Wind turbine generator fault prediction method and system based on multi-modal feature fusion
By using a multimodal feature fusion method, a benchmark sensor is selected, correlation coefficients are calculated, and thresholds are adjusted to form a standard association combination. This solves the data lag problem in wind turbine fault prediction and achieves efficient and accurate fault early warning and prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-03-20
AI Technical Summary
In existing wind turbine fault prediction, the data information of a single mode is limited in dimension and ignores the influence of environmental factors and the correlation of multiple components, resulting in prediction lag, low accuracy, and difficulty in early warning of fault risks.
By selecting a benchmark sensor, defining an effective time limit, calculating the sensor correlation coefficient, setting a judgment threshold, forming an initial association combination, adjusting the association combination, analyzing the similarity, merging into a standard association combination, predicting operating data, and outputting an alarm signal.
It enables accurate prediction of wind turbine faults in dynamic environments, improving the accuracy and timeliness of predictions and ensuring the stable operation and efficient maintenance of wind turbines.
Smart Images
Figure CN121706009A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine fault prediction technology, and more specifically, to a wind turbine fault prediction method and system based on multimodal feature fusion. Background Technology
[0002] Offshore wind turbines are exposed to extreme environments such as high salt spray, strong waves, and typhoons for extended periods, which can easily lead to multiple failures. High salt spray accelerates the corrosion of metal components and the aging of composite materials, strong waves cause foundation swaying and abnormal loads on the transmission system, and strong typhoon winds can cause aerodynamic damage to the blades, ultimately leading to problems such as tower corrosion, blade surface damage, and gearbox lubrication failure. Furthermore, since offshore operation and maintenance relies on specialized vessels and high-altitude operation equipment, it is greatly restricted by weather conditions. The core components of offshore wind turbines are mostly located in high-altitude nacelles or underwater foundations, making manual inspection difficult and time-consuming. This results in offshore operation and maintenance costs being significantly higher than onshore costs and a substantial reduction in troubleshooting efficiency. Thanks to the development of high weather-resistant sensor technology and data monitoring technology (which can stably collect data in harsh marine environments and solve the problems of easy damage and data distortion of traditional sensors), by deploying sensors on different components of the unit to collect operating data and comparing it with preset anomaly judgment rules, it can be determined whether each component is abnormal. If an anomaly is found, the system will output a signal containing the location of the abnormal component, prompting maintenance personnel to handle it in time and avoid the fault from worsening and causing downtime.
[0003] Due to the highly dynamic nature of the marine environment (such as random changes in wind speed and direction, and foundation swaying caused by tides), and the need for the generator unit to adjust its power output according to wind conditions, the correlation between sensors changes dynamically, and the combination of correlations varies at different times. For example: In the gearbox monitoring of wind turbines, the vibration sensor serves as the reference sensor. Its correlation with other sensors such as temperature sensor and oil sensor will vary depending on factors such as the gearbox's running time, load changes, and ambient temperature. Within different effective time limits (such as the time intervals corresponding to the equipment start-up phase, stable operation phase, and high load phase), different correlation combinations will be presented. Therefore, there is an urgent need for a wind turbine fault prediction method and system based on multimodal feature fusion that can fully consider the dynamic changes and effective time limits of sensor association combinations, thereby accurately predicting the operating data of the benchmark sensor. Summary of the Invention
[0004] The purpose of this invention is to solve the problems in existing wind turbine fault prediction, such as limited dimensions of single-mode data information, neglect of the influence of environmental factors and the correlation of multiple components, reliance on manual analysis or ineffective integration of multi-source data, resulting in prediction lag, low accuracy, and difficulty in early warning of fault risks.
