Hydroelectric equipment rapid maintenance decision-making method based on big data

Abstract

The invention discloses a hydroelectric equipment rapid maintenance decision-making method based on big data, and relates to the technical field of hydroelectric equipment maintenance, and the method comprises the steps: obtaining a historical maintenance record of hydroelectric equipment, and uploading the historical maintenance record to a hydroelectric maintenance platform; determining key parameters of each piece of hydroelectric equipment according to the characteristic parameters of the library corresponding to each piece of hydroelectric equipment in the hydroelectric maintenance platform; a training period is set, and a dynamic weight distribution model is constructed based on data of key parameters and environment parameters in the training period; real-time monitoring and maintenance decision making are carried out based on a dynamic weight distribution model, manual intervention is not needed, the automation level of equipment operation and maintenance is improved, the workload of operation and maintenance personnel is relieved, subjectivity of experience-based judgment is avoided through a maintenance strategy generated through real-time health score and threshold judgment, a maintenance plan is more scientific and reasonable, and the maintenance efficiency is improved. The equipment utilization rate can be effectively balanced.

Description

Technical Field

[0001] This invention belongs to the field of hydropower equipment maintenance technology, and in particular relates to a rapid maintenance decision-making method for hydropower equipment based on big data. Background Technology

[0002] Existing power equipment maintenance decision-making methods are mostly based on single indicators or manually set fixed weights for evaluation. This approach can be applied in simple scenarios, but in complex scenarios involving multiple factors such as operating status, environmental conditions, historical records, and risk levels, fixed weights often lack flexibility and are prone to result deviations. Due to the lack of a dynamic adjustment mechanism, traditional methods are unable to adaptively optimize weights according to real-time changes in operating conditions and differences in task requirements, thus affecting the rationality and scientific nature of maintenance plans. In addition, manually set weights are highly subjective and difficult to guarantee consistency and objectivity. Therefore, it is necessary to design a rapid maintenance decision-making method for hydropower equipment based on big data to solve the above problems. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a rapid maintenance decision-making method for hydropower equipment based on big data. It aims to solve the problems of lack of flexibility under the fixed weight of the existing technology, which is prone to result deviation and lacks dynamic adjustment mechanism. It has the characteristics of intelligent fusion of multi-source factors and dynamic weight allocation, making maintenance decisions more targeted and flexible.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A rapid maintenance decision-making method for hydropower equipment based on big data includes the following steps: S1, retrieve historical maintenance records of hydropower equipment and upload them to the hydropower maintenance platform; S2, Based on the characteristic parameters of the corresponding database for each hydropower equipment in the hydropower maintenance platform, determine the key parameters of each hydropower equipment; S3 sets the training period and constructs a dynamic weight allocation model based on the data of key parameters and environmental parameters within the training period; S4, based on a dynamic weight allocation model, performs real-time monitoring and maintenance decisions.

[0005] Preferably, step S1 includes the following steps: Several operation monitoring devices and surrounding environment acquisition devices are deployed in each hydropower equipment. The monitored operation data and environmental data are synchronized to the hydropower maintenance platform and merged to generate the working data of each hydropower equipment. The maintenance cycle of the water and electricity equipment is preset. If the operating data of a certain water and electricity equipment does not meet the preset standard threshold range within the maintenance cycle, the operating data is set as abnormal data. The working data of all water and electricity equipment is obtained, and the executed maintenance plan is collected. The fault record of the water and electricity equipment is generated by combining the working data and the maintenance plan. The fault record is uploaded to the water and electricity maintenance platform to establish a data anomaly database and determine the characteristic parameters of the data anomaly database. If the operating data of all hydropower equipment meets the preset standard threshold range during the maintenance cycle, the maintenance results are collected during the maintenance process. If the maintenance results show that there is a hidden fault, the fault record of the hydropower equipment is generated and uploaded to the hydropower maintenance platform to establish a data degradation library and determine the characteristic parameters of the data degradation library. If a sudden shutdown of hydropower equipment occurs before the maintenance cycle has expired and no abnormal data has been triggered in advance, the fault record will be uploaded to the hydropower maintenance platform to establish a sudden database and determine the characteristic parameters of the sudden database.

[0006] Preferably, determining the characteristic parameters of the data anomaly database includes the following steps: In the data anomaly database, obtain the historical fault records of a certain hydropower equipment, collect the abnormal operating parameters of the hydropower equipment in a certain historical fault record, extract the monitoring values ​​of the abnormal operating parameters, and plot the operating change graph of the abnormal operating parameters with the values ​​as the vertical axis and time as the horizontal axis. In the operation change graph, a preset monitoring time period is set, the volatility of the operating parameters in each monitoring time period is calculated, a volatility threshold is set, and the monitoring time period exceeding the volatility threshold is set as the starting time period. The starting time period is used as the starting point, and the maintenance time point is used as the ending point to calculate the volatility change value. Summarize the volatility changes of a certain abnormal operating parameter, calculate the average volatility change of the abnormal operating parameter, and set the number of feature parameters in the data anomaly database to 'a'. Sort the abnormal operating parameters in descending order of their average volatility changes, and select the first 'a' abnormal operating parameters as feature parameters of the data anomaly database.

[0007] Preferably, determining the characteristic parameters of the data degradation library includes the following steps: In the data degradation database, historical fault records of a certain hydropower equipment are obtained, and the historical fault records are classified according to different latent faults to obtain a set of historical fault records for each latent fault. In a historical fault record set of a certain latent fault, the working data of the previous n maintenance cycles of a certain historical fault record are obtained. The monitoring value of a certain operating parameter in each maintenance cycle is extracted and the average monitoring value in the maintenance cycle is calculated. The average monitoring value of adjacent maintenance cycles is collected and the difference between the average monitoring values ​​is calculated. If the difference is less than 0, the absolute value of the difference is taken as the decrease. If the difference is greater than or equal to 0, the difference is set to zero. The decreases of all maintenance cycles are accumulated to obtain the cumulative decrease of the operating parameter. A degradation curve of the operating parameters is plotted. Based on the degradation curve, the average slope of the operating parameters is calculated. In the degradation curve, a preset window length of b is used, and a maintenance cycle of c is selected to obtain the maintenance cycle set of the front window as [cb, c-b+1, ..., c] and the maintenance cycle set of the back window as [c, c+1, ..., c+b]. The slopes of the front window and the back window are calculated. A preset inflection point slope threshold is set. If the slope of the back window is greater than the slope of the front window plus the inflection point slope threshold, then the maintenance cycle c is marked as an accelerated degradation inflection point. The number of accelerated degradation inflection points of the operating parameters is counted. The cumulative decrease, average slope, and number of accelerated degradation inflection points for each operating parameter are normalized. The normalized cumulative decrease, average slope, and number of accelerated degradation inflection points are then weighted and summed to obtain the degradation sensitivity score for each operating parameter. Step S135: Assign a value to each latent fault and calculate the overall degradation score according to the following formula: ; Where P represents the overall degradation score of the operating parameters, and A ie Let B represent the degradation sensitivity score of the operating parameters in the e-th historical fault record within the i-th latent fault. i Let represent the value of the i-th latent fault, gi represent the total number of historical fault records in the i-th latent fault, and r represent the total number of latent faults. The number of feature parameters in the preset data degradation library is f. The running parameters are sorted from largest to smallest according to the comprehensive degradation score, and the top f running parameters are selected as the feature parameters of the data degradation library.

