Hydroelectric generating unit condition-based maintenance system and method based on multi-source situation awareness

Abstract

This invention discloses a condition-based maintenance system and method for hydropower units based on multi-source situational awareness, applicable to the field of hydropower unit condition monitoring technology. This invention performs data quality identification on multi-source condition data, identifies the operating condition switching boundaries of hydropower units, and sets different response stabilization waiting times for the response characteristics of measurement points corresponding to different types of condition quantities. It establishes a transition isolation window matching the response characteristics of measurement points for the operating condition switching boundaries, forms a stable sample candidate set outside the transition isolation window, and performs sample purification to generate a health benchmark. Based on the health benchmark, it performs evidence classification to generate an evidence set, forming sample deviations, gradual variables, sudden changes, and coupled anomalies as evidence objects and constructing an abnormal evidence chain. Feedback updates are performed based on the maintenance execution results. This helps to solve the problem that the determination of maintenance objects during condition-based maintenance of hydropower units is easily interfered with by operating condition switching data, making it difficult to accurately distinguish the state fluctuations caused by operating condition disturbances.

Description

Technical Field

[0001] This invention relates to the field of hydropower unit condition monitoring technology, and in particular to a hydropower unit condition-based maintenance system and method based on multi-source situational awareness. Background Technology

[0002] With the increasing complexity of hydropower unit operating conditions and the extension of continuous operating time, traditional maintenance methods relying on fixed-cycle shutdowns for inspection or judgment based on single monitoring parameters are no longer sufficient to accurately reflect the real-time deterioration status of key components. To reduce the risk of sudden failures and improve the relevance of maintenance plans, existing technologies are increasingly adopting multi-source situational awareness methods for comprehensive monitoring and analysis of hydropower unit operating status. This involves collecting various types of status data generated during unit operation to build a status identification foundation for equipment health assessment and maintenance decision-making. Based on this, existing maintenance methods typically include the following: During the operation of the hydropower unit, multi-source operational data, including vibration, temperature, pressure, speed, oscillation, oil condition, and historical maintenance records, are collected. The data from different sources are then time-aligned, format-converted, anomaly-removed, and feature-extracted to form condition-aware data that characterizes the unit's operating status. Subsequently, combined with preset condition evaluation rules, fault diagnosis models, or operational experience thresholds, the health status of key components such as the turbine, generator, bearings, speed control system, and auxiliary equipment is comprehensively assessed to identify whether the unit exhibits abnormal vibration, abnormal temperature rise, abnormal pressure fluctuations, decreased efficiency, or component deterioration trends. Based on this, the corresponding risk level, fault type, and maintenance priority are determined according to the condition assessment results, and maintenance suggestions such as routine inspections, shutdown checks, component replacement, or parameter adjustments are generated to achieve a shift from periodic maintenance to condition-based maintenance.

[0003] In existing technologies, some solutions have been researched regarding the health status evaluation, maintenance strategy optimization, and maintenance level recommendation of hydropower units. For example, Chinese invention patent application CN115577808B discloses a method and system for evaluating the health status of hydropower units based on multi-source information fusion, which includes: constructing a hierarchical analysis structure for evaluating the health status of the unit; determining the weights of each layer and performing consistency checks; determining the state classification thresholds for each layer; constructing index scoring standards by combining fuzzy comprehensive evaluation method and industry standards; acquiring multi-source information on the unit status through a data acquisition module; calculating and marking anomalies at the index layer; calculating and marking anomalies at the project layer; calculating and marking anomalies at the system layer; and providing overall unit status evaluation and operation and maintenance guidance suggestions. Another example is Chinese invention patent application CN121352762A, which discloses a method, equipment, medium, and product for optimizing the maintenance strategy of a hydropower unit speed control system, including: based on... This method can calculate state-based reliability using degraded data; calculate preventative maintenance reliability based on historical operation and maintenance data; determine the optimal maintenance method for the speed control system based on reliability; and determine the optimal maintenance interval by combining the determined optimal maintenance method with power plant maintenance management evaluation and component importance. Another example is a method, device, and storage medium for intelligent recommendation of maintenance levels for hydropower units disclosed in Chinese invention patent application CN119090292A, which includes: creating a maintenance item database; generating a preliminary maintenance item table for each unit based on its remaining defects, technical upgrade plans, and maintenance plans; importing the maximum number of units that can be or are planned to be maintained throughout the plant, counting them sequentially according to different maintenance levels to obtain the number of units at different maintenance levels; performing a state score on each unit; and intelligently recommending the maintenance level for each unit based on the number of units planned for maintenance throughout the plant, maintenance items, state score results, and the last maintenance completion time.

[0004] However, in the actual operation of hydropower units, the switching processes of operating conditions such as start-up and shutdown, variable load operation, guide vane opening adjustment, and crossing vibration-sensitive areas are relatively frequent. The sampling frequency, response speed, and settling time corresponding to different types of measuring points are not consistent. Near the boundary of operating condition switching, multi-source state data can easily contain normal transitional fluctuations, measuring point response lags, and real abnormal changes in equipment. If the data of this stage is still directly included in the construction of health samples or dynamic threshold calculation, the health benchmark is easily contaminated by transitional data, making subsequent sample deviation judgment, abnormal trend identification, and dynamic threshold judgment lack reliable basis.

[0005] Meanwhile, abnormal states of hydropower units are not always manifested as instantaneous exceedances at a single measuring point, but may simultaneously exhibit multiple forms of evidence, such as slow increases in vibration, sudden changes in pressure pulsations, delayed temperature increases, and abnormal coupling between swing and load changes. Existing comprehensive scoring or single health index evaluation methods typically compress abnormal information from different measuring points, time scales, and component sources into a uniform score, making it difficult to retain the formation time, change type, measuring point attribution, and component correlation of abnormal evidence. This results in unclear correspondences between abnormal evidence and target components, candidate fault types, and maintenance levels, and a lack of traceability in the maintenance decision-making process. If transitional state data at the boundary of operating condition switching is not effectively identified and isolated, the constructed health benchmark itself may be biased. Consequently, at the boundary of operating condition switching such as hydropower unit start-up and shutdown, load changes, and crossing vibration-sensitive areas, there is difficulty in distinguishing between normal operating condition transition responses and abnormal equipment responses, inconsistent response times of multiple source measuring points, and unclear health sample boundaries. This makes it easy for the identification of maintenance targets during condition-based maintenance of hydropower units to be interfered with by operating condition switching data, making it difficult to accurately distinguish state fluctuations caused by operating condition disturbances. Summary of the Invention

[0006] This invention provides a condition-based maintenance system and method for hydropower units based on multi-source situational awareness: Firstly, a condition-based maintenance system for hydropower units based on multi-source situational awareness is provided, including: a multi-source data access module, used to acquire multi-source status data, operating condition data, and maintenance-related data during the operation of the hydropower unit; establishing status data objects based on the multi-source status data; performing data quality identification on the multi-source status data; obtaining objectified status data with quality identification; an operating condition boundary identification module, used to identify the operating condition switching boundaries of the hydropower unit based on the operating condition data; setting different response stabilization waiting times for different types of state variables and corresponding measurement point response characteristics; and a health benchmark purification module, used to establish a transition isolation window matching the measurement point response characteristics for the operating condition switching boundaries; selecting objectified status data outside the transition isolation window to form a stable sample candidate set; and purifying the stable sample candidate set. The selection process involves sample purification to generate health benchmarks associated with corresponding quality labels, operating condition ranges, corresponding components, and corresponding measurement point types. Within a transitional isolation window, data within that window is prohibited from entering the health sample. The evidence classification module classifies evidence based on the real-time status values ​​corresponding to the health benchmarks to generate an evidence set, including sample deviation, gradual variables, sudden changes, and coupling anomalies. The evidence chain mapping module maps sample deviation, gradual variables, sudden changes, and coupling anomalies into evidence objects and constructs an abnormal evidence chain, determining the condition-based maintenance decision based on the abnormal evidence chain. The feedback update module obtains the maintenance execution results after the condition-based maintenance decision is implemented and updates the operating condition health benchmarks, transitional isolation windows, and abnormal evidence discrimination rules based on the maintenance execution results.