[0005] To achieve the above objectives, the present invention provides S1: Selecting a sensor as a reference sensor; retrieving the effective time limit of the corresponding component during normal operation of the reference sensor and defining it as a corresponding time interval; retrieving the operating data of the reference sensor and other sensors within the effective time limit according to each time interval; Calculate the correlation coefficient between the reference sensor and other sensors at each time interval, set a judgment threshold to determine the correlation between the reference sensor and other sensors, and sort the other sensors according to the size of the correlation coefficient to form the initial association combination for each corresponding time interval; S2. Adjust the time interval. During the adjustment process, multiple adjusted association combinations are obtained. Calculate the similarity between the adjusted association combination and the corresponding initial association combination. Analyze the changes of each adjusted association combination. Determine multiple identical initial association combinations and their corresponding adjusted association combinations, and then merge them into the same association set. The most similar adjusted association combination within the same association set is defined as the standard association combination corresponding to the same association set. Retrieve the standard association combination within each identical association set, select other identical sensors, and adjust the judgment threshold based on the minimum correlation coefficient between the other identical sensors and the benchmark sensor; And select the adjustment period corresponding to the standard associated combination, and define it as the selection period; S3. Predict the operating data of the reference sensor during the prediction period: Select the effective data of the reference sensor and other sensors in the current period, recalculate the correlation coefficient, and use the adjusted judgment threshold to judge other sensors that are related to the reference sensor, and then form a real-time candidate combination. Calculate the matching degree between real-time candidate combinations and standard associated combinations, select the standard associated combination with the highest matching degree, and predict the operating data of the benchmark sensor within the corresponding prediction period; sense the corresponding anomaly judgment rule, compare the predicted operating data with the anomaly judgment rule, and output an alarm signal if an anomaly is judged.
[0006] As a further improvement to this technical solution, S1 receives historical operation and maintenance data, and sequentially retrieves the repair completion timestamp and fault occurrence timestamp corresponding to the same abnormal component; by subtracting the fault occurrence timestamp from adjacent repair completion timestamps, multiple valid time limits are obtained; the operating data detected by each sensor within each valid time limit is defined as valid data. Select the first sensor used to detect component operation data as the reference sensor, and retrieve the valid data of the reference sensor within each valid time limit; Using the effective time limit of the reference sensor as the reference time period, retrieve the effective data of other sensors besides the reference sensor within the reference time period; Calculate the correlation coefficient between the reference sensor and other sensors in each valid time limit; set a judgment threshold, and if the correlation coefficient is greater than the judgment threshold, it is determined that there is a correlation between the reference sensor and other sensors in the valid time limit.
[0007] As a further improvement to this technical solution, in S1, the operating data of the reference sensor within the effective time limit is defined as the first data sequence, and the operating data of other sensors within the corresponding effective time limit is defined as the second data sequence. The numerator for calculating the correlation coefficient is obtained by successively subtracting the mean of the first data sequence from each running data point in the first data sequence, and subtracting the mean of the second data sequence from each running data point in the second data sequence, multiplying the resulting differences, and summing them. The denominator is obtained by multiplying the square root of the sum of squares of each data point in the first data sequence minus its mean by the square root of the sum of squares of each data point in the second data sequence minus its mean. Then, the correlation coefficient between the reference sensor and other sensors within the effective time limit is obtained by dividing the numerator by the denominator.
[0008] As a further improvement to this technical solution, S2 receives multiple valid time limits of the abnormal component and retrieves the minimum time period; For the effective time limit greater than the minimum time period, a fixed step size is set, and the effective time limit greater than the minimum time period is gradually reduced as the adjustment period until the adjustment period equals the minimum time period. During the process of adjusting the effective time limit with a fixed step size, each adjustment reduces the running data within the effective time limit by one. When the effective time limit is gradually adjusted to be greater than the minimum time period, multiple other sensors that are related to the reference sensor within the adjustment time period are calculated in turn and constructed as corresponding adjustment association combinations. Each effective time limit corresponds to multiple adjustment association combinations. As a further improvement to this technical solution, in step S2, the changes in the adjustment association combination corresponding to each successively decreasing adjustment period are calculated and analyzed. If the similarity between multiple adjustment association combinations and a certain initial association combination is an ordered change, then it is determined that the initial association combination is the same as the initial association combination corresponding to multiple adjustment association combinations. After determining that they are the same, the next initial association combination is synchronously adjusted using a fixed step size. The next initial association combination is set as the initial association combination corresponding to the adjusted association combination, and the similarity is calculated and the judgment is continued until the similarity between the adjusted association combinations changes in an unordered manner. Then, the initial association combination when the unordered change occurs is retrieved, and the judgment is repeated to determine whether they are the same. In this way, multiple identical initial association combinations and their corresponding adjusted association combinations are merged into the same association set.