[0008] Preferably, determining the characteristic parameters of the burst database includes the following steps: In the emergency database, a fault time window is preset, and historical fault records of a certain hydropower equipment are obtained. In each historical fault record, the operating data within the fault time window before and after the shutdown time is extracted as an analysis sample for the emergency shutdown. In the analysis sample, the average value of each running data point in the pre-fault window and post-fault window is calculated, and the relative and absolute mutation amplitudes are calculated according to the following formulas: ; Where Q1 represents the relative mutation magnitude, Q2 represents the absolute mutation magnitude, L1 represents the average value of the previous fault window, and L2 represents the average value of the subsequent fault window. Calculate the standard deviation of the operating data within the previous fault window, and divide the mean value within the previous fault window by the standard deviation to obtain the coefficient of variation. Then, calculate the stability index using the following formula: ; Where H represents the stability index, G represents the coefficient of variation, and G' represents the preset reference value for the coefficient of variation; For each fault type, the relative mutation amplitude, absolute mutation amplitude, and stability index of each operational data point are statistically analyzed and normalized. Simultaneously, each fault type is assigned a value, and the burst score of each operational data point is calculated using the following formula: ; Where K represents the burst score, Q1' nj Let Q2' be the normalized relative abrupt change magnitude of the j-th historical fault record in the n-th fault type. nj H' represents the normalized absolute abrupt change magnitude of the j-th historical fault record within the n-th fault type. nj Let F1, F2, and F3 represent the normalized stability index value of the j-th historical fault record in the n-th fault type, respectively. D represents the relative mutation amplitude, absolute mutation amplitude, and the weights of the stability index. n Let represent the weight of the nth fault type, hn represent the total number of historical fault records in the nth fault type, and m represent the total number of fault types. The number of feature parameters in the emergency database is set to u. The operating parameters are sorted from largest to smallest according to the emergency score, and the top u operating parameters are selected as the feature parameters of the emergency database.

[0009] Preferably, determining the key parameters of each hydroelectric device in step S2 includes the following steps: Obtain the data anomaly database, data degradation database, and burst database for a certain hydropower equipment, and extract the feature parameters from each database; The feature parameters in each database are intersected to obtain common feature parameters that exist simultaneously in the data anomaly database, data degradation database, and burst database. These common feature parameters are then identified as the key parameters of the hydropower equipment. If the common feature parameter is zero, the values ​​corresponding to the feature parameters in each library are normalized and assigned to each library. The key score is then calculated according to the following formula: ; Where M represents the key score, N x V represents the normalized feature parameter value in the x-th library. x Let y represent the weight of the x-th library, and y represent the total number of libraries. The number of key parameters is preset to s. The feature parameters are sorted from largest to smallest according to the key score, and the first s feature parameters are selected as key parameters.

[0010] Preferably, the construction of the dynamic weight allocation model in step S3 includes the following steps: Select several consecutive days as the training period, obtain data on key parameters and environmental parameters of each hydropower equipment within the training period, collect historical maintenance records of each hydropower equipment, preset a fixed window length of W and a step size of R, perform sliding window slicing on the time series, and generate a training sample set. The data in the training sample set is normalized to generate key parameter sequences and environmental parameter sequences. The historical maintenance records in the training sample set are used to construct sentence vectors through a language model. The sentence vectors are normalized to generate historical record feature vectors. Z days after the end of each window are used as prediction windows to obtain the fault status of water and electricity equipment within the prediction window, assign values, and generate labels. The key parameter sequence and environmental parameter sequence are mapped into a sequence vector matrix through an encoder, and the historical feature vectors are stacked to generate an input vector matrix. The input vector matrix is ​​then mapped through a multi-head attention mechanism to generate a query vector, a key vector, and a value vector. The relevance score is calculated through the dot product operation of the attention mechanism, and the relevance score is normalized through the Softmax function to obtain the attention weight. Finally, the attention weight is obtained by weighted summation with the value vector to obtain the fusion vector. The fused vector is input into the multilayer perceptron model, and the label is used as a supervision signal. During the training process, the multilayer perceptron model dynamically updates the weight parameters based on the loss function calculation results to construct a dynamic weight allocation model.

[0011] Preferably, step S4 includes the following steps: The real-time operating data and real-time environmental data of each hydropower device are obtained and input into the dynamic weight allocation model to calculate the real-time weight value and the real-time fault prediction probability. A preset fault prediction probability threshold is set. If the real-time fault prediction probability exceeds the threshold, a fault repair strategy is generated, published to the water and electricity maintenance platform, and staff are notified. If the real-time fault prediction probability does not exceed the threshold, a real-time health score is calculated by weighting the real-time weight value with the real-time operating data and real-time environmental data. A preset health score threshold is set. If the real-time health score is less than the health score threshold, the monitoring strategy continues to be executed. If the real-time health score is greater than or equal to the health score threshold, a preventive maintenance strategy is generated, published to the water and electricity maintenance platform, and staff are notified.

[0012] The beneficial effects of this invention are as follows: 1. By establishing a data anomaly database, a data degradation database, and a sudden incident database, this invention can cover abnormal situations of hydropower equipment under different operating scenarios, which is more comprehensive than traditional single data monitoring and improves the accuracy and completeness of fault identification.

[0013] 2. This invention employs a feature parameter intersection and weighted scoring mechanism to automatically select the key parameters that have the greatest impact on equipment health status, avoiding the subjectivity and limitations of existing technologies that rely on manual experience for parameter selection, thereby improving the model's reliability and generalization ability. By integrating operating parameters, environmental parameters, and historical maintenance records, and combining a multi-head attention mechanism to achieve adaptive weight allocation, the importance of different modal data can be dynamically adjusted, enhancing the model's ability to predict fault trends under complex operating conditions. A real-time health scoring system is constructed, which can generate a health index based on real-time monitoring data and automatically generate preventive maintenance strategies when thresholds are exceeded, effectively avoiding serious faults or downtime caused by delayed maintenance. Compared with traditional methods that rely on fixed maintenance cycles or reactive maintenance, this invention can provide early warning and rapid problem location, significantly shortening maintenance decision time, reducing unplanned downtime of hydropower equipment, and improving equipment operation safety. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0015] Example 1: like Figure 1 As shown, a rapid maintenance decision-making method for hydropower equipment based on big data is proposed, the method including: Step S100: Obtain historical maintenance records of hydropower equipment and upload them to the hydropower maintenance platform; Step S200: Determine the key parameters of each hydroelectric device based on the characteristic parameters of the corresponding database for each hydroelectric device in the hydroelectric maintenance platform; Step S300: Set the training period and construct a dynamic weight allocation model based on the data of key parameters and environmental parameters within the training period; Step S400: Real-time monitoring and maintenance decisions are made based on the dynamic weight allocation model.