[0007] Secondly, the condition-based maintenance method for hydropower units based on multi-source situational awareness includes: acquiring multi-source status data, operating condition data, and maintenance-related data during the operation of the hydropower unit; establishing status data objects based on the multi-source status data; performing data quality labeling on the multi-source status data to obtain objectified status data with quality labels; identifying the operating condition switching boundaries of the hydropower unit based on the operating condition data; setting different response stabilization waiting times for different types of status variables based on the response characteristics of the measuring points; establishing a transition isolation window that matches the response characteristics of the measuring points for the operating condition switching boundaries; selecting objectified status data outside the transition isolation window to form a stable sample candidate set; and performing quality control on the stable sample candidate set. Sample purification involves generating health benchmarks associated with corresponding quality labels, operating condition ranges, corresponding components, and corresponding measurement point types. Within the transition isolation window, data within the window is prohibited from entering the healthy sample. Evidence is categorized based on the real-time status values ​​corresponding to the health benchmarks to generate an evidence set, including sample deviation, gradual variables, sudden changes, and coupling anomalies. Sample deviation, gradual variables, sudden changes, and coupling anomalies are formed into evidence objects, and an abnormal evidence chain is constructed. The condition-based maintenance decision is determined based on the abnormal evidence chain. The maintenance execution results after the condition-based maintenance decision are executed are obtained, and the operating condition health benchmarks, transition isolation windows, and abnormal evidence discrimination rules are updated based on the maintenance execution results.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By objectifying and quality-labeling data from different monitoring systems through a multi-source data access module, it helps avoid low-quality data such as missing data, communication anomalies, and sensor anomalies from directly participating in health assessments, thus improving the reliability of basic data. 2. By establishing transition isolation windows during operating condition switching phases such as start-up / shutdown, load changes, guide vane opening adjustments, and crossing vibration-sensitive zones through a condition boundary identification module and a health benchmark purification module, it prevents normal operating condition fluctuations from contaminating health samples, making the health benchmark more reflective of stable operating conditions. 3. By classifying abnormal states into sample deviation, gradual change, abrupt change, and coupled anomalies through an evidence classification module, it preserves the anomaly formation process, change type, and measurement point source, helping to avoid the compression of anomaly information by traditional single comprehensive scoring. 4. By binding the anomaly evidence chain to specific components, candidate faults, and maintenance items through an evidence chain mapping module, it improves the accuracy of maintenance object location and the traceability of maintenance decisions. 5. By using the feedback update module to back-update the health benchmark, transition isolation window, and anomaly evidence discrimination rules with actual maintenance results, it reduces the risk of subsequent false alarms and missed alarms, enabling the hydropower unit condition-based maintenance system to have continuous optimization capabilities.

[0009] 2. By acquiring operating data and analyzing changes in head, load, guide vane opening, and rotational speed of the hydropower unit, the operating condition switching boundaries during the unit's operation are identified. Furthermore, the fluctuation decay time of the same measuring point before and after a preset number of operating condition switches is statistically analyzed. The time required for the state change amplitude to recover to a preset stable fluctuation range is taken as the response stabilization waiting time for that measuring point. Its advantage over existing methods is that it simultaneously considers the timing of the operating condition switch and the measuring point's own response characteristics, avoiding the limitations of existing methods that rely solely on fixed times or uniform empirical rules to determine the operating condition transition phase. Since different measuring points experience different recovery times after start-up, load changes, guide vane opening adjustments, or crossing vibration-sensitive zones, determining the response stabilization waiting time through historical fluctuation decay times helps to more accurately distinguish the data stability boundaries corresponding to fast and slow response quantities. This reduces the possibility of unstable transitional data being mistakenly used as healthy samples, improving the accuracy of subsequent health benchmark construction.

[0010] 3. By generating a transition isolation window for the corresponding measuring point based on the operating condition switching boundary, the response stabilization waiting time of the corresponding measuring point, the pre-boundary preparation time, and the stabilization confirmation time, the objectified state data located within the transition isolation window is marked as operating condition transition data, and the operating condition transition data is prohibited from entering the healthy sample. The objectified state data located outside the transition isolation window is used as the candidate set of stable samples. Its advantages over existing methods are that it can proactively eliminate normal transition fluctuations caused by operating condition switching such as start-up, shutdown, load change, and guide vane adjustment before constructing the healthy sample. At the same time, the pre-boundary preparation time covers the influence of control actions before the operating condition switching, and the stabilization confirmation time prevents short-term declines from being misjudged as stable, making the candidate set of stable samples closer to the actual stable operating state of the unit.

[0011] 4. When the transition duration and fluctuation decay rate of the same measuring point are different under different load change amplitudes and different guide vane opening change rates, by monitoring the load change amplitude and guide vane opening change rate, when the load change amplitude is greater than the preset load change amplitude threshold or the guide vane opening change rate is greater than the preset guide vane opening change rate threshold, the transition isolation window is divided into sub-intervals. When the output is a stable transition isolation window, the objectified state data outside the stable transition isolation window is used as a candidate set of stable samples. When the output is an unstable transition isolation window, the objectified state data within the window is prohibited from entering the healthy sample, and the objectified state data outside the window is used as data to be verified, and an abnormal working condition transition prompt is sent. Its advantage over the existing methods is that it can dynamically distinguish the stability of data inside and outside the transition window according to the strength of the working condition switch and the intensity of the adjustment action, avoiding the entry of undamaged vibration, pressure pulsation or noise response into the healthy sample due to the fixed isolation window being too short, and also avoiding the long-term exclusion of effective data that has recovered to stability due to the fixed isolation window being too long. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0013] Figure 1 This is a flowchart of a hydropower unit condition-based maintenance system based on multi-source situational awareness provided in an embodiment of the present invention; Figure 2 This is a system network architecture diagram provided in an embodiment of the present invention; Figure 3 This is a diagram of the working condition switching boundary and transition isolation window provided in an embodiment of the present invention; Figure 4 This is a mapping matrix diagram of evidence conflict handling and inspection objects provided in the embodiments of the present invention; Figure 5 This is a flowchart of a hydropower unit condition-based maintenance method based on multi-source situational awareness provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0015] The first embodiment, as follows Figure 1The schematic diagram shown illustrates the structure of a hydropower unit condition-based maintenance system based on multi-source situational awareness. The system, as provided in this embodiment, includes the following components: a multi-source data access module: acquiring multi-source status data, operating condition data, and maintenance-related data during the hydropower unit's operation; establishing status data objects based on the multi-source status data; performing data quality identification on the multi-source status data to obtain objectified status data with quality identification; an operating condition boundary identification module: identifying the operating condition switching boundaries of the hydropower unit based on the operating condition data; setting different response stabilization waiting times for different types of status quantities based on the measurement point response characteristics; and a health benchmark purification module: establishing a transition isolation window matching the measurement point response characteristics for the operating condition switching boundaries; and selecting objectified status data outside the transition isolation window. A stable sample candidate set is generated, and the stable sample candidate set is purified to generate a health benchmark associated with the corresponding quality label, operating condition range, corresponding component, and corresponding measurement point type. Within the transition isolation window, data within the transition isolation window is prohibited from entering the health sample. Evidence classification module: Based on the real-time status value corresponding to the health benchmark, evidence is classified to generate an evidence set, including sample deviation, gradual variable, sudden change, and coupling anomaly. Evidence chain mapping module: The sample deviation, gradual variable, sudden change, and coupling anomaly are formed into evidence objects and an abnormal evidence chain is constructed. The condition maintenance decision result is determined based on the abnormal evidence chain. Feedback update module: The maintenance execution result after the condition maintenance decision result is executed is obtained, and the operating condition health benchmark, transition isolation window, and abnormal evidence discrimination rules are updated based on the maintenance execution result.

[0016] For example, taking the No. 1 turbine generator unit of a hydropower plant as an example, the rated load of this unit is 170MW and the rated speed is 187.5r / min. During a certain operation, the unit adjusted from 70MW to 95MW, and the load change reached 25MW within 5 minutes, exceeding the preset load change threshold of 8.5MW. At the same time, the guide vane opening was adjusted from 38% to 47%. Therefore, the operating condition boundary identification module determined that there was a variable load boundary at this moment and recorded the corresponding operating condition switching type as variable load operating condition switching.