[0009] As a further improvement to this technical solution, the adjustment of the association combination corresponding to the initial association combination is as follows: sort multiple valid time limits, select the largest time limit as the priority to decrease the time limit, and gradually decrease it towards the second valid time limit corresponding to the second largest time limit; When gradually reducing the maximum time limit to the second effective time limit, the initial association combination corresponding to the association combination is adjusted during the reduction process, which is the initial association combination corresponding to the second effective time limit; If the adjustment period corresponding to the maximum time limit equals the second effective time limit, then the second effective time limit is reduced simultaneously. At this time, the adjustment association combination corresponding to the maximum time limit and the second effective time limit is obtained. The initial association combination corresponding to the adjustment association combination is the initial association combination corresponding to the third effective time limit.
[0010] As a further improvement to this technical solution, in step S2, the similarity between the initial association combination and each adjusted association combination in each identical association set is retrieved, and the adjusted association combination corresponding to the maximum similarity is retrieved and defined as the standard association combination. Each identical association set corresponds to a standard association combination. Filter other sensors that are the same in multiple standard association combinations, and select the smallest correlation coefficient between the same other sensor and the benchmark sensor as the threshold for judging whether there is a correlation; filter the adjustment period corresponding to the standard association combination, and define it as the selection period.
[0011] As a further improvement to this technical solution, in step S3, when predicting the operating data of the reference sensor during the prediction period, the effective data in the current period is retrieved, and the correlation coefficient between the reference sensor and each other sensor is recalculated. Then, using the judgment thresholds corresponding to other sensors, other sensors that are related to the benchmark sensor are judged. Other related sensors are retrieved to form real-time candidate combinations. The matching degree between these combinations and each standard associated combination is then determined, and the standard associated combination with the highest matching degree is selected.
[0012] As a further improvement to this technical solution, the operating data of other sensors that are related to the benchmark sensor are selected during the selected time period corresponding to the highest matching degree standard association combination. If the operating data of all associated other sensors exist within the selected time period, the operating data of other sensors are used to predict the operating data of the benchmark sensor within the corresponding prediction time period. The anomaly judgment rule corresponding to the sensing reference sensor is to compare the operating data of the reference sensor in the predicted period with the anomaly judgment rule in turn. If there is any operating data that does not meet the anomaly judgment rule, an alarm signal is output. If the operating data of all associated other sensors within the selected time period is missing, the next selected time period corresponding to the standard association combination is selected in descending order of matching degree, until the operating data of all associated sensors and the corresponding sensors in the standard association combination are available.
[0013] A wind turbine fault prediction system based on multimodal feature fusion includes: The initial association combination construction module selects a reference sensor, analyzes multiple effective time limits corresponding to the reference sensor and defines them as time intervals, selects the running data in each time interval, calculates the correlation coefficient between the reference sensor and other sensors in each time interval, sets a judgment threshold to judge the reference sensor and other sensors that have a correlation relationship, and constructs the initial association combination for each corresponding time interval. The threshold adjustment module determines multiple adjusted correlation combinations by adjusting the time interval; it calculates the similarity between the adjusted correlation combination and the corresponding initial correlation combination, analyzes the changes in each adjusted correlation combination, identifies multiple identical initial correlation combinations and their corresponding adjusted correlation combinations, and then merges them into a common correlation set; it analyzes the standard correlation combination corresponding to the most similar adjusted correlation combination within the common correlation set; it retrieves the standard correlation combination within each common correlation set, selects other identical sensors, and adjusts the judgment threshold based on the minimum correlation coefficient between the other identical sensors and the baseline sensor; and it selects the adjustment time period corresponding to the standard correlation combination, defining it as the selection time period. The running data prediction output module predicts the running data of the reference sensor within the prediction period: selects the effective data of the reference sensor and other sensors within the current period, recalculates the correlation coefficient, and uses the adjusted judgment threshold to judge other sensors that are related to the reference sensor, and then forms a real-time candidate combination. Calculate the matching degree between real-time candidate combinations and standard associated combinations, select the standard associated combination with the highest matching degree, and predict the operating data of the benchmark sensor within the corresponding prediction period; sense the corresponding anomaly judgment rule, compare the predicted operating data with the anomaly judgment rule, and output an alarm signal if an anomaly is judged.