[0016] Furthermore, step S100 includes: Step S110: Deploy several operation monitoring devices and surrounding environment acquisition devices in each hydropower equipment, synchronize the monitored operation data and environmental data to the hydropower maintenance platform, and merge the operation data and environmental data to generate the working data of each hydropower equipment; Step S120: Preset the maintenance cycle of water and electricity equipment. If the operating data of a certain water and electricity equipment does not meet the preset standard threshold range within the maintenance cycle, the operating data is set as abnormal data. Obtain the working data of all water and electricity equipment and collect the executed maintenance plan. Combine the working data and the maintenance plan to generate the fault record of the water and electricity equipment and upload the fault record to the water and electricity maintenance platform to establish a data anomaly database and determine the characteristic parameters of the data anomaly database. Step S130: If the operating data of all hydropower equipment meets the preset standard threshold range during the maintenance cycle, the maintenance results are collected during the maintenance process. If the maintenance results show that there is a hidden fault, the fault record of the hydropower equipment is generated and uploaded to the hydropower maintenance platform to establish a data degradation library and determine the characteristic parameters of the data degradation library. Step S140: If a sudden shutdown of hydropower equipment occurs before the maintenance cycle has arrived and no abnormal data has been triggered in advance, the fault record will be uploaded to the hydropower maintenance platform to establish a sudden database and determine the characteristic parameters of the sudden database. By deploying and operating monitoring equipment and environmental acquisition equipment in each hydropower unit, the operating status of the equipment and the surrounding environmental conditions can be collected synchronously and integrated to generate complete working data. This allows for a more comprehensive and accurate reflection of the actual operating status of the equipment, avoiding misjudgments caused by a single data dimension. During the maintenance cycle, if the equipment operation data is abnormal, the corresponding fault record will be generated in conjunction with the maintenance plan, and a data anomaly database and characteristic parameter model will be established. This will enable the summarization and classification of different types of anomalies, and improve the system's subsequent automatic diagnosis and predictive maintenance capabilities. During the maintenance process, if a hidden fault is found, even if the operating data does not exceed the threshold, it will be recorded and archived in the data degradation database. This helps to accumulate characteristic data of the degradation and evolution process, identify potential risks in advance, and thus achieve precise management of equipment life cycle and health status. If equipment suddenly shuts down during a non-maintenance period, the system will immediately upload and establish an emergency database. By summarizing the characteristic parameters of the emergency, a basis for emergency response and source tracing analysis can be formed, reducing downtime and improving the safety and reliability of power plant operation. By establishing a data anomaly database, a data degradation database, and a sudden incident database, a multi-dimensional database system covering three types of situations—anomalies, degradation, and sudden incidents—has been formed. Corresponding characteristic parameters have been determined, which can provide data support for optimizing maintenance plans and adjusting operational strategies.

[0017] Furthermore, the characteristic parameters for determining the data anomaly database in step S120 include: Step S121: In the data anomaly database, obtain the historical fault record of a certain hydropower equipment, collect the abnormal operating parameters of the hydropower equipment in a certain historical fault record, extract the monitoring value of the abnormal operating parameters, and plot the operating change graph of the abnormal operating parameters with the value as the vertical axis and time as the horizontal axis. Step S122: In the operation change graph, a preset monitoring time period is set, the volatility of the operating parameters in each monitoring time period is calculated, a volatility threshold is set, and the monitoring time period exceeding the volatility threshold is set as the starting time period. The starting time period is used as the starting point, and the maintenance time point is used as the ending point to calculate the volatility change value. Step S123: Summarize the volatility change values ​​of a certain abnormal operating parameter, calculate the average volatility change of the abnormal operating parameter, preset the number of feature parameters in the data anomaly database to be a, sort the abnormal operating parameters in descending order of average volatility change, and select the first a abnormal operating parameters as feature parameters of the data anomaly database. By plotting the values ​​of abnormal operating parameters over time as an operating change graph, the operating trend before and after the fault can be intuitively reflected, providing a reliable graphical basis for subsequent feature analysis. By setting a monitoring period and calculating the volatility, the significant fluctuation range of operating parameters before the occurrence of anomalies can be accurately captured. Compared with methods that rely solely on threshold judgment, this method can more sensitively and effectively identify potential fault signs. By averaging the volatility changes and sorting the parameters, the most representative outlier parameters can be automatically selected in a data-driven manner, ensuring that the characteristic parameters of the established data anomaly database are both representative and avoid redundant information interference.

[0018] Furthermore, the characteristic parameters of the data degradation library determined in step S130 include: Step S131: In the data degradation database, obtain the historical fault records of a certain hydropower equipment, and classify the historical fault records according to different latent faults to obtain a set of historical fault records for each latent fault. Step S132: In the historical fault record set of a certain latent fault, obtain the working data of the previous n maintenance cycles of a certain historical fault record, extract the monitoring value of a certain operating parameter in each maintenance cycle, and calculate the average monitoring value in the maintenance cycle. Collect the average monitoring value of adjacent maintenance cycles, calculate the difference between the average monitoring values. If the difference is less than 0, take the absolute value of the difference as the decrease. If the difference is greater than or equal to 0, take the difference as zero. Add up the decrease of all maintenance cycles to obtain the cumulative decrease of the operating parameter. Step S133: Plot the degradation curve of the operating parameters, and calculate the average slope of the operating parameters based on the degradation curve. In the degradation curve, the preset window length is b, and the maintenance cycle is selected as c. The maintenance cycle set of the front window is [cb, c-b+1, ..., c], and the maintenance cycle set of the back window is [c, c+1, ..., c+b]. The slope of the front window and the slope of the back window are calculated. A preset inflection point slope threshold is set. If the slope of the back window is greater than the slope of the front window plus the inflection point slope threshold, the maintenance cycle c is marked as an accelerated degradation inflection point. The number of accelerated degradation inflection points of the operating parameters is counted. Step S134: Normalize the cumulative decrease, average slope, and number of accelerated degradation inflection points for each operating parameter. Then, perform a weighted sum of the normalized cumulative decrease, average slope, and number of accelerated degradation inflection points to calculate the degradation sensitivity score for each operating parameter. Step S135: Assign a value to each latent fault and calculate the overall degradation score according to the following formula: ; Where P represents the overall degradation score of the operating parameters, and A ie Let B represent the degradation sensitivity score of the operating parameters in the e-th historical fault record within the i-th latent fault. i Let represent the value of the i-th latent fault, gi represent the total number of historical fault records in the i-th latent fault, and r represent the total number of latent faults. Step S136: The number of feature parameters in the preset data degradation library is f. Sort the running parameters from largest to smallest according to the comprehensive degradation score, and select the first f running parameters as the feature parameters of the data degradation library. By classifying historical fault records according to different latent faults, it is possible to achieve orderly management of latent fault data, avoid data mixing, and improve the pertinence and effectiveness of subsequent degradation feature analysis. By introducing a method for calculating the cumulative decrease, the performance degradation of operating parameters over multiple maintenance cycles can be quantitatively reflected. Compared with single-point data analysis, this method can more accurately reflect the overall trend of the degradation process. By plotting degradation curves and calculating average slopes, and by setting inflection point thresholds in conjunction with window slope differences, the key time points that accelerate the degradation of operating parameters can be automatically identified, providing a reliable basis for early warning and life prediction of equipment. By normalizing and weighting the cumulative decline, average slope, and number of accelerated degradation inflection points, a comprehensive quantification of the degradation sensitivity of different operating parameters can be achieved, avoiding bias caused by a single indicator and ensuring that the results are more scientific and reasonable. By assigning values ​​to each type of latent fault and calculating a comprehensive degradation score in conjunction with the degradation sensitivity score, we can not only reflect the degree of degradation of individual operating parameters, but also reveal the importance of latent faults in the overall degradation process. This enables a global quantitative evaluation from the parameter level to the fault level. By establishing a degradation sensitivity score and comprehensive degradation score system based on multi-dimensional indicators, we can more scientifically characterize the degradation patterns of latent faults and enhance the accuracy and reliability of equipment life prediction and degradation trend early warning.