[0017] The multi-source data access module synchronously acquires data on the unit's head, load, guide vane opening, and rotational speed, as well as status data such as upper guide bearing vibration, main shaft runout, pressure pulsation, unit noise, generator rotor temperature, slip ring carbon brush temperature, top cover bolt stress, and stator partial discharge. For the upper guide bearing vibration measurement point, the system binds its effective vibration velocity value, acquisition time, sampling frequency, measurement point type, and associated component as a status data object; if the measurement point data is continuous and complete and does not exceed physical boundaries, a quality qualified tag is added. For a noise measurement point, if a communication interruption occurs, the noise data is tagged with a quality abnormality and recorded in the quality management record, and is not included in the healthy sample.

[0018] Since vibration, sway, pressure pulsation, and unit noise are fast-response quantities, the system sets a short response stabilization waiting time for them, such as 3 minutes. Generator rotor temperature, slip ring carbon brush temperature, top cover bolt stress, and stator partial discharge are slow-response quantities, and the system sets a longer response stabilization waiting time for them, such as 20 minutes. The health benchmark purification module establishes transition isolation windows for different measuring points based on the load change boundary, the preparation time before the boundary, the response stabilization waiting time, and the stabilization confirmation time. For the upper guide bearing vibration measuring point, the time period from 1 minute before the load change boundary to 3 minutes after the boundary, after stabilization confirmation, is marked as the transition isolation window; for the rotor temperature measuring point, the time period from 1 minute before the load change boundary to 20 minutes after the boundary, after stabilization confirmation, is marked as the transition isolation window. Although the data within these windows can be used for process analysis, they are not included in the health sample library.

[0019] Outside the transition isolation window, the system selects qualified and stable upper guide bearing vibration samples under medium head and medium load conditions as a stable sample candidate set. Start-up / shutdown samples, load shedding samples, maintenance / test samples, forced excitation samples, abrupt change samples, and samples with continuous unidirectional drift are removed to obtain purified healthy samples. These purified healthy samples are then input into the operating condition health generation model to generate a health benchmark for the upper guide bearing vibration measurement points within that operating condition range. For example, the model outputs a health center of 1.80 mm / s, a health fluctuation range of 1.68 mm / s to 1.92 mm / s, 1200 samples, and a sample time span of 30 days.

[0020] Subsequently, after the unit stabilized at 95MW, the evidence classification module acquired the real-time vibration value of the upper guide bearing. If the effective value of the upper guide bearing vibration velocity increased to 2.25mm / s within a certain time slice, exceeding the healthy fluctuation range, the system generated a sample deviation; if the vibration value gradually increased from 1.85mm / s to 2.25mm / s within 30 minutes, the system generated a gradual variable; if the stress of the top cover bolts suddenly dropped from the original stable value of 120MPa to 105MPa in a short period of time, the system generated a sudden change; if the difference between the current Pearson correlation coefficient and the preset correlation coefficient under healthy conditions between the effective vibration velocity value sequence and the unit noise sound pressure level sequence belongs to the preset abnormal range, the system generated a coupling anomaly.

[0021] After maintenance is completed, the feedback update module obtains the maintenance execution results. If the maintenance confirms slight wear or abnormal clearance in the upper guide bearing, the system writes multi-source status data before, during, and after the abnormality trigger into the fault sample library. If the unit operates stably within the same operating condition range after maintenance, and the vibration of the upper guide bearing recovers to the range of 1.75 mm / s to 1.88 mm / s, then the data at this stage can be used as a highly reliable health sample to update the corresponding operating condition health benchmark. If no fault is found during maintenance, the system writes the corresponding evidence object into the false alarm record and checks whether there is an excessively short transition isolation window, insufficient confidence of the health benchmark, or amplification of a single measurement point abnormality, thereby correcting the subsequent abnormal evidence discrimination rules.

[0022] As can be seen from the above, this system does not simply output maintenance suggestions based on whether a certain measuring point exceeds a fixed threshold. Instead, it first identifies the boundaries of operating condition switching, such as variable load, and isolates transitional samples. Then, it generates a health baseline based on the purified stable samples. Subsequently, it breaks down real-time anomalies into deviations, gradual changes, abrupt changes, and coupled anomaly evidence, and maps the evidence chain to specific maintenance objects. Finally, it uses actual maintenance results to update the baseline and rules in reverse. Therefore, this system can reduce the contamination of the health baseline by operating condition switching transition data, improve the accuracy of anomaly identification, the accuracy of maintenance object location, and the traceability of maintenance decisions.

[0023] Furthermore, objectified status data with quality identifiers is obtained. The specific process is as follows: the data interface of the hydropower unit in the online monitoring system is connected to obtain multi-source status data, operating condition data and maintenance-related data respectively; the online monitoring system includes a stability monitoring system, an air gap monitoring system, a partial discharge monitoring system, a noise monitoring system, a rotor temperature monitoring system, a collector ring carbon brush temperature monitoring system and a top cover bolt status monitoring system.

[0024] like Figure 2 The diagram shows the system network architecture provided in this application. Various online monitoring systems are deployed in Security Zone II to comprehensively perceive the unit's status and obtain unit status information. Data from each system is collected through the communication server and network equipment in Zone II and then sent to the IV big data platform for storage and management. In addition to status perception via sensors, this system also communicates with the monitoring system. Since the monitoring system is located in Zone I, a hardware firewall is installed between the monitoring system and the communication server in accordance with network security specifications. The data platform is deployed in Zone IV to facilitate data management and access. The platform equipment includes four data servers, one algorithm server, one application server, and one engineer workstation. Data from Zone II is transmitted to Zone IV via the network for management and storage. In accordance with network security specifications, a one-way network isolation device is deployed between Zone II and Zone IV, ensuring network security as data flows unidirectionally from Zone II to Zone IV.

[0025] Performing unified time reference, measurement point coding, and unit normalization processing on multi-source state data, and establishing state data objects based on the processed multi-source state data, helps eliminate data correlation difficulties caused by inconsistencies in sampling time, measurement point naming methods, and measurement units between different monitoring systems. This enables data from different sources to be compared and analyzed under the same time axis, the same measurement point identifier, and the same dimensional standard, providing a unified data foundation for subsequent working condition boundary identification, health benchmark construction, and abnormal evidence generation. The process is as follows: Read the basic measurement point information of the corresponding measurement point according to the measurement point code. The basic measurement point information includes one or more of the following: measurement point number, measurement point name, measurement point type, monitoring system, component, installation location, monitoring direction, sampling frequency, and measurement unit; match the multi-source state data after unified time reference processing with the measurement point code and the basic measurement point information to determine each multi-source state data object. The source status data includes the measurement point type and its associated component. Based on the unit-normalized measurement point data, the measurement point values ​​within the corresponding sampling time or statistical time slice are extracted and bound to the corresponding timestamp, sampling frequency, measurement point type, and associated component. This means that the status parameter value formed by a measurement point within the corresponding sampling time or statistical time slice is recorded together with the acquisition time, data acquisition cycle, monitoring parameter category, and corresponding hydropower unit component. For example, for the upper guide bearing vibration measurement point, the effective vibration value or peak-to-peak value extracted within a certain time slice is bound to the timestamp, vibration data sampling frequency, vibration measurement point type, and upper guide bearing component identifier for that time slice. For the generator rotor temperature measurement point, the acquired rotor temperature value is bound to the corresponding acquisition time, temperature sampling frequency, temperature measurement point type, and generator rotor component identifier. Through this binding relationship, subsequent health benchmark matching, abnormal evidence generation, and maintenance object mapping can accurately locate specific measurement points and specific components. The status data object includes the measurement point value, measurement point type, associated component, sampling frequency, timestamp, and data quality label.