[0014] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description
[0015] Figure 1 This is a schematic diagram illustrating the selection of effective data in this invention; Figure 2 This is a schematic diagram of the initial association combination of the present invention; Figure 3 This is a schematic diagram illustrating the working principle of the adjusted associated combination of the present invention; Figure 4This is a schematic diagram illustrating the working principle of related association set analysis and threshold adjustment in this invention; Figure 5 This is a schematic diagram illustrating the working principle of analyzing and predicting data in this invention.
[0016] The meanings of the labels in the diagram are as follows: 1. Initial association combination construction module; 2. Judgment threshold adjustment module; 3. Running data prediction output module. Detailed Implementation
[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] refer to Figures 1-5 As shown, a wind turbine fault prediction method based on multimodal feature fusion includes the following steps: When deploying sensors to collect operational data in different components of a wind turbine, the sensors involved include vibration sensors, temperature sensors, oil sensors, etc., and each sensor is responsible for detecting data for one component. For example, in gearbox monitoring, a vibration sensor can be selected as the reference sensor to detect gearbox vibration data, while a temperature sensor is used to detect gearbox oil temperature data and an oil sensor to detect the viscosity or contamination of the oil in the gearbox. Each sensor collects data for different operating indicators of the gearbox. This invention considers that during the operation of different components within a wind turbine, due to the long-term exposure of wind turbines (especially offshore wind turbines) to extreme environments such as high salt spray, strong waves, and typhoons, and the need to dynamically adjust power generation according to wind conditions, the operating states of different components are interconnected (e.g., extreme environments accelerate component aging or cause abnormal loads, and power adjustments change component loads). This leads to mutual influence between data detected by different sensors, resulting in dynamic changes in the correlation between the data detected by sensors deployed on different components (e.g., in gearbox monitoring, the correlation between vibration sensors, temperature sensors, and oil sensors changes synchronously with factors such as gearbox operating time, load changes, and ambient temperature). To accurately analyze other sensors that interact with the sensors, this invention: S1. Select a sensor (the core reference object in the fault prediction process) as the reference sensor; retrieve the effective time limit of the corresponding component during normal operation and define it as the corresponding time interval; retrieve the operating data of the reference sensor and other sensors within the effective time limit according to each time interval, and eliminate the interference of component failure and maintenance process on the operating data detected by the sensor through the above effective time limit, so as to avoid interference of the subsequent judgment of correlation due to the operating data detected in other time limits; To quantify the correlation between data from different sensors, the correlation coefficient between the reference sensor and other sensors is calculated for each time interval. A judgment threshold is set to determine the correlation between the reference sensor and other sensors. The other sensors are then sorted according to the magnitude of the correlation coefficient to form the initial correlation combination for each corresponding time interval. The specific steps are as follows: S1.1, sequentially retrieve the repair completion timestamps corresponding to the same abnormal components. Fault occurrence timestamp Calculate the repair completion timestamp Fault occurrence timestamp Valid time limit Valid period Used to describe the time interval between when a component is in normal operating condition from the completion of maintenance to the time before it fails again, eliminating interference from the fault and maintenance process on sensor detection; Define the validity period The operational data detected by each sensor is valid data: ; in For the first Each sensor within its effective time limit Inner The operational data from this test (such as vibration amplitude, temperature reading, oil viscosity, etc.); S1.2, Select the first A sensor used to detect component operating data serves as a reference sensor. And bring up the reference sensor. In each valid time limit Valid data within the specified range is denoted as valid data. ,in For the effective time limit Internal reference sensor The number of data collected during operation As a reference sensor within the validity period Inner The valid data collected this time; With reference sensor Validity period As a benchmark period Activate other sensors ( During the base period Valid data within ,in Base period Other internal sensors The number of data collected during operation For other sensors During the base period Inner The valid data collected this time; Calculate the reference sensor sequentially Other sensors Correlation coefficient within each valid time period Valid period Corresponding correlation coefficient for: ; in As a reference sensor Corresponding valid data The mean, For other sensors During the base period Internal valid data The mean; Set judgment threshold If the correlation coefficient > Judgment threshold Then it is determined that it is within the valid time limit. Internal reference sensor Other sensors There is a related relationship; S1.3, Valid Reception Time Limit Internal and reference sensors Multiple other sensors that are correlated, and each of these other sensors and the reference sensor. Correlation coefficient between ,in For the effective time limit Inner Other sensors With reference sensor The correlation coefficient between them; The other sensors are sorted according to the correlation coefficient within each valid time limit to form an initial association combination corresponding to each valid time limit. The corresponding initial association combination is .