[0019] Furthermore, the characteristic parameters of the burst database determined in step S140 include: Step S141: In the emergency database, a fault time window is preset, and historical fault records of a certain hydropower equipment are obtained. In each historical fault record, the operating data within the fault time window before and after the shutdown time is extracted as the analysis sample for the emergency shutdown. Step S142: In the analysis sample, calculate the average value of each running data in the previous fault window and the subsequent fault window, and calculate the relative mutation amplitude and absolute mutation amplitude according to the following formula: ; Where Q1 represents the relative mutation magnitude, Q2 represents the absolute mutation magnitude, L1 represents the average value of the previous fault window, and L2 represents the average value of the subsequent fault window. Step S143: Calculate the standard deviation of the running data within the previous fault window, and divide the average value within the previous fault window by the standard deviation to calculate the coefficient of variation. Calculate the stability index according to the following formula: ; Where H represents the stability index, G represents the coefficient of variation, and G' represents the preset reference value for the coefficient of variation; Step S144: Calculate the relative mutation amplitude, absolute mutation amplitude, and stability index of each operational data point for each fault type, and perform normalization calculations. Assign values ​​to each fault type and calculate the burst score for each operational data point according to the following formula: ; Where K represents the burst score, Q1' njLet Q2' be the normalized relative abrupt change magnitude of the j-th historical fault record in the n-th fault type. nj H' represents the normalized absolute abrupt change magnitude of the j-th historical fault record within the n-th fault type. nj Let F1, F2, and F3 represent the normalized stability index value of the j-th historical fault record in the n-th fault type, respectively. D represents the relative mutation amplitude, absolute mutation amplitude, and the weights of the stability index. n Let represent the weight of the nth fault type, hn represent the total number of historical fault records in the nth fault type, and m represent the total number of fault types. Step S145: Preset the number of feature parameters in the outbreak database to be u. Sort the running parameters according to the outbreak score from largest to smallest, and select the first u running parameters as feature parameters of the outbreak database. By setting a fault time window, it is possible to capture sudden changes in equipment operating parameters before and after shutdown, ensuring that the analysis samples are highly correlated with the sudden event and improving the relevance and effectiveness of the data; By calculating the relative and absolute magnitude of mutations, we can not only reflect the proportion of changes in operating parameters, but also characterize the absolute degree of change, thus achieving a comprehensive representation of the intensity of the impact of emergencies. By introducing a stability index and combining the coefficient of variation with a reference value, the fluctuation of operating data before a sudden event can be effectively assessed, thereby revealing the stability level of the equipment before a sudden shutdown. By normalizing the relative mutation amplitude, absolute mutation amplitude, and stability index, and combining them with weights to calculate the sudden score, a unified measurement can be achieved among multi-dimensional indicators, ensuring the comparability between different parameters and different fault types.

[0020] Furthermore, step S200 includes: Step S201: Obtain the data anomaly database, data degradation database, and burst database of a certain hydropower equipment, and extract the feature parameters from each database; Step S202: Perform an intersection operation on the feature parameters in each database to obtain common feature parameters that exist simultaneously in the data anomaly database, the data degradation database, and the burst database, and determine the common feature parameters as the key parameters of the hydropower equipment; Step S203: If the common feature parameter is zero, normalize the values ​​corresponding to the feature parameters in each library, assign values ​​to each library, and calculate the key score according to the following formula: ; Where M represents the key score, N x V represents the normalized feature parameter value in the x-th library.x Let y represent the weight of the x-th library, and y represent the total number of libraries. The number of key parameters is preset to s. The feature parameters are sorted from largest to smallest according to the key score, and the first s feature parameters are selected as key parameters. By performing intersection operations on the feature parameters of the three databases, we can filter out common parameters that are significant in all three types of features: anomaly, degradation, and burst, thus ensuring that the identified key parameters have the highest comprehensive representativeness and stability. For cases where the intersection is empty, the feature parameters of each library are normalized and key scores are calculated by combining weights. This achieves a mechanism that allows for the scientific selection of key parameters even when there are no common parameters, ensuring the integrity and applicability of the method. By normalizing and weighted summing the characteristic parameters of different databases, it is possible to integrate data from different sources and with different dimensions under a unified measurement system, avoiding the one-sidedness of single database analysis. Since the key parameters are the result of integrating multi-dimensional features of anomalies, degradation, and sudden changes, they are more sensitive and representative of the equipment operating status, thereby significantly improving the comprehensiveness and accuracy of fault diagnosis and operation early warning.

[0021] Furthermore, step S300 includes: Step S301: Select several consecutive days as the training period, obtain the data of key parameters and environmental parameters of each hydropower equipment within the training period, collect the historical maintenance records of each hydropower equipment, preset the fixed window length as W and the step size as R, perform sliding window slicing on the time series, and generate a training sample set; Step S302: Normalize the data in the training sample set to generate key parameter sequences and environmental parameter sequences. Use a language model to construct sentence vectors from the historical maintenance records in the training sample set. Normalize the sentence vectors to generate historical record feature vectors. Use Z days after the end of each window as a prediction window to obtain the fault status of water and electricity equipment within the prediction window, assign values, and generate labels. Step S303: Map the key parameter sequence and environmental parameter sequence into a sequence vector matrix through the encoder, and stack the historical feature vectors to generate an input vector matrix. Map each vector in the input vector matrix through a multi-head attention mechanism to generate a query vector, a key vector, and a value vector. Calculate the relevance score through the dot product operation of the attention mechanism, and normalize the relevance score through the Softmax function to obtain the attention weight. Finally, perform a weighted summation with the value vector to obtain the fusion vector. Step S304: Input the fused vector into the multilayer perceptron model and use the label as a supervision signal. During the training process, the multilayer perceptron model dynamically updates the weight parameters based on the loss function calculation results to construct a dynamic weight allocation model. By setting a fixed window length W and step size R to perform sliding window slicing on time series data, the temporal characteristics of key parameters and environmental parameters can be effectively preserved. At the same time, training sample sets are generated by combining historical maintenance records, making the training data comprehensive and structured. By normalizing the key parameter sequences and environmental parameter sequences, and introducing a language model to construct sentence vectors for maintenance records, unified quantification of numerical parameters and textual records was achieved, ensuring that multi-source heterogeneous data can be collaboratively modeled under the same framework. By setting a prediction window of Z days, clear labels can be generated based on the equipment failure status after the window, making the model training highly consistent with the actual operation scenario and improving the interpretability and practicality of the model prediction results. By introducing a multi-head attention mechanism, the importance of different modal features is dynamically modeled, enabling key parameters, environmental parameters, and historical data features to be adaptively weighted during the fusion process, thereby highlighting key influencing factors and avoiding information redundancy or bias. By inputting the fused vector into a multilayer perceptron and dynamically updating the weight parameters based on the supervision signal, the weight relationship between features can be continuously optimized to form a dynamic weight allocation model, thereby improving the accuracy and robustness of prediction.