[0026] The measured value represents the state parameter value collected at the corresponding measured point within the sampling time or statistical time slice. For example, the measured value of a vibration measured point can be vibration displacement, vibration velocity, vibration acceleration, RMS value, or peak-to-peak value. When multiple sampled values ​​exist for the same measured point within the same statistical time slice, the corresponding statistical features are extracted as measured point values ​​based on the measured point type. For example, for vibration, sway, and pressure pulsation data, RMS value, peak value, peak-to-peak value, dominant frequency amplitude, or fluctuation amplitude can be extracted as measured point values. A unified time reference means aligning multi-source state data collected by different monitoring systems along the same time axis, so that data with different sampling frequencies, different upload cycles, and different recording formats can correspond to the same operating time or the same time slice. The measured point code indicates the value based on the system to which the measured point belongs, its classification, and its corresponding statistical features. Each measuring point is assigned a unique identification number based on its component and installation location to distinguish status data from different sources and monitoring objects. Unit normalization refers to converting similar physical quantities output by different monitoring systems into a unified unit of measurement or a unified dimensional expression. For example, different temperature units are unified to degrees Celsius, and swing or pressure pulsation is converted into a preset standard unit or normalized state value. The measuring point type is used to characterize the category of monitoring parameters corresponding to the measuring point, such as vibration measuring point, swing measuring point, pressure pulsation measuring point, temperature measuring point, air gap measuring point, magnetic flux measuring point, bolt stress measuring point, or partial discharge measuring point. The component to which the measuring point belongs is used to characterize the specific object of the hydropower unit corresponding to the measuring point, such as turbine, generator, main shaft, upper guide bearing, lower guide bearing, water guide bearing, thrust bearing, rotor, stator, water guiding mechanism, or speed control system.

[0027] The system determines whether the bound multi-source status data conforms to any data quality anomaly type. If the corresponding multi-source status data conforms to any data quality anomaly type, it records the corresponding multi-source status data in the quality governance record and prohibits it from entering the health sample. This helps to avoid low-quality data, abnormally collected data, or sensor failure data from contaminating the health sample library, reducing misjudgments, false alarms, and health benchmark deviations caused by data anomalies, and improving the reliability of subsequent health benchmarks. If the corresponding multi-source status data does not conform to any data quality anomaly type, it records the timestamp, component, and quality label corresponding to the multi-source status data. It combines the measurement point value, measurement point type, component, sampling frequency, timestamp, and data quality label to form a status data object. The multi-source status data is then used as objectified status data that can participate in subsequent operating condition boundary identification, health benchmark construction, or anomaly evidence generation. This can convert scattered raw measurement point data into objectified status data with quality identification, component affiliation, and time relationship, enabling it to directly participate in subsequent operating condition boundary identification, health benchmark construction, and anomaly evidence generation, thereby improving the standardization and automation of the data processing process.

[0028] Specifically, data quality anomalies include missing data, duplicate data, communication interruption, exceeding physical boundaries, or sensor anomalies. Exceeding physical boundaries means that the measurement data exceeds the physical value range that the corresponding monitored object can take during normal operation, allowable fluctuations, or within the equipment protection range of the hydropower unit. For example, the speed data exceeds the upper limit of the unit's allowable speed. Sensor anomalies indicate that the noise sensor, infrared temperature sensor, rotor temperature sensor, ultrasonic sensor, or partial discharge sensor has a self-diagnostic alarm.

[0029] Multi-source state data includes: vibration, sway, pressure pulsation, air gap, magnetic flux density, unit noise, generator rotor temperature, slip ring carbon brush temperature, turbine top cover bolt stress, and generator stator partial discharge, among other parameters. Specifically, hydropower unit vibration includes one or more of vibration displacement, vibration velocity, and vibration acceleration; sway includes one or more of upper guide sway, lower guide sway, water guide sway, and peak sway; pressure pulsation includes one or more of pressure pulsation amplitude and pressure pulsation peak value; air gap includes one or more of the air gap distance between the stator and rotor, minimum air gap, maximum air gap, and average air gap; magnetic flux density includes one or more of the air gap magnetic flux density, stator core magnetic flux density, and rotor pole magnetic flux density; unit noise includes one or more of the noise peak value, noise spectrum amplitude, and sound intensity; generator rotor temperature includes one or more of the rotor winding temperature, rotor pole temperature, and rotor core temperature; slip ring carbon brush temperature includes one or more of the slip ring surface temperature, carbon brush contact surface temperature, and carbon brush body temperature; and turbine top cover bolt stress. Stress includes one or more of the following: axial stress of the top cover bolts, bolt preload, and bolt tension. Partial discharge of the generator stator includes one or more of the following: partial discharge amplitude, number of discharge pulses, and partial discharge intensity. Specifically, vibration, sway, and pressure pulsation of the hydropower unit can be collected by a stability monitoring system, which may include vibration sensors, sway sensors, and pressure pulsation sensors. Air gap can be collected by an air gap monitoring system. Magnetic flux density can be monitored by a magnetic flux sensor, Hall sensor, magnetic field probe, or magnetic flux measurement coil located on the generator stator side, in the air gap, or at the corresponding position of the rotor magnetic poles. Unit noise can be collected by a noise monitoring system. Slip ring carbon brush temperature can be collected by a slip ring carbon brush temperature monitoring system. Turbine top cover bolt stress can be collected by a top cover bolt condition monitoring system. Generator stator partial discharge can be collected by a partial discharge monitoring system. Maintenance-related data includes one or more of the following based on historical maintenance time, maintenance location, fault records, and defect records obtained from the system log, used to provide historical maintenance basis for sample screening of multi-source status data, construction of health benchmarks, identification of abnormal evidence, and mapping of maintenance objects.

[0030] In this embodiment, by uniformly accessing, aligning, objectifying, standardizing units, and labeling data from multiple sources of hydropower unit status data, operational data from different monitoring systems, sampling frequencies, and physical quantity types are converted into objectified status data with quality labels. This allows subsequent condition-based maintenance processes to no longer directly rely on scattered, heterogeneous, and undefined raw data, but rather to be executed based on data objects with unified time relationships, clear measurement point attributions, clear component associations, and clear quality statuses. This effectively improves the usability of hydropower unit status data and provides reliable data support for subsequent operating condition switching boundary identification, transition isolation window establishment, health baseline purification, anomaly evidence classification, and maintenance object mapping.

[0031] Furthermore, the operating condition switching boundaries of hydropower units are identified based on operating condition data. The specific process is as follows: Operating condition data is acquired, and based on changes in the hydropower unit's head, load, guide vane opening, and rotational speed, the operating condition switching boundaries during operation are identified. This includes: when the hydropower unit's rotational speed moves from the zero-speed range into the preset grid-connected speed range, a start-up boundary is determined to exist; if the rotational speed does not move from the zero-speed range into the preset grid-connected speed range, no start-up boundary is determined to exist. When the hydropower unit's rotational speed decreases from the preset grid-connected speed range to the preset low-speed range or zero-speed range, a shutdown boundary is determined to exist; if the rotational speed does not decrease from the preset grid-connected speed range to the preset low-speed range or zero-speed range, no shutdown boundary is determined to exist. This helps to accurately distinguish between start-up and shutdown processes using speed changes, ensuring that vibrations, sway, temperature, or pressure fluctuations generated during the unit's speed-up, grid connection, disconnection, and speed-down phases are not mistaken for abnormal data under stable operating conditions, thereby improving the accuracy of healthy sample boundary identification.

[0032] When the load of a hydropower unit changes by more than a preset load change threshold within a set time period, a load change boundary is determined to exist. Conversely, if the load change does not exceed the threshold, the hydropower unit is determined not to have experienced a load change process that meets the boundary identification conditions within the corresponding set time period, and therefore no load change boundary is determined to exist. During the load change process, load changes may cause synchronous changes in state variables such as hydraulic excitation, pressure pulsation, shaft vibration, and unit noise. Therefore, identifying the load change boundary helps to distinguish between normal responses caused by load regulation and abnormal responses caused by equipment deterioration, thereby reducing the risk of false alarms.

[0033] The preset grid-connected speed range, or other ranges or ranges stored in the database in this application, is usually determined by the design parameters of the hydropower unit or the power plant's operation and maintenance experience. For example, taking a hydropower unit with a rated speed of 187.5 r / min and a rated load of 170 MW as an example, the preset grid-connected speed range can be determined based on the rated speed and the allowable grid-connected speed deviation. When the allowable grid-connected speed deviation is set to ±1%, 185.625 r / min to 189.375 r / min can be set as the preset grid-connected speed range. When the unit speed gradually increases from the zero-speed range of 0 r / min to 15 r / min and enters the range of 185.625 r / min to 189.375 r / min and maintains it for a preset confirmation time, it is determined that the unit has entered the grid-connected speed state and is identified as the start-up boundary. The zero-speed range refers to the range of speeds from when the hydropower unit is completely stationary, such as 0 r / min, to when it has just left the stationary state but the speed is still at a preset low level, such as ≤15 r / min.