[0019] Furthermore, this invention takes into account that as the wind turbine operates for a long time, different components will have multiple effective time limits. And valid data within each valid time limit It may also exhibit different patterns of change (for example, the vibration, temperature, and other data fluctuation patterns of the gearbox may differ at different effective time limits due to variations in lubrication conditions and load), thus affecting the reference sensor at different times. Unlike other sensors There is a relationship between them; The effective time limit of different components of a wind turbine It is based on the repair completion timestamps of the same abnormal components in historical operation and maintenance data. Subtract the fault occurrence timestamp Certainly, each valid time limit The length difference makes the reference sensor... Other sensors Different valid time limits Different correlation coefficients and different effective time limits may appear within the same period. Internal and reference sensors Other related sensor combinations may have originally been identical, but due to time constraints... The differences manifest in different correlations, thus affecting the baseline sensor. Unlike other sensors The determination of a correlation can have an impact, therefore, to avoid the above situation: S2, Receive multiple valid time limits of the faulty component in S1.1 Adjust the minimum time period ; For > minimum time period Validity period Set a fixed step size and gradually decrease it to the minimum time interval. Validity period As adjustment period Until the time period is adjusted =Minimum time period ; Adjust the validity period During the process, each adjustment extends the effective time limit. The number of operational data points within the system is reduced by one, thereby gradually refining different effective time limits. Valid data within; During the gradual adjustment, the adjustment periods are calculated sequentially. Internal and reference sensors Multiple other sensors that are correlated are configured to adjust the correlation combination accordingly (each valid time limit). There are multiple possible adjustments and combinations, not just limited to the existing effective time limit. Through multiple adjustment periods To achieve different effective time limits Comprehensive capture of the correlation between sensors; Analyze each successively decreasing adjustment period Corresponding adjustments to the associated combinations: Valid reception time limit The corresponding initial association combination is Adjusting the time period The corresponding adjustment association combination is as follows Then the initial association combination Adjustment of related combinations The similarity is: ; in For the initial association combination The Middle The weight of each position sensor (set according to the rule that the sensor is more important the closer it is to the front). For indicator functions, if (If the same position sensor is used), then the indication function is... Conversely, it indicates the function. ; As a normalization factor, it makes the matching degree range within between; If the similarity between multiple adjusted association combinations and a certain initial association combination changes in an ordered manner (with the adjustment period)... When the change is synchronously increased or decreased, the orderly change reflects that multiple adjustment association combinations and the initial association combination have the same core characteristics of sensor association. Therefore, it is determined that the initial association combination (the initial association combination before the fixed step size adjustment) is the same as the initial association combination corresponding to multiple adjustment association combinations. After determining if they are identical, the next initial association combination is synchronously adjusted using a fixed step size. The next initial association combination is set as the initial association combination corresponding to the adjusted association combination, and similarity is calculated to continue the judgment until the similarity between the adjusted association combinations changes without order. Then, the initial association combination with the unordered change is retrieved, and the judgment is repeated to determine if they are identical. This merges multiple identical initial association combinations and their corresponding adjusted association combinations into a common association set, avoiding different effective time limits. Regarding the impact of correlation analysis, we focus more on the essential correlation logic between different sensors. In the subsequent prediction of wind turbine operating status, accurate correlation relationships can improve the accuracy and efficiency of judgment, and help to more reliably ensure the stable operation of wind turbines. The above-mentioned adjusted association combination corresponds to the initial association combination as follows: sorting multiple valid time limits. Select the maximum time limit As a priority for reducing time limits, it is directed towards the next lowest time limit. Corresponding validity period Gradually decrease; Gradually reduce the maximum time limit Until the expiration date During the reduction process, the initial association combination corresponding to the association combination is adjusted, which is the effective time limit. The corresponding initial association combination; If the maximum time limit Corresponding adjustment period =Valid Time Limit If so, the effective time limit will be reduced simultaneously. At this point, the maximum time limit is obtained. Validity period The corresponding adjustment of the associated combination, at this time the initial associated combination corresponding to the adjusted associated combination, is the effective time limit. The corresponding initial association combination.