[0022] Furthermore, step S400 includes: Step S401: Obtain real-time operating data and real-time environmental data for each hydropower device, and input the real-time operating data and real-time environmental data into the dynamic weight allocation model to calculate the real-time weight value and real-time fault prediction probability. Step S402: Preset a fault prediction probability threshold. If the real-time fault prediction probability exceeds the fault prediction probability threshold, generate a fault repair strategy, publish it to the water and electricity maintenance platform, and remind staff. If the real-time fault prediction probability does not exceed the fault prediction probability threshold, calculate the real-time health score by weighting and summing the real-time weight value with the real-time operating data and real-time environmental data. Step S403: Set a health score threshold. If the real-time health score is less than the health score threshold, continue to execute the monitoring strategy. If the real-time health score is greater than or equal to the health score threshold, generate a preventive maintenance strategy, publish it to the water and electricity maintenance platform, and remind the staff. By inputting real-time operating data and environmental data into the dynamic weight allocation model, the weights of each feature can be adaptively adjusted according to the real-time status of the device, thereby achieving a real-time and accurate assessment of the device's health status. The operational and environmental data are weighted and summed based on real-time weight values ​​to obtain an intuitive health score. This transforms the complex, multi-dimensional characteristics of equipment operation status into a single, measurable indicator, making it easier for maintenance personnel to understand and monitor. By setting health score thresholds, when the score reaches or exceeds the threshold, the system can proactively generate preventive maintenance strategies and push them to the maintenance platform, realizing the transformation from "post-event maintenance" to "pre-event prevention" and significantly reducing the rate of sudden equipment failures. This method can automatically complete data collection, scoring calculation and strategy generation without manual intervention, which improves the automation level of equipment operation and maintenance, reduces the workload of operation and maintenance personnel, and avoids the subjectivity of experience-based judgment by generating maintenance strategies through real-time health scoring and threshold determination, making maintenance plans more scientific and reasonable, and effectively balancing equipment utilization.

[0023] Example 2: This embodiment provides a rapid maintenance decision-making method for hydropower equipment based on big data, the method including: Step S100: Obtain historical maintenance records of hydropower equipment and upload them to the hydropower maintenance platform; Step S100 includes: Step S110: Deploy several operation monitoring devices and surrounding environment acquisition devices in each hydropower equipment, synchronize the monitored operation data and environmental data to the hydropower maintenance platform, and merge the operation data and environmental data to generate the working data of each hydropower equipment; Step S120: Preset the maintenance cycle of water and electricity equipment. If the operating data of a certain water and electricity equipment does not meet the preset standard threshold range within the maintenance cycle, the operating data is set as abnormal data. Obtain the working data of all water and electricity equipment and collect the executed maintenance plan. Combine the working data and the maintenance plan to generate the fault record of the water and electricity equipment and upload the fault record to the water and electricity maintenance platform to establish a data anomaly database and determine the characteristic parameters of the data anomaly database. The step S120, which involves determining the characteristic parameters of the data anomaly database, includes the following steps: Step S121: In the data anomaly database, obtain the historical fault record of a certain hydropower equipment, collect the abnormal operating parameters of the hydropower equipment in a certain historical fault record, extract the monitoring value of the abnormal operating parameters, and plot the operating change graph of the abnormal operating parameters with the value as the vertical axis and time as the horizontal axis. Step S122: In the operation change graph, a preset monitoring time period is set, the volatility of the operating parameters in each monitoring time period is calculated, a volatility threshold is set, and the monitoring time period exceeding the volatility threshold is set as the starting time period. The starting time period is used as the starting point, and the maintenance time point is used as the ending point to calculate the volatility change value. Step S123: Summarize the volatility change values ​​of a certain abnormal operating parameter, calculate the average volatility change of the abnormal operating parameter, preset the number of feature parameters in the data anomaly database to be a, sort the abnormal operating parameters in descending order of average volatility change, and select the first a abnormal operating parameters as feature parameters of the data anomaly database. For example, the hydroelectric equipment is a turbine generator set. The historical fault record shows that the stator winding temperature was abnormal, and the abnormality occurred on the 15th day. Therefore, the data from the 15 days are summarized. Assuming a monitoring period of 2 days, the volatility is calculated as the maximum value minus the minimum value, divided by the average value, and then multiplied by 100%. For the first monitoring period, the temperature range is 72-74°C, with an average of 73°C, resulting in a volatility of 2.7%. For the second monitoring period, the temperature range is 74-77°C, with an average of 75.5°C, resulting in a volatility of 4.0%. For the third monitoring period, the temperature range is 77-82°C, with an average of 79.5°C, resulting in a volatility of 6.3%. ... For the eighth monitoring period, the temperature range is 108-112°C, with an average of 110°C, resulting in a volatility of 3.6%. Assuming the threshold is 5%, the third monitoring period exceeds the limit. The starting time period is the third monitoring period, and the calculated volatility change value is 2.7%. The fluctuation rate of stator winding temperature is 2.7%, the fluctuation rate of bearing vibration is 4.5%, the fluctuation rate of cooling water pressure is 1.2%, the fluctuation rate of stator current is 3.6%, and the fluctuation rate of generator terminal voltage is 0.8%. Assuming a is 3, the characteristic parameters of the data anomaly database are bearing vibration, stator current, and stator winding temperature.