[0034] The preset load change threshold, or other thresholds or preset values ​​stored in the database in this application, can be determined by historical anomaly records of the hydropower unit's design parameters and power plant operation and maintenance experience, and can be corrected based on subsequent maintenance feedback. The preset load change threshold can be determined based on the unit's rated load and the normal load fluctuation range under stable operating conditions. When the rated load of a hydropower unit is 170MW, and the natural load fluctuation within 5 minutes under historical stable operating conditions usually does not exceed 3.5MW, 8.5MW, which is significantly higher than the normal fluctuation range, can be set as the preset load change threshold. At this time, if the unit load changes from 70MW to 82MW within 5 minutes, with a change of 12MW, it is determined that the load change boundary identification condition is met; if the load only changes from 70MW to 75MW, with a change of 5MW, it is determined that the load change boundary identification condition is not met.

[0035] When the duration of continuous change in the guide vane opening of a hydroelectric generator exceeds a preset duration, it is determined that there is a guide vane opening adjustment boundary; otherwise, it is determined that there is no guide vane opening adjustment boundary. The duration of continuous change in the guide vane opening of a hydroelectric generator represents the time elapsed from the start of continuous increase or decrease in the guide vane opening until the change in the guide vane opening stops. This can incorporate the influence of the guide vane adjustment process on the water flow pattern, pressure pulsation, runner force, and vibration response into the operating condition switching identification range, so that subsequent analysis no longer relies solely on load changes to determine the operating condition, thereby improving the completeness of the identification of the turbine hydraulic regulation process.

[0036] When the unit's operating point enters or leaves the preset vibration-sensitive zone, it is determined that there is a crossing of the vibration-sensitive zone boundary; otherwise, it is determined that there is no crossing of the vibration-sensitive zone boundary. It can identify the operating stage of the hydropower unit that may experience vibration, sway, pressure pulsation, or increased noise under specific combinations of head, load, guide vane opening, or speed. This allows the system to distinguish the process of crossing the vibration-sensitive zone from the normal stable operation process, avoiding health baseline contamination or abnormal misjudgment caused by the operating point entering the sensitive zone.

[0037] It needs to be explained that the preset vibration sensitive zone refers to the operating range that is pre-determined based on historical operating data, vibration tests, or maintenance experience of the hydropower unit, which is prone to increased vibration, sway, pressure pulsation, or noise. The unit operating point refers to the operating state position of the hydropower unit at a certain moment or within a certain statistical time slice, determined by the operating condition data. For example, the unit operating point can be determined by using head and load as two-dimensional coordinates, or by using head, load, guide vane opening, and speed as multi-dimensional operating condition coordinates. The operating condition switching boundary includes one or more of the following: start-up boundary, shutdown boundary, load change boundary, guide vane opening adjustment boundary, and boundary crossing the vibration sensitive zone. The operating condition data includes... The parameters of a hydropower unit include one or more of the following: head, load, guide vane opening, and rotational speed. Head refers to the effective water level difference used by the hydropower unit to perform work during operation. Specifically, it can be the height difference between the upstream water level and the downstream tailwater level. It can be monitored by a pressure-type water level sensor and is used to help determine whether the operating condition has changed and whether the unit's operating point has entered or left the preset vibration sensitive zone. The load of the hydropower unit represents the active power of the hydropower unit and can be monitored by a generator active power transmitter. The guide vane opening can be monitored by a guide vane opening sensor. The rotational speed of the hydropower unit represents the rotational speed of the rotating parts of the turbine generator unit per unit time and can be monitored by a speed sensor or a speed encoder.

[0038] In one optional calculation method, the operating condition switching intensity is calculated based on the operating condition data, as shown in the following formula: Where K(t) is the operating condition switching intensity; H, P, G, and n are the head, load, guide vane opening, and rotational speed, respectively; N_H, N_P, N_G, and N_n are normalization functions obtained by expert experience based on historically stable hydropower unit operating condition data scales; limit is the boundary constraint function; and K_max is the upper limit value. The operating condition switching boundary is identified based on the operating condition switching intensity, and the specific process is as follows: The operating condition switching intensity corresponding to each sampling moment is calculated on the continuous time axis and compared with a preset switching intensity threshold; when the operating condition switching intensity changes from below the preset switching intensity threshold to above or equal to the preset switching intensity threshold, the corresponding moment is determined. The candidate operating condition switching start boundary is defined as follows: when the operating condition switching intensity is continuously lower than the preset switching intensity threshold and remains below the preset stable confirmation time, the corresponding moment is defined as the candidate operating condition switching end boundary; combined with the load change direction, guide vane opening change direction, speed range, and whether the unit operating point enters or leaves the preset vibration sensitive zone, the type of the candidate operating condition switching boundary is determined to obtain the start boundary, shutdown boundary, load change boundary, guide vane opening adjustment boundary, or boundary crossing the vibration sensitive zone. For example, when the speed of the hydropower unit enters the preset grid-connected speed range from the zero speed range around the time corresponding to the candidate operating condition switching start boundary, the candidate operating condition switching boundary is determined as the start boundary.

[0039] By defining the moment when the intensity of the operating condition switching changes from below the preset switching intensity threshold to above or equal to the preset switching intensity threshold as the candidate operating condition switching start-up boundary, it is possible to promptly capture the starting moment when the hydropower unit enters the process of starting up, stopping, changing load, adjusting guide vane opening, or crossing the vibration sensitive zone from a stable operating state, providing an accurate basis for determining the starting point of the subsequent transition isolation window. By defining the moment when the intensity of the operating condition switching remains below the preset switching intensity threshold and exceeds the preset stability confirmation time as the candidate operating condition switching end boundary, it is possible to avoid prematurely determining the end of the operating condition switching due to short-term decline or instantaneous fluctuation of the operating condition variables, making the end boundary of the operating condition switching process more stable and reliable, thereby reducing the risk of transitional data mistakenly entering healthy samples. By combining the load change direction, the guide vane opening change direction, the speed range, and whether the unit operating point enters or leaves the preset vibration sensitive zone, the candidate operating condition switching boundary is classified, which can further distinguish between the start-up boundary, the shutdown boundary, the changing load boundary, the guide vane opening adjustment boundary, and the boundary of crossing the vibration sensitive zone, so that different types of subsequent operating condition switching correspond to different response stability waiting times and transition isolation strategies, improving the pertinence of transition sample isolation.

[0040] To address the response characteristics of different types of state variables at measurement points, different response stabilization waiting times are set. The specific process is as follows: Based on the timer statistics of the fluctuation decay time of the same measurement point before and after a preset number of operating condition changes, the time required for the state change amplitude to recover to the preset stable fluctuation range is taken as the response stabilization waiting time for that measurement point; the response stabilization waiting time is determined based on the historical fluctuation decay time of the measurement point, so that the transition data of different measurement points after operating condition changes such as start-up, shutdown, load change, guide vane opening adjustment, or crossing of vibration sensitive areas can be differentiated and isolated, thereby improving the accuracy and adaptability of the transition isolation window setting.

[0041] In one implementation, the fluctuation attenuation time of the same measuring point during multiple historical operating condition switching processes can be statistically analyzed, and its median, mean, or preset quantile can be used as the response stabilization waiting time of the measuring point. The state change amplitude represents the absolute value of the difference between the measuring point values ​​before and after the operating condition switching. For example, taking the operating condition switching corresponding to the variable load boundary as an example, the measuring point value of the corresponding measuring point within the stable load range before the load change is used as the reference value, and the difference between the measuring point value corresponding to each sampling time after the load change and the reference value is calculated to obtain the state change amplitude corresponding to each sampling time. When the state change amplitude gradually decreases from greater than the preset stable fluctuation range to the preset stable fluctuation range, the state change amplitude is calculated. Within a preset stable confirmation time, if the load changes and the measured point remains within a preset stable fluctuation range, the time from the moment of load change to the moment the stable state is formed is defined as the fluctuation decay time of that measured point under that load change. The fluctuation decay time represents the time elapsed from the moment of the change in operating conditions until the amplitude of the state change at the corresponding measured point remains within the preset stable fluctuation range and meets the preset stable confirmation time. For fast-response quantities such as vibration, sway, pressure pulsation, and unit noise, since they usually respond and tend to stabilize more quickly after load changes, guide vane opening adjustments, or hydraulic excitation changes, a relatively short response stabilization waiting time is set. For slow-response quantities such as generator rotor temperature, slip ring carbon brush temperature, turbine top cover bolt stress, and generator stator partial discharge, since they are affected by factors such as heat conduction, structural stress release, and changes in electrical insulation status, they usually have a longer hysteresis response process, so a relatively long response stabilization waiting time is set.