[0020] Find the similarity between the initial association combination and each adjusted association combination in each identical association set, find the adjusted association combination with the highest similarity, and define it as the standard association combination. Each identical association set corresponds to one standard association combination. Filter out other sensors that are the same in multiple standard association combinations, and select the same other sensors and the reference sensor. The minimum correlation coefficient between the sensors is used as the threshold for determining whether they are correlated, and a corresponding threshold is assigned to each other sensor. The adjustment period corresponding to the combination of screening criteria. Defined as a time period; thus solving the problem of limited validity time. The problem of bias in association judgment caused by differences is addressed by standard association combinations and precise judgment thresholds. The determination of this method enables a complete closed loop from effective data processing to the construction of fault prediction basis, effectively improving the systematicness and reliability of wind turbine sensor correlation analysis, and providing efficient and accurate technical support for the accurate prediction of wind turbine faults.
[0021] S3, Predictive Reference Sensor When forecasting operational data for a given period, retrieve the valid data for the current period (reference sensor). (and valid data corresponding to all other sensors), then recalculate the reference sensor. Correlation coefficient with each of the other sensors; Then, the judgment thresholds corresponding to other sensors are used. Judgment and reference sensor Other sensors that are related; Other related sensors are retrieved to form real-time candidate combinations, and the matching degree between these combinations and each standard associated combination is determined: Real-time candidate combinations are received. Standard association combination Calculate the corresponding matching degree: ,in For real-time candidate combinations Associated with standard Same number of sensors For real-time candidate combinations Associated with standard Count all sensors, remove duplicates; Select the combination of criteria with the highest matching degree, that is... This allows for the rapid identification of the most representative sensor correlation patterns under the current operating conditions, providing a reliable basis for subsequent operational data prediction. The selection period is associated with the combination based on the highest matching degree. Selecting a reference sensor Operating data from other related sensors during the selected time period, if the selected time period... Operating data of all related other sensors If both exist, then the operating data of other sensors will be used to predict the corresponding prediction period for the reference sensor. Operational data: Based on historical data, linear regression is used to calculate the weight and intercept of each other sensor. Then, the operational data of each other sensor is multiplied by its corresponding weight, and all products are summed and the intercept is added to obtain the benchmark sensor for the period to be predicted. Operational data; Sensing reference sensor The corresponding anomaly detection rules are based on the reference sensors during the period to be predicted. The running data is compared with the anomaly judgment rules in sequence. If there is running data that does not meet the anomaly judgment rules, an alarm signal is output. If you select a time period Operating data of all related other sensors If a data point is missing, the next selection period corresponding to the standard association combination is selected in descending order of matching degree until the operating data of all associated sensors and the corresponding sensors in the standard association combination are available. This solves the real-time problem of the association relationship caused by the dynamic changes in the operating conditions of wind turbine units. Furthermore, through accurate prediction and timely alarm, the accuracy and timeliness of fault prediction are greatly improved, effectively ensuring the efficient operation and maintenance and safe power generation of wind turbine units.
[0022] The weight of each of the other sensors mentioned above: Receiver reference sensor Historical operating data Other sensors Historical operating data And due to the reference sensor Other sensors ( There is a correlation between them, therefore the reference sensor Other sensors The two sides satisfy a linear relationship: ; The weights are solved using the least squares method. and intercept : ; in The number of historical data samples. and For the first Historical data for each sample.
[0023] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for predicting wind turbine faults based on multimodal feature fusion, characterized in that, Includes the following steps: S1. Select a reference sensor; Retrieve the effective time limit of the reference sensor and define it as the corresponding time interval; retrieve the operating data of the reference sensor and other sensors within the effective time limit in each time interval; Calculate the correlation coefficient between the reference sensor and other sensors at each time interval, set a judgment threshold to judge the reference sensor and other sensors that have a correlation relationship, and establish the initial correlation combination for each corresponding time interval; S2. Adjust the time interval to obtain multiple adjusted association combinations; calculate the similarity between the adjusted association combination and the corresponding initial association combination, and analyze the adjusted association combination and the initial association combination to obtain the same association set; The adjusted association combination with the highest similarity within the same association set is defined as the standard association combination, and the adjustment time period corresponding to the standard association combination is defined as the selection time period; Retrieve the standard association combination within each identical association set, select other identical sensors, and adjust the judgment threshold based on the minimum correlation coefficient between the other identical sensors and the benchmark sensor; S3. Predict the operating data of the reference sensor during the prediction period: Select the effective data of the reference sensor and other sensors in the current period, recalculate the correlation coefficient, and determine the real-time candidate combination of other sensors that are related to the reference sensor. Calculate the matching degree between real-time candidate combinations and standard associated combinations, select the standard associated combination with the highest matching degree, and predict the operating data of the benchmark sensor within the corresponding prediction period; sense the corresponding anomaly judgment rule, compare the predicted operating data with the anomaly judgment rule, and output an alarm signal if an anomaly is judged.
