complete.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning and diagnosis of wind turbine faults by integrating multimodal time-series data, characterized in that, include: Vibration time-series data, current time-series data, and temperature time-series data of the wind turbine are acquired respectively, and the first characteristic sequence, the second characteristic sequence, and the third characteristic sequence are calculated respectively. Fault warning signals are generated based on a three-state polarity synchronous counting mechanism; Calculate the dispersion statistics and perform chronic fault diagnosis; When the fan is under low load steady-state operation, active lubrication status detection and feedback are performed; While calculating the first feature sequence, concurrent oil contamination diagnosis is performed simultaneously. The ambient temperature is determined using temperature time series data, and low-temperature adaptive compensation is performed under low-temperature conditions.

2. The wind turbine fault early warning and diagnosis method according to claim 1, characterized in that, The calculation of the first feature sequence, the second feature sequence, and the third feature sequence specifically includes: Based on the vibration time series data, a first feature sequence is calculated to characterize the disorder of the vibration time series data; based on the current time series data, a second feature sequence is calculated to characterize the deviation of the current time series data; based on the temperature time series data, a third feature sequence is calculated to characterize the changing trend of the temperature time series data.

3. The wind turbine fault early warning and diagnosis method according to claim 2, characterized in that, The first feature sequence is obtained by calculating the Shannon entropy from the amplitude distribution of vibration time series data; the second feature sequence is obtained by calculating the KL divergence between current time series data and healthy current template; the third feature sequence is obtained by calculating the temperature gradient sequence by calculating the difference between the current temperature and the temperature at the previous time point.

4. The wind turbine fault early warning and diagnosis method according to claim 1, characterized in that, The fault warning signal generation based on the three-state polarity synchronous counting mechanism specifically includes: Within one calculation cycle, the incremental polarity of the first feature sequence, the second feature sequence, and the third feature sequence is determined; before accumulating a synchronous count value, a spectral symmetry check is performed, a spectral analysis is performed on the current vibration time series data, and the symmetry of its spectral energy distribution on both sides of the center frequency is calculated. When the incremental polarity of the first feature sequence, the second feature sequence, and the third feature sequence are all positive, and the symmetry of the spectral energy distribution satisfies the symmetry condition, the synchronization count value is accumulated; when the synchronization count value exceeds the warning threshold within a time window, a fault warning signal is generated.

5. The wind turbine fault early warning and diagnosis method based on multimodal time series data according to claim 4, characterized in that, The symmetry condition for the spectral symmetry check is as follows: the ratio of the total energy of the spectrum in the left half-band to the total energy in the right half-band of the center frequency falls within a predetermined interval containing the value 1.

6. The wind turbine fault early warning and diagnosis method according to claim 4, characterized in that, The warning thresholds include a first warning threshold and a second warning threshold, wherein the second warning threshold is greater than the first warning threshold. When the synchronization count value exceeds the first warning threshold, a first-level warning signal is generated, and when the synchronization count value exceeds the second warning threshold, a second-level warning signal is generated.

7. The wind turbine fault early warning and diagnosis method according to claim 1, characterized in that, The calculation of the dispersion statistics and the execution of chronic fault diagnosis specifically include: Within a long-period sliding window, a time series consisting of synchronization count values ​​is recorded, and the standard deviation of the time series is calculated as the dispersion statistic. When the dispersion statistic is lower than the chronic fault threshold, an independent chronic fault warning signal, distinct from the fault warning signal, is generated.

8. The wind turbine fault early warning and diagnosis method according to claim 1, characterized in that, The active lubrication status detection and feedback when the wind turbine is under low load steady-state operation specifically includes: When the wind turbine is in a low-load steady-state condition, a standardized micro-disturbance current signal is injected into the motor by controlling the converter. During the injection, the current signal is collected and the phase difference between the fundamental wave and the higher harmonics is calculated. When the instantaneous jump of the phase difference exceeds the phase difference threshold, the weight W of the synchronization count value in the accumulation operation is temporarily increased, where W>1.

9. The wind turbine fault early warning and diagnosis method according to claim 1, characterized in that, The simultaneous execution of accompanying oil contamination diagnosis during the calculation of the first feature sequence specifically includes: In the process of processing vibration time series data to calculate the first feature sequence, high-frequency residual signals are separated; and based on the impact pulse events in the high-frequency residual signals, the time delay entropy of the time interval distribution between adjacent impact pulse events is calculated; when the time delay entropy exceeds the oil contamination threshold, a maintenance suggestion signal indicating abnormal gearbox lubrication status is generated.

10. The wind turbine fault early warning and diagnosis method according to claim 1, characterized in that, The process of using temperature time-series data to determine ambient temperature and performing low-temperature adaptive compensation under low-temperature conditions specifically includes: When the ambient temperature is detected to be lower than the low temperature threshold, the standard deviation of the fundamental phase of the current time series data within the first time period is calculated; when the standard deviation is less than the phase stability threshold, the zero reference used to determine that the polarity of the first characteristic sequence increment is positive is adjusted to a negative reference.

Images