[0024] Step S130: If the operating data of all hydropower equipment meets the preset standard threshold range during the maintenance cycle, the maintenance results are collected during the maintenance process. If the maintenance results show that there is a hidden fault, the fault record of the hydropower equipment is generated and uploaded to the hydropower maintenance platform to establish a data degradation library and determine the characteristic parameters of the data degradation library. The step S130, which involves determining the characteristic parameters of the data degradation library, includes the following steps: Step S131: In the data degradation database, obtain the historical fault records of a certain hydropower equipment, and classify the historical fault records according to different latent faults to obtain a set of historical fault records for each latent fault. Step S132: In the historical fault record set of a certain latent fault, obtain the working data of the previous n maintenance cycles of a certain historical fault record, extract the monitoring value of a certain operating parameter in each maintenance cycle, and calculate the average monitoring value in the maintenance cycle. Collect the average monitoring value of adjacent maintenance cycles, calculate the difference between the average monitoring values. If the difference is less than 0, take the absolute value of the difference as the decrease. If the difference is greater than or equal to 0, take the difference as zero. Add up the decrease of all maintenance cycles to obtain the cumulative decrease of the operating parameter. Step S133: Plot the degradation curve of the operating parameters, and calculate the average slope of the operating parameters based on the degradation curve. In the degradation curve, the preset window length is b, and the maintenance cycle is selected as c. The maintenance cycle set of the front window is [cb, c-b+1, ..., c], and the maintenance cycle set of the back window is [c, c+1, ..., c+b]. The slope of the front window and the slope of the back window are calculated. A preset inflection point slope threshold is set. If the slope of the back window is greater than the slope of the front window plus the inflection point slope threshold, the maintenance cycle c is marked as an accelerated degradation inflection point. The number of accelerated degradation inflection points of the operating parameters is counted. Step S134: Normalize the cumulative decrease, average slope, and number of accelerated degradation inflection points for each operating parameter. Then, perform a weighted sum of the normalized cumulative decrease, average slope, and number of accelerated degradation inflection points to calculate the degradation sensitivity score for each operating parameter. Step S135: Assign a value to each latent fault and calculate the overall degradation score according to the following formula: ; Where P represents the overall degradation score of the operating parameters, and A ie Let B represent the degradation sensitivity score of the operating parameters in the e-th historical fault record within the i-th latent fault. i Let represent the value of the i-th latent fault, gi represent the total number of historical fault records in the i-th latent fault, and r represent the total number of latent faults. Step S136: The number of feature parameters in the preset data degradation library is f. Sort the running parameters from largest to smallest according to the comprehensive degradation score, and select the first f running parameters as the feature parameters of the data degradation library. For example, the types of latent faults in hydro-generator units include: Insulation aging; Bearing wear; Cooling system blockage; In the data degradation database, there are 40 historical fault records for insulation aging, 30 historical fault records for bearing wear, and 20 historical fault records for cooling system blockage. Taking the latent fault of insulation aging as an example, the stator insulation resistance parameter is taken, and monitoring data for the first n=5 maintenance cycles are obtained. The average value in the first maintenance cycle is 100, the average value in the second maintenance cycle is 95, the average value in the third maintenance cycle is 93, the average value in the fourth maintenance cycle is 91, and the average value in the fifth maintenance cycle is 88. The difference between adjacent cycles is calculated: Period 2 - Period 1: 95 - 100 = -5, decrease = 5; Period 3 - Period 2: 93 - 95 = -2, decrease = 2; Period 4 - Period 3: 91 - 93 = -2, decrease = 2; Period 5 - Period 4: 88 - 91 = -3, decrease = 3; The cumulative decrease is 5 + 2 + 2 + 3 = 12; Set the window length b=2 and the maintenance cycle c=3: Front window [1,2,3]: slope is -3.5; Back window [3,4,5]: slope is -2.5; Set the inflection point slope threshold to 1.0; -2.5 = -3.5 + 1.0, therefore no inflection point of accelerated degradation has appeared; Assuming the cumulative decrease in stator insulation resistance is normalized to 0.6, the average slope is normalized to 0.5, the number of inflection points is normalized to 0, and the weighted sum is 0.44; The cumulative decrease in bearing vibration after normalization is 0.8, the average slope after normalization is 0.7, the number of inflection points after normalization is 1, and the weighted sum is 0.78. The cumulative decrease in cooling water flow rate is normalized to 0.5, the average slope is normalized to 0.4, the number of inflection points is normalized to 0, and the weighted sum is 0.36. The calculated comprehensive degradation score for stator insulation resistance is 0.264, for bearing vibration it is 0.624, and for cooling water flow rate it is 0.18. With f set to 2, the characteristic parameters of the data degradation library are stator insulation resistance and bearing vibration.

[0025] Step S140: If a sudden shutdown of hydropower equipment occurs before the maintenance cycle has arrived and no abnormal data has been triggered in advance, the fault record will be uploaded to the hydropower maintenance platform to establish a sudden database and determine the characteristic parameters of the sudden database. Step S140, which involves determining the characteristic parameters of the burst database, includes the following steps: Step S141: In the emergency database, a fault time window is preset, and historical fault records of a certain hydropower equipment are obtained. In each historical fault record, the operating data within the fault time window before and after the shutdown time is extracted as the analysis sample for the emergency shutdown. Step S142: In the analysis sample, calculate the average value of each running data in the previous fault window and the subsequent fault window, and calculate the relative mutation amplitude and absolute mutation amplitude according to the following formula: ; Where Q1 represents the relative mutation magnitude, Q2 represents the absolute mutation magnitude, L1 represents the average value of the previous fault window, and L2 represents the average value of the subsequent fault window. Step S143: Calculate the standard deviation of the running data within the previous fault window, and divide the average value within the previous fault window by the standard deviation to calculate the coefficient of variation. Calculate the stability index according to the following formula: ; Where H represents the stability index, G represents the coefficient of variation, and G' represents the preset reference value for the coefficient of variation; Step S144: Calculate the relative mutation amplitude, absolute mutation amplitude, and stability index of each operational data point for each fault type, and perform normalization calculations. Assign values ​​to each fault type and calculate the burst score for each operational data point according to the following formula: ; Where K represents the burst score, Q1' nj Let Q2' be the normalized relative abrupt change magnitude of the j-th historical fault record in the n-th fault type. nj H' represents the normalized absolute abrupt change magnitude of the j-th historical fault record within the n-th fault type. nj Let F1, F2, and F3 represent the normalized stability index value of the j-th historical fault record in the n-th fault type, respectively. D represents the relative mutation amplitude, absolute mutation amplitude, and the weights of the stability index. n Let represent the weight of the nth fault type, hn represent the total number of historical fault records in the nth fault type, and m represent the total number of fault types. Step S145: Set the number of feature parameters in the emergency database to u. Sort the operating parameters according to the emergency score from largest to smallest, and select the top u operating parameters as feature parameters of the emergency database.

[0026] Step S200: Determine the key parameters of each hydroelectric device based on the characteristic parameters of the corresponding database for each hydroelectric device in the hydroelectric maintenance platform; Step S200 includes: Step S201: Obtain the data anomaly database, data degradation database, and burst database of a certain hydropower equipment, and extract the feature parameters from each database; Step S202: Perform an intersection operation on the feature parameters in each database to obtain common feature parameters that exist simultaneously in the data anomaly database, the data degradation database, and the burst database, and determine the common feature parameters as the key parameters of the hydropower equipment; Step S203: If the common feature parameter is zero, normalize the values ​​corresponding to the feature parameters in each library, assign values ​​to each library, and calculate the key score according to the following formula: ; Where M represents the key score, N x V represents the normalized feature parameter value in the x-th library. x Let y represent the weight of the x-th library, and y represent the total number of libraries. The number of key parameters is preset to s. The feature parameters are sorted from largest to smallest according to the key score, and the first s feature parameters are selected as key parameters. For example, the characteristic parameters of the data degradation library are stator insulation resistance and bearing vibration; The characteristic parameters of the data anomaly database are bearing vibration, stator current, and stator winding temperature; The characteristic parameters of the burst database are bearing vibration and stator current; The intersection is bearing vibration, therefore the key parameter of the hydro-generator unit is bearing vibration.