[0042] In this embodiment, by performing boundary identification on the hydropower unit's operating data, the operating condition switching processes, such as start-up, shutdown, load change, guide vane opening adjustment, and crossing vibration-sensitive zones, are clearly separated from the unit's continuous operating data. This allows subsequent health baseline construction, transition isolation window setting, and anomaly evidence generation to be performed based on clearly defined operating condition stages. Consequently, the interference of normal transition responses during operating condition switching on health samples and dynamic thresholds can be reduced, improving the accuracy of health baseline construction, the reliability of anomaly detection, and the traceability of maintenance object location during hydropower unit condition-based maintenance.

[0043] Furthermore, a transition isolation window matching the response characteristics of the measurement points is established for the operating condition switching boundary. The specific process is as follows: Based on the operating condition switching boundary, the response stabilization waiting time of the corresponding measurement point, the preparation time before the boundary, and the stabilization confirmation time, a transition isolation window for the corresponding measurement point is generated. The transition isolation window covers the preparation time before the operating condition switching boundary, the response stabilization waiting time after the operating condition switching boundary, and the stabilization confirmation time. Objectified state data located within the transition isolation window are marked as operating condition transition data, and operating condition transition data is prohibited from entering healthy samples. Objectified state data located outside the transition isolation window are used as a candidate set of stable samples. Figure 3 The diagram shown illustrates the operating condition switching boundary and transition isolation window provided in this application. The horizontal axis represents time, the left axis represents normalized state characteristics, and the right axis represents operating condition variables. The yellow area represents the transition isolation window, and the green area represents the stable sample candidate area. The diagram shows that unstable samples within the transition isolation window do not enter the healthy baseline.

[0044] Because different measuring points respond differently to operating condition changes, the settling times of fast-response quantities such as vibration, sway, and pressure pulsation are not consistent with those of slow-response quantities such as temperature, partial discharge, and bolt stress. Therefore, generating a transition isolation window for the corresponding measuring point based on the response stabilization waiting time, the boundary preparation time, and the stabilization confirmation time can avoid the problem of some unstable data entering the healthy sample or excessive exclusion of stable data caused by using a uniform isolation time. Furthermore, by marking the objectified state data within the transition isolation window as operating condition transition data and prohibiting it from entering the healthy sample, it is possible to prevent state fluctuations caused by normal operating condition changes from contaminating the healthy sample library, thus ensuring the health base generated subsequently. The accuracy better reflects the true health status of the hydropower unit under stable operating conditions; the response stabilization waiting time is used to indicate the waiting time required for the corresponding measuring point to recover from the preset transition fluctuation state to the preset stable state after the operating condition switching boundary occurs; the preparation time before the operating condition switching boundary indicates the time interval extended forward to cover the start of guide vane opening adjustment, load regulation command issuance, speed change precursors, or early operation control actions before the operating condition switching boundary is identified; the stabilization confirmation time indicates the continuous observation time required to confirm that the state of the measuring point has not deviated from the preset stable fluctuation range and remains relatively stable after the state change amplitude of the corresponding measuring point enters the preset stable fluctuation range.

[0045] The stable sample candidate set is purified to generate health benchmarks associated with corresponding quality labels, operating condition ranges, corresponding components, and corresponding measurement point types. The specific process is as follows: maintenance test samples, start-up and shutdown samples, load shedding samples, forced excitation samples, mutation samples, and continuous unidirectional drift samples are removed from the stable sample candidate set to obtain purified samples; the purified samples are input into the preset operating condition health generation model to generate health benchmarks. The health benchmarks include health center, health fluctuation range, number of samples, sample time span, and sample confidence level.

[0046] The maintenance test sample represents the status sample collected during the maintenance of the hydropower unit. The start-up and shutdown sample represents the status sample collected during the start-up or shutdown of the hydropower unit. The load shedding sample represents the status sample collected during the process of the hydropower unit's load dropping to zero load within a preset change time period. The forced excitation sample represents the status sample collected when the generator excitation system is under forced excitation. The sudden change sample represents the status sample where the measured value exceeds the preset sudden change judgment condition by jumping, dropping, peaking, or stepping within a preset change time period. The sudden change judgment condition can be determined based on the change amplitude of adjacent sampled values. The continuous unidirectional drift sample represents the status sample where the measured value continuously changes in the direction of increase or decrease within a preset continuous time slice and does not recover to the preset stable fluctuation range within a preset observation time.

[0047] The "health center" refers to the typical health status value represented by the purified sample under the corresponding quality label, operating condition range, component, and measuring point type. This value can be determined based on the mean, median, or cluster center of the purified sample. For example, for a purified sample with a quality label of "qualified," an operating condition range of medium head and medium load, an upper guide bearing as the component, and a vibration measuring point type, the mean or median of the effective vibration value of the upper guide bearing within that operating condition range can be used as the corresponding health center. The operating condition range represents the range of operating states defined by pre-defined personnel based on the hydropower unit's operating data, used to characterize the hydropower unit under the same or similar operating conditions. The stable operating range under near-hydraulic and output conditions, for example, the head can be divided into low head range, medium head range, and high head range. The health fluctuation range represents the normal allowable fluctuation range of the purified sample around the health center under the corresponding quality label, corresponding operating condition range, corresponding component, and corresponding measurement point type, which can be determined based on the sample standard deviation or quantile interval. The sample number represents the number of purified samples participating in the generation of the corresponding health benchmark. The sample time span represents the coverage of the purified samples participating in the generation of the corresponding health benchmark on the time axis. The sample confidence represents the confidence level of the purified samples participating in the generation of the corresponding health benchmark.

[0048] In addition, the pre-defined operating condition health generation model can be established using a random forest regression model. The model establishment process includes training sample construction, model training, model validation, and health benchmark output. First, based on healthy samples from historical operating data that have undergone transitional isolation window removal and sample purification, model training samples are constructed. Each training sample includes input features and output targets; the input features include head, load, guide vane opening, rotational speed, operating condition interval identifier, component code, measuring point type code, sampling frequency, and data quality label; the output target is the health benchmark of the corresponding measuring point under the operating condition interval. Then, the training samples are divided into training and validation sets, and the random forest regression model is trained using the training set, enabling the model to learn the distribution relationship of health benchmarks under different head, load, guide vane opening, rotational speed, component, and measuring point type conditions. The random forest regression model consists of multiple decision trees, each trained based on a portion of the training samples and a portion of the input features, and outputs the predicted health value of the corresponding measuring point under the current operating condition; the average of the outputs of multiple decision trees is used as the candidate health center value for that measuring point under the corresponding operating condition interval.

[0049] During model validation, purified healthy samples from the validation set are input into the trained random forest regression model to obtain predicted health values. These predicted health values ​​are then compared with the actual measurement values ​​in the validation set, and the prediction error, residual distribution, and stable sample coverage are calculated. If the prediction error is within a preset allowable range and a preset proportion of the validation samples fall within the health fluctuation range generated by the model, the working condition health generation model is deemed to meet the validation requirements. If the validation results do not meet the requirements, the training sample range, working condition interval division method, feature input items, or model parameters are readjusted, and the model is retrained.

[0050] After successful model validation, the purified samples under the current operating condition range, corresponding quality labels, corresponding components, and corresponding measurement point types are input into the operating condition health generation model to generate a health baseline. The health center can be determined by the predicted health value output by the random forest regression model; the health fluctuation range can be determined by the residual distribution or quantile intervals of the training and validation samples; the sample size is the number of purified samples participating in the generation of the operating condition health baseline; the sample time span is the coverage of the purified samples participating in the generation of the operating condition health baseline on the time axis; and the sample confidence level is determined based on the model validation error.