2. The wind turbine fault prediction method based on multimodal feature fusion according to claim 1, characterized in that: S1 receives historical maintenance data and sequentially retrieves the repair completion timestamp and fault occurrence timestamp corresponding to the same abnormal component. Multiple valid time limits are obtained by subtracting the fault occurrence time limit from the adjacent maintenance completion timestamp; the operational data detected by each sensor within each valid time limit is defined as valid data. Select a sensor used to detect component operation data as a reference sensor, and retrieve the valid data of the reference sensor within each valid time limit; Using the effective time limit of the reference sensor as the reference time period, retrieve the effective data of other sensors besides the reference sensor within the reference time period; Calculate the correlation coefficients between the reference sensor and other sensors in each effective time limit; Set a judgment threshold. If the correlation coefficient is greater than the judgment threshold, it is determined that there is a correlation between the reference sensor and other sensors within the effective time limit.
3. The wind turbine fault prediction method based on multimodal feature fusion according to claim 2, characterized in that: The correlation coefficient in S1 is defined as follows: the operating data of the reference sensor within the effective time limit is the first data sequence, and the operating data of other sensors within the corresponding effective time limit is the second data sequence. The numerator for calculating the correlation coefficient is obtained by successively subtracting the mean of the first data sequence from each running data point in the first data sequence, and subtracting the mean of the second data sequence from each running data point in the second data sequence, multiplying the resulting differences, and summing them. The denominator is obtained by multiplying the square root of the sum of squares of each data point in the first data sequence minus its mean by the square root of the sum of squares of each data point in the second data sequence minus its mean. Then, the correlation coefficient between the reference sensor and other sensors within the effective time limit is obtained by dividing the numerator by the denominator.
4. The wind turbine fault prediction method based on multimodal feature fusion according to claim 2, characterized in that: S2 receives multiple valid time limits from the abnormal component and retrieves the minimum time period; For the effective time limit greater than the minimum time period, a fixed step size is set, and the effective time limit greater than the minimum time period is gradually reduced as the adjustment period until the adjustment period equals the minimum time period. During the process of adjusting the effective time limit with a fixed step size, each adjustment reduces the running data within the effective time limit by one. When the effective time limit is gradually adjusted to be greater than the effective time limit of the minimum time period, multiple other sensors that are related to the reference sensor within the adjustment period are calculated in turn and constructed into corresponding adjustment association combinations. Each effective time limit corresponds to multiple adjustment association combinations.
5. The wind turbine fault prediction method based on multimodal feature fusion according to claim 4, characterized in that: S2 calculates and analyzes the changes in the adjustment association combination corresponding to each successively decreasing adjustment period. If the similarity between multiple adjustment association combinations and a certain initial association combination changes in an ordered manner, then it is determined that the initial association combination is the same as the initial association combination corresponding to multiple adjustment association combinations. After determining that they are the same, the next initial association combination is synchronously adjusted using a fixed step size. The next initial association combination is set as the initial association combination corresponding to the adjusted association combination, and the similarity is calculated and the judgment is continued until the similarity between the adjusted association combinations changes in an unordered manner. Then, the initial association combination when the unordered change occurs is retrieved, and the same judgment is repeated to determine whether they are the same, thereby merging multiple identical initial association combinations and their corresponding adjusted association combinations into an identical association set.