[0027] Step S300: Set the training period and construct a dynamic weight allocation model based on the data of key parameters and environmental parameters within the training period; Step S300 includes: Step S301: Select several consecutive days as the training period, obtain the data of key parameters and environmental parameters of each hydropower equipment within the training period, collect the historical maintenance records of each hydropower equipment, preset the fixed window length as W and the step size as R, perform sliding window slicing on the time series, and generate a training sample set; Step S302: Normalize the data in the training sample set to generate key parameter sequences and environmental parameter sequences. Use a language model to construct sentence vectors from the historical maintenance records in the training sample set. Normalize the sentence vectors to generate historical record feature vectors. Use Z days after the end of each window as a prediction window to obtain the fault status of water and electricity equipment within the prediction window, assign values, and generate labels. Step S303: Map the key parameter sequence and environmental parameter sequence into a sequence vector matrix through the encoder, and stack the historical feature vectors to generate an input vector matrix. Map each vector in the input vector matrix through a multi-head attention mechanism to generate a query vector, a key vector, and a value vector. Calculate the relevance score through the dot product operation of the attention mechanism, and normalize the relevance score through the Softmax function to obtain the attention weight. Finally, perform a weighted summation with the value vector to obtain the fusion vector. Step S304: Input the fused vector into the multilayer perceptron model and use the label as a supervision signal. During the training process, the multilayer perceptron model dynamically updates the weight parameters based on the loss function calculation results to construct a dynamic weight allocation model. For example, select 60 consecutive days of historical data as the sample pool. Sliding window parameters: window length is 1 day, step size is 0.5 days, prediction window Z is 7 days; Key parameter sequence: bearing vibration; Environmental parameter sequence: inlet water flow rate, unit load, inlet water temperature, ambient temperature, oil temperature, etc.; Historical maintenance records: maintenance logs for each unit, using language models to generate sentence vectors; Normalization: Perform min-max normalization on the sequence according to its features; The key parameter sequence and the environment sequence are encoded into sequence vector matrices using the time series encoder TransformerEncode; The historical maintenance sentence vector is normalized once through a fully connected layer and then stacked with the sequence vector using a tile. Assume the key parameter sequence is as follows: The environmental parameter sequence is The historical feature vector is The input matrix obtained by stacking is ; In the multi-head attention mechanism, there are three types of vectors: query vector, key vector, and value vector. These three are obtained by matrix multiplication of the input data. The parameter matrices of the query vector, key vector, and value vector are optimized through backpropagation during training. Specifically, the parameter matrices are randomly initialized and automatically updated through backpropagation and gradient descent during training to learn appropriate values ​​and generate the final parameter matrices. Suppose that the parameter matrix of the query vector is The parameter matrix of the key vector is The parameter matrix of the value vector is , The input matrix is ​​multiplied by the parameter matrix of each vector to obtain the query vector. The key vector is The value vector is ; The correlation score is calculated using the following formula: The correlation scores for the first row are calculated as [0, 0.707, 0.707, 0.707], the second row as [0.707, 0, 0.707, 0.707], the third row as [-0.707, 0.707, 0, 0], and the fourth row as [0, 0.707, 0.707, 0.707]. This generates a correlation score matrix. Normalized using the Softmax function, the calculated values ​​are: row 1 [0.163, 0.279, 0.279, 0.279], row 2 [0.279, 0.163, 0.279, 0.279], row 3 [0.163, 0.442, 0.197, 0.197], and row 4 [0.163, 0.279, 0.279, 0.279]. The fusion vector is obtained by weighted summation of the value vectors. ; The fused vector is pooled and then fed into a multilayer perceptron to output the fault probability.

[0028] Step S400: Real-time monitoring and maintenance decision-making based on the dynamic weight allocation model; Step S400 includes: Step S401: Obtain real-time operating data and real-time environmental data for each hydropower device, and input the real-time operating data and real-time environmental data into the dynamic weight allocation model to calculate the real-time weight value and real-time fault prediction probability. Step S402: Preset a fault prediction probability threshold. If the real-time fault prediction probability exceeds the fault prediction probability threshold, generate a fault repair strategy, publish it to the water and electricity maintenance platform, and remind staff. If the real-time fault prediction probability does not exceed the fault prediction probability threshold, calculate the real-time health score by weighting and summing the real-time weight value with the real-time operating data and real-time environmental data. Step S403: Set a health score threshold. If the real-time health score is less than the health score threshold, continue to execute the monitoring strategy. If the real-time health score is greater than or equal to the health score threshold, generate a preventive maintenance strategy, publish it to the water and electricity maintenance platform, and remind the staff.

[0029] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A rapid maintenance decision-making method for hydropower equipment based on big data, characterized in that, Includes the following steps: S1, retrieve historical maintenance records of hydropower equipment and upload them to the hydropower maintenance platform; S2, Based on the characteristic parameters of the corresponding database for each hydropower equipment in the hydropower maintenance platform, determine the key parameters of each hydropower equipment; S3 sets the training period and constructs a dynamic weight allocation model based on the data of key parameters and environmental parameters within the training period; S4, based on a dynamic weight allocation model, performs real-time monitoring and maintenance decisions.

2. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 1, characterized in that, Step S1 includes the following steps: Several operation monitoring devices and surrounding environment acquisition devices are deployed in each hydropower equipment. The monitored operation data and environmental data are synchronized to the hydropower maintenance platform and merged to generate the working data of each hydropower equipment. The maintenance cycle of the water and electricity equipment is preset. If the operating data of a certain water and electricity equipment does not meet the preset standard threshold range within the maintenance cycle, the operating data is set as abnormal data. The working data of all water and electricity equipment is obtained, and the executed maintenance plan is collected. The fault record of the water and electricity equipment is generated by combining the working data and the maintenance plan. The fault record is uploaded to the water and electricity maintenance platform to establish a data anomaly database and determine the characteristic parameters of the data anomaly database. If the operating data of all hydropower equipment meets the preset standard threshold range during the maintenance cycle, the maintenance results are collected during the maintenance process. If the maintenance results show that there is a hidden fault, the fault record of the hydropower equipment is generated and uploaded to the hydropower maintenance platform to establish a data degradation library and determine the characteristic parameters of the data degradation library. If a sudden shutdown of hydropower equipment occurs before the maintenance cycle has expired and no abnormal data has been triggered in advance, the fault record will be uploaded to the hydropower maintenance platform to establish a sudden database and determine the characteristic parameters of the sudden database.

3. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 2, characterized in that, The determination of the characteristic parameters of the data anomaly database includes the following steps: In the data anomaly database, obtain the historical fault records of a certain hydropower equipment, collect the abnormal operating parameters of the hydropower equipment in a certain historical fault record, extract the monitoring values ​​of the abnormal operating parameters, and plot the operating change graph of the abnormal operating parameters with the values ​​as the vertical axis and time as the horizontal axis. In the operation change graph, a preset monitoring time period is set, the volatility of the operating parameters in each monitoring time period is calculated, a volatility threshold is set, and the monitoring time period exceeding the volatility threshold is set as the starting time period. The starting time period is used as the starting point, and the maintenance time point is used as the ending point to calculate the volatility change value. Summarize the volatility changes of a certain abnormal operating parameter, calculate the average volatility change of the abnormal operating parameter, and set the number of feature parameters in the data anomaly database to 'a'. Sort the abnormal operating parameters in descending order of their average volatility changes, and select the first 'a' abnormal operating parameters as feature parameters of the data anomaly database.

4. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 2, characterized in that, The determination of the characteristic parameters of the data degradation library includes the following steps: In the data degradation database, historical fault records of a certain hydropower equipment are obtained, and the historical fault records are classified according to different latent faults to obtain a set of historical fault records for each latent fault. In a historical fault record set of a certain latent fault, the working data of the previous n maintenance cycles of a certain historical fault record are obtained. The monitoring value of a certain operating parameter in each maintenance cycle is extracted, and the average monitoring value in the maintenance cycle is calculated. The average monitoring value of adjacent maintenance cycles is collected, and the difference between the average monitoring values ​​is calculated. If the difference is less than 0, the absolute value of the difference is taken as the decrease. If the difference is greater than or equal to 0, the difference is set to zero. The decreases of all maintenance cycles are accumulated to obtain the cumulative decrease of the operating parameter.

5. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 4, characterized in that, The determination of the characteristic parameters of the data degradation library also includes the following steps: A degradation curve of the operating parameters is plotted. Based on the degradation curve, the average slope of the operating parameters is calculated. In the degradation curve, a preset window length of b is used, and a maintenance cycle of c is selected to obtain the maintenance cycle set of the front window as [cb, c-b+1, ..., c] and the maintenance cycle set of the back window as [c, c+1, ..., c+b]. The slopes of the front window and the back window are calculated. A preset inflection point slope threshold is set. If the slope of the back window is greater than the slope of the front window plus the inflection point slope threshold, then the maintenance cycle c is marked as an accelerated degradation inflection point. The number of accelerated degradation inflection points of the operating parameters is counted. The cumulative decrease, average slope, and number of accelerated degradation inflection points for each operating parameter are normalized. The normalized cumulative decrease, average slope, and number of accelerated degradation inflection points are then weighted and summed to obtain the degradation sensitivity score for each operating parameter.

6. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 5, characterized in that, The determination of the characteristic parameters of the data degradation library also includes the following steps: Assign a value to each latent fault and calculate the overall degradation score according to the following formula: ； Where P represents the overall degradation score of the operating parameters, and A ie Let B represent the degradation sensitivity score of the operating parameters in the e-th historical fault record within the i-th latent fault. i Let represent the value of the i-th latent fault, gi represent the total number of historical fault records in the i-th latent fault, and r represent the total number of latent faults. The number of feature parameters in the preset data degradation library is f. The running parameters are sorted from largest to smallest according to the comprehensive degradation score, and the top f running parameters are selected as the feature parameters of the data degradation library.

7. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 2, characterized in that, Determining the characteristic parameters of the burst database includes the following steps: In the emergency database, a fault time window is preset, and historical fault records of a certain hydropower equipment are obtained. In each historical fault record, the operating data within the fault time window before and after the shutdown time is extracted as an analysis sample for the emergency shutdown. In the analysis sample, the average value of each running data point in the pre-fault window and post-fault window is calculated, and the relative and absolute mutation amplitudes are calculated according to the following formulas: ； Where Q1 represents the relative mutation magnitude, Q2 represents the absolute mutation magnitude, L1 represents the average value of the previous fault window, and L2 represents the average value of the subsequent fault window. Calculate the standard deviation of the operating data within the previous fault window, and divide the mean value within the previous fault window by the standard deviation to obtain the coefficient of variation. Then, calculate the stability index using the following formula: ； Where H represents the stability index, G represents the coefficient of variation, and G' represents the preset reference value for the coefficient of variation; For each fault type, the relative mutation amplitude, absolute mutation amplitude, and stability index of each operational data point are statistically analyzed and normalized. Simultaneously, each fault type is assigned a value, and the burst score of each operational data point is calculated using the following formula: ； Where K represents the burst score, Q1' nj Let Q2' be the normalized relative abrupt change magnitude of the j-th historical fault record in the n-th fault type. nj H' represents the normalized absolute abrupt change magnitude of the j-th historical fault record within the n-th fault type. nj Let F1, F2, and F3 represent the normalized stability index value of the j-th historical fault record in the n-th fault type, respectively. D represents the relative mutation amplitude, absolute mutation amplitude, and the weights of the stability index. n Let represent the weight of the nth fault type, hn represent the total number of historical fault records in the nth fault type, and m represent the total number of fault types. The number of feature parameters in the emergency database is set to u. The operating parameters are sorted from largest to smallest according to the emergency score, and the top u operating parameters are selected as the feature parameters of the emergency database.

8. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 1, characterized in that, The determination of key parameters for each hydroelectric device in step S2 includes the following steps: Obtain the data anomaly database, data degradation database, and burst database for a certain hydropower equipment, and extract the feature parameters from each database; The feature parameters in each database are intersected to obtain common feature parameters that exist simultaneously in the data anomaly database, data degradation database, and burst database. These common feature parameters are then identified as the key parameters of the hydropower equipment. If the common feature parameter is zero, the values ​​corresponding to the feature parameters in each library are normalized and assigned to each library. The key score is then calculated according to the following formula: ； Where M represents the key score, N x V represents the normalized feature parameter value in the x-th library. x Let y represent the weight of the x-th library, and y represent the total number of libraries. The number of key parameters is preset to s. The feature parameters are sorted from largest to smallest according to the key score, and the first s feature parameters are selected as key parameters.

9. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 8, characterized in that, The construction of the dynamic weight allocation model in step S3 includes the following steps: Select several consecutive days as the training period, obtain data on key parameters and environmental parameters of each hydropower equipment within the training period, collect historical maintenance records of each hydropower equipment, preset a fixed window length of W and a step size of R, perform sliding window slicing on the time series, and generate a training sample set. The data in the training sample set is normalized to generate key parameter sequences and environmental parameter sequences. The historical maintenance records in the training sample set are used to construct sentence vectors through a language model. The sentence vectors are normalized to generate historical record feature vectors. Z days after the end of each window are used as prediction windows to obtain the fault status of water and electricity equipment within the prediction window, assign values, and generate labels. The key parameter sequence and environmental parameter sequence are mapped into a sequence vector matrix through an encoder, and the historical feature vectors are stacked to generate an input vector matrix. The input vector matrix is ​​then mapped through a multi-head attention mechanism to generate a query vector, a key vector, and a value vector. The relevance score is calculated through the dot product operation of the attention mechanism, and the relevance score is normalized through the Softmax function to obtain the attention weight. Finally, the attention weight is obtained by weighted summation with the value vector to obtain the fusion vector. The fused vector is input into the multilayer perceptron model, and the label is used as a supervision signal. During the training process, the multilayer perceptron model dynamically updates the weight parameters based on the loss function calculation results to construct a dynamic weight allocation model.

10. The rapid maintenance decision-making method for hydropower equipment based on big data according to claim 9, characterized in that, Step S4 includes the following steps: The real-time operating data and real-time environmental data of each hydropower device are obtained and input into the dynamic weight allocation model to calculate the real-time weight value and the real-time fault prediction probability. A preset fault prediction probability threshold is set. If the real-time fault prediction probability exceeds the threshold, a fault repair strategy is generated, published to the water and electricity maintenance platform, and staff are notified. If the real-time fault prediction probability does not exceed the threshold, a real-time health score is calculated by weighting the real-time weight value with the real-time operating data and real-time environmental data. A preset health score threshold is set. If the real-time health score is less than the health score threshold, the monitoring strategy continues to be executed. If the real-time health score is greater than or equal to the health score threshold, a preventive maintenance strategy is generated, published to the water and electricity maintenance platform, and staff are notified.

Images