[0051] For example, within the medium head and medium load operating range of a certain hydropower unit, 1200 cleaned samples were selected, labeled as qualified, belonging to the upper guide bearing, and classified as vibration measuring points. The sample time span was 30 consecutive days. Input features included head, load, guide vane opening, rotational speed, component code, and measuring point type code. The output target was the effective value of the vibration velocity of the upper guide bearing. 960 samples were used as the training set, and 240 samples as the validation set to train and validate the random forest regression model. If the validation results showed that the average absolute error between the model's predicted effective vibration velocity and the actual effective vibration velocity was 0.03 mm / s, and 95% of the validation sample residuals were within ±0.12 mm / s, then the model was considered validated successfully. After successful model validation, the purified samples under the current operating conditions are input into the operating condition health generation model. The model outputs a predicted health value of 1.80 mm / s for the vibration measurement point of the upper guide bearing. This 1.80 mm / s is then taken as the health center for this measurement point under medium head and medium load operating conditions. Based on the residual distribution of the training and validation samples, the health fluctuation range is determined to be 1.68 mm / s to 1.92 mm / s, or expressed as the health center ±0.12 mm / s. The sample size is recorded as 1200, and the sample time span is recorded as 30 days. If the model validation error is less than the preset error threshold of 0.05 mm / s, and the sample size is greater than the preset number threshold of 1000, then the sample confidence level of this health benchmark is determined to be high confidence.

[0052] For example, for generator rotor temperature measurement points under the same operating condition range, if the number of purified samples is 800, the sample time span is 20 days, the predicted healthy temperature output by the random forest regression model is 58℃, the mean absolute error of the validation set is 0.6℃, and 95% of the validation sample residuals are distributed within the range of ±1.5℃, then 58℃ can be regarded as the healthy center, and 56.5℃ to 59.5℃ can be regarded as the healthy fluctuation range.

[0053] Evidence is categorized based on the real-time status value corresponding to the health benchmark to generate an evidence set. The specific process is as follows: The process of generating the sample deviation D_i is as follows: The difference between the real-time status value and the health center within the operating condition interval is calculated, and the absolute value of the result is taken. The ratio of this result to the maximum value in the health fluctuation range and the denominator protection value is used as the sample deviation, as follows: Where D_i(t) is the sample deviation of measurement point i at time t, v_i(t) is the real-time status value; B_i(z) is the health center within the working condition interval z, S_i(z) is the health fluctuation range; ε is the denominator protection value, where the denominator protection value represents the preset minimum value used to avoid abnormal calculation of sample deviation when the health fluctuation range is zero or fails to be effectively statistically analyzed.

[0054] The gradual variable T_i is represented by the changing trend of the measuring point within a continuous time window. In one optional implementation, the gradual variable can be represented by the slope of the state change of the measuring point within the current time window, that is, the difference between the measuring point value at the end of the current time window and the measuring point value at the beginning of the window, divided by the corresponding time length, to obtain the trend change of the measuring point. Specifically, for the generator rotor temperature measuring point, the gradual variable can be represented as the temperature rise rate of the generator rotor temperature during continuous operation. The abrupt change J_i is represented by the instantaneous change amplitude of the measuring point within adjacent sampling moments or short time windows. In one optional implementation, the abrupt change can be represented by the absolute value of the difference between the measuring point value at the current moment and the previous moment, or by the difference between the maximum and minimum measuring point values ​​within a short time window. The coupling anomaly C_i represents the difference between the correlation coefficient of two measuring point values ​​and the preset correlation coefficient. The correlation coefficient represents the Pearson correlation coefficient of the two measuring point values ​​within the current time window, calculated based on a preset correlation analysis algorithm, such as the Pearson correlation coefficient algorithm.

[0055] Specifically, sample deviation, gradual variables, sudden changes, and coupled anomalies are formed into evidence objects and anomaly evidence chains are constructed. Based on these anomaly evidence chains, condition-based maintenance decisions are determined. The specific process is as follows: Sample deviation, gradual variables, sudden changes, and coupled anomalies are encapsulated as evidence objects, enabling the conversion of different types of anomaly calculation results into structured evidence units in a unified format. Each piece of anomaly evidence retains information such as the component it belongs to, evidence type, trigger time, duration, real-time status value, health baseline, evidence value, and candidate fault indication. Figure 4 The figure shown is a mapping matrix of evidence conflict handling and maintenance objects provided in this application. The horizontal axis represents the consistency and quality status of evidence, the vertical axis represents the type of evidence, and the cell represents the maintenance response. The figure shows the differences in the handling of different abnormal evidence under the conditions of single-point triggering, multi-source consistency, mechanism consistency, and invalid data quality. The color from light to dark indicates that the maintenance response level gradually increases. The light-colored area corresponds to observation or data governance, and the darker area corresponds to planned inspection, special inspection, or emergency escalation.

[0056] Therefore, abnormal information is no longer compressed into a single comprehensive score, but can retain the source, type, occurrence time, and component attribution of the abnormality, which is beneficial for building a traceable chain of abnormal evidence. The evidence object includes at least the component to which it belongs, the type of evidence, the trigger time, the duration, the real-time status value, and the health benchmark. Encapsulation means that the sample deviation, gradual variable, sudden change, and coupled anomaly are converted into structured evidence objects, that is, the source, value, time, and component attribution of each type of abnormal evidence are recorded according to a unified data field format. For example, for the sample deviation, the deviation degree of the corresponding measurement point, the real-time status value, the health center, the health fluctuation range, the component to which it belongs, the trigger time, and the data quality label are written into the corresponding sample deviation evidence object. The real-time status value represents the measurement point value corresponding to the measurement point at the evidence trigger time or the current time. Among them, the evidence corresponding to the sample deviation describes the deviation of the real-time value from the health benchmark; the evidence corresponding to the gradual variable describes the continuous rise or fall of indicators such as temperature, vibration, or partial discharge; the evidence corresponding to the sudden change describes short-term anomalies such as sudden drop in bolt preload and sudden vibration jump; the evidence corresponding to the coupled anomaly describes the changes in the relationship between vibration and noise, swing and bearing temperature, rotor temperature and magnetic flux.

[0057] The system invokes a pre-defined fault knowledge base to map evidence objects to measurement point anomalies, component states, candidate faults, and maintenance objects, constructing an abnormal evidence chain. The specific process is as follows: Evidence objects representing real-time generated sample deviations, gradual variables, sudden changes, and coupled anomalies are stored in a unified format. Within a pre-defined time window, evidence objects are clustered and grouped according to component affiliation, enabling the association of scattered multi-source abnormal evidence to the same component or the same fault mechanism. Each evidence group is sequentially matched against matching rules in the pre-defined fault knowledge base. Matching rules include, but are not limited to, evidence type combinations, measurement point combinations, and trigger time difference thresholds for measurement point-subtracted evidence objects. If a match is successful, the successfully matched evidence group is associated as an abnormal evidence chain. The abnormal evidence chain is input into a pre-defined maintenance decision result generation model, outputting a state-based maintenance decision result. This state-based maintenance decision result is then bound to the corresponding specific component of the hydropower unit, such as the generator rotor or generator stator. If a match fails, the evidence objects in that evidence group are marked as unmatched evidence, and an audit alert is pushed.

[0058] The condition-based maintenance decision-making results include: maintenance object, candidate faults, primary evidence, auxiliary evidence, evidence to be observed, recommended maintenance level, maintenance items, and required tools and spare parts. Primary evidence, auxiliary evidence, and evidence to be observed represent: evidence directly pointing to candidate faults in the abnormal evidence chain; evidence belonging to the same evidence chain as the primary evidence but insufficient to trigger the current maintenance recommendation on its own; and evidence in the abnormal evidence chain that needs continuous monitoring in subsequent monitoring but does not change the current maintenance decision. The pre-built fault knowledge base represents a pre-constructed structured knowledge set describing the mapping relationship between multi-source evidence combinations and fault modes, component states, and maintenance operations. The abnormal evidence chain describes the process from the generation and development of an anomaly to its pointing to the maintenance object. The complete process; in which, when there is a conflict between evidence objects, that is, when the same component generates multiple pieces of evidence within a preset time, but the different evidence points to different fault types or maintenance urgency levels, for example: vibration deviation indicates an urgent anomaly, while temperature gradual change evidence indicates that observation is sufficient; or coupled anomalies show a deviation in the relationship, but the deviation of a single measuring point does not exceed the limit, the system processes according to the evidence stratification rules, which include: evidence objects corresponding to sudden changes are given priority to trigger emergency inspections or rapid verification; evidence objects corresponding to gradual changes are given priority to enter planned maintenance or continuous observation; evidence objects corresponding to coupled anomalies are given priority over single score increases; unqualified evidence objects do not trigger maintenance level upgrades; a single score increase indicates that the sample deviation of a certain measuring point has increased.