6. The wind turbine fault prediction method based on multimodal feature fusion according to claim 4, characterized in that: The adjustment of the association combination corresponding to the initial association combination is as follows: sort multiple valid time limits, select the largest time limit as the priority to decrease the time limit, and gradually decrease it to the second valid time limit that is second only to the largest time limit; When gradually reducing the maximum time limit to the second effective time limit, the initial association combination corresponding to the association combination is adjusted during the reduction process, which is the initial association combination corresponding to the second effective time limit; If the adjustment period corresponding to the maximum time limit equals the second effective time limit, then the second effective time limit is reduced simultaneously. At this time, the adjustment association combination corresponding to the maximum time limit and the second effective time limit is obtained. The initial association combination corresponding to the adjustment association combination is the initial association combination corresponding to the third effective time limit.
7. The wind turbine fault prediction method based on multimodal feature fusion according to claim 5, characterized in that: S2: retrieve the similarity between the initial association combination and each adjusted association combination in each identical association set, retrieve the adjusted association combination corresponding to the maximum similarity, and define it as the standard association combination. Each identical association set corresponds to a standard association combination. Filter other sensors that are the same in multiple standard association combinations, and select the smallest correlation coefficient between the same other sensor and the benchmark sensor as the threshold for judging whether there is a correlation; filter the adjustment period corresponding to the standard association combination, and define it as the selection period.
8. The wind turbine fault prediction method based on multimodal feature fusion according to claim 3, characterized in that: In step S3, when predicting the operating data of the reference sensor during the prediction period, the effective data in the current period is retrieved, and the correlation coefficient between the reference sensor and each other sensor is calculated again. Then, using the judgment thresholds corresponding to other sensors, other sensors that are related to the benchmark sensor are judged. Other related sensors are retrieved to form real-time candidate combinations. The matching degree between these combinations and each standard associated combination is then determined, and the standard associated combination with the highest matching degree is selected.
9. The wind turbine fault prediction method based on multimodal feature fusion according to claim 8, characterized in that: By using the highest matching degree standard associated combination corresponding to the selected time period, the operating data of other sensors that are associated with the benchmark sensor are selected during the selected time period. If the operating data of all associated other sensors exist within the selected time period, the operating data of other sensors are used to predict the operating data of the benchmark sensor within the corresponding prediction time period. The anomaly judgment rule corresponding to the sensing reference sensor is to compare the operating data of the reference sensor in the predicted period with the anomaly judgment rule in turn. If there is any operating data that does not meet the anomaly judgment rule, an alarm signal is output. If the operating data of all associated other sensors within the selected time period is missing, the next selected time period corresponding to the standard association combination is selected in descending order of matching degree, until the operating data of all associated sensors and the corresponding sensors in the standard association combination are available.
10. A wind turbine fault prediction system based on multimodal feature fusion, applied to the wind turbine fault prediction method based on multimodal feature fusion as described in any one of claims 1-9, characterized in that, include: Initial association combination construction module (1): Select a reference sensor, analyze multiple effective time limits corresponding to the reference sensor and define them as time intervals, select the running data in each time interval, calculate the correlation coefficient between the reference sensor and other sensors in each time interval, set a judgment threshold to judge the reference sensor and other sensors that have a correlation relationship, and construct the initial association combination for each corresponding time interval. The threshold adjustment module (2) adjusts the time interval to obtain multiple adjustment association combinations; calculates the similarity between the adjustment association combination and the corresponding initial association combination, analyzes the change of each adjustment association combination, judges multiple identical initial association combinations and their corresponding adjustment association combinations, and then merges them into the same association set; analyzes the standard association combination corresponding to the largest similar adjustment association combination in the same association set; retrieves the standard association combination in each same association set, selects the same other sensors, and adjusts the judgment threshold by the smallest correlation coefficient between the same other sensors and the benchmark sensor; And select the adjustment period corresponding to the standard associated combination, and define it as the selection period; The running data prediction output module (3) predicts the running data of the reference sensor in the prediction period: selects the effective data of the reference sensor and other sensors in the current period, recalculates the correlation coefficient, and uses the adjusted judgment threshold to judge other sensors that are related to the reference sensor, and then forms a real-time candidate combination. Calculate the matching degree between real-time candidate combinations and standard associated combinations, select the standard associated combination with the highest matching degree, and predict the operating data of the benchmark sensor within the corresponding prediction period; sense the corresponding anomaly judgment rule, compare the predicted operating data with the anomaly judgment rule, and output an alarm signal if an anomaly is judged.