[0059] The operating condition health baseline, transition isolation window, and abnormal evidence discrimination rules are updated based on the maintenance execution results. The specific process is as follows: The maintenance execution results after execution according to the condition-based maintenance decision are obtained. These results include states such as confirmed fault, no fault found, normal operation after maintenance, and fault found but no early warning. When the maintenance execution result shows a confirmed fault, multi-source state data before, during, and after the fault trigger is written into the fault sample library, which enhances the identification capability under the same fault type. When the maintenance execution result shows no fault found, the corresponding evidence object is written into the false alarm record, and a prompt to reduce the confidence of the evidence object is sent, which helps reduce subsequent false alarms. When the maintenance execution result shows normal operation after maintenance, the corresponding multi-source state data is added to the health sample library as a high-confidence health sample, and the corresponding feedback update target is updated. When the maintenance execution result shows a fault found but no early warning, the state data before the fault is traced back, and a check is sent to examine whether there are gaps in the transition isolation window, health baseline, trend evidence, and coupling evidence. Note: Feedback update targets include health baseline, health sample confidence, transition isolation window, evidence object confidence, and rule branches. The transition isolation window represents the dynamic time interval established by the generator set at the boundary of operating condition switching based on the response stabilization waiting time of different types of measuring points. The rule branch is interpreted as follows: in the preset fault knowledge base, for the same component or the same fault mode, there may be multiple different combinations of evidence that can all point to the same or similar fault conclusions. Each combination of evidence and its corresponding mapping relationship, including evidence type, measuring point combination, time constraint, quantitative index threshold, output fault type, and maintenance suggestion, is called a rule branch. Among them, multi-source state data in the abnormal development stage before maintenance must not be included in the health sample database.

[0060] like Figure 5The flowchart shown is a process flow of the hydropower unit condition-based maintenance method based on multi-source situational awareness provided in this application, including: S1, Multi-source data access: Acquiring multi-source state data, operating condition data, and maintenance-related data during the operation of the hydropower unit; establishing state data objects based on the multi-source state data; performing data quality identification on the multi-source state data to obtain objectified state data with quality identification; S2, Operating condition boundary identification: Identifying the operating condition switching boundary of the hydropower unit based on the operating condition data; setting different response stabilization waiting times for different types of state quantities corresponding to the measurement point response characteristics; S3, Health benchmark purification: Establishing a transition isolation window that matches the measurement point response characteristics for the operating condition switching boundary; selecting objectified state data outside the transition isolation window to form a stable sample candidate set; and stabilizing... The sample candidate set undergoes sample purification to generate a health benchmark associated with the corresponding quality label, operating condition range, corresponding component, and corresponding measurement point type. Within the transition isolation window, data within the transition isolation window is prohibited from entering the health sample. S4, Evidence Classification: Evidence is classified according to the real-time status value corresponding to the health benchmark to generate an evidence set, including sample deviation, gradual variable, sudden change, and coupling anomaly. S5, Evidence Chain Mapping: Sample deviation, gradual variable, sudden change, and coupling anomaly are formed into evidence objects and an abnormal evidence chain is constructed. The condition maintenance decision result is determined based on the abnormal evidence chain. S6, Feedback Update: The maintenance execution result after the condition maintenance decision result is executed is obtained, and the operating condition health benchmark, transition isolation window, and abnormal evidence discrimination rules are updated based on the maintenance execution result.

[0061] In this embodiment, a continuous processing flow—including transition isolation window establishment, health sample purification, operating condition health benchmark generation, abnormal evidence objectification, abnormal evidence chain mapping, and maintenance feedback updates—transforms the condition-based maintenance of hydropower units from traditional single-point threshold judgment or comprehensive scoring judgment into a traceable maintenance decision-making method based on operating condition boundaries, measurement point response characteristics, health benchmarks, and multi-source evidence chains. This scheme effectively avoids contamination of health samples by normal transitional fluctuations during operating condition switching, reducing the risk of health benchmark distortion; it retains different evidence forms such as sample deviation, gradual change, abrupt change, and coupling anomalies, improving the hierarchy and interpretability of anomaly identification; it maps abnormal evidence chains to specific components, candidate faults, and maintenance objects, improving the location accuracy of maintenance decisions; and it continuously corrects health benchmarks, transition isolation windows, and evidence discrimination rules based on actual maintenance results, thereby reducing the risk of false alarms and missed alarms and improving the reliability, accuracy, and continuous optimization capability of condition-based maintenance of hydropower units.

[0062] The second embodiment adds the following method to the first embodiment: the transition duration and fluctuation decay rate of the same measuring point are not the same under different load change amplitudes and different guide vane opening change rates. If existing condition-based maintenance methods use a uniform time length or fixed rules to isolate the operating condition switching phase, the isolation window may be too short or too long. If the isolation window is too short, unstable transitional data will still enter the healthy sample set, contaminating the health baseline for each operating condition. If the isolation window is too long, it will exclude effective operating data that has recovered to stability, resulting in an insufficient number of healthy samples and affecting the timeliness of dynamic threshold updates and abnormal trend identification. In this case, a transition isolation window matching the response characteristics of the measurement point is established for the operating condition switching boundary. This also includes dynamically adjusting the response stabilization waiting time based on the load change amplitude and guide vane opening change rate. The specific process is as follows: Monitor the load change amplitude and guide vane opening change rate. When the load change amplitude is greater than the preset load change amplitude threshold or the guide vane opening change rate is greater than the preset guide vane opening change rate threshold, input the corresponding load change amplitude and guide vane opening change rate into the hierarchical model of the transition isolation window to divide the transition isolation window into sub-intervals, including stable transition isolation windows and unstable transition isolation windows.

[0063] It should be noted that the hierarchical model of the transition isolation window and the preset maintenance decision result generation model can be constructed in a similar way to the preset working condition health generation model. That is, training samples are constructed based on historical operating data, maintenance record data and manual annotation results, and trained and validated with the help of random forest model or other machine learning models, such as gradient boosting tree model, support vector machine model or neural network model, one or more of them.

[0064] The training samples for the transition isolation window grading model include input features and output labels. Input features include one or more of the load change amplitude and guide vane opening change rate. Output labels represent the grading results for the corresponding transition isolation window, which include stable and unstable transition isolation sub-windows. During model training, historical load change processes and guide vane opening adjustment processes can be used as sample sources. Each time slice is labeled as belonging to a stable or unstable transition isolation sub-window based on manual verification results or historical stability criteria. After training, the load change amplitude and guide vane opening change rate during the current operating condition switching process are input into the transition isolation window grading model, outputting the grading results for each time slice or sub-interval within the corresponding transition isolation window.

[0065] The training samples corresponding to the pre-defined maintenance decision generation model include input features and output labels; among them, the input features include abnormal evidence chains, and the output labels include one or more of the following: maintenance object, candidate fault, main evidence, auxiliary evidence, evidence to be observed, suggested maintenance level, maintenance items, and required tools and spare parts.

[0066] When the output is a stable transition isolation window, the objectified state data outside the stable transition isolation window is used as a candidate set of stable samples. When the output is an unstable transition isolation window, the objectified state data within the window is prohibited from entering the healthy sample, and the objectified state data outside the window is used as data to be verified, and an abnormal operating condition transition prompt is sent. Among them, the load change amplitude value represents the absolute value of the change in active power of the hydropower unit within a preset time, such as 1 second or 1 minute. The larger the load change amplitude value, the stronger the impact load on the unit, and the corresponding increase in the transition response time and fluctuation amplitude of each measuring point. The guide vane opening change rate represents the change in guide vane opening within a preset time. The guide vane opening change rate reflects the severity of the action of the regulating system. The higher the rate, the more significant the transition process such as water hammer effect, pressure fluctuation and mechanical vibration.

[0067] In this embodiment, by further introducing the load change amplitude and guide vane opening change rate based on the first embodiment, the transition isolation window is dynamically graded, which can modify the transition isolation strategy of different measuring points according to the severity of the operating condition switch. Since the state response duration and fluctuation decay rate of the same measuring point are different under small load adjustment, rapid load change, slow guide vane adjustment and rapid guide vane action, this embodiment no longer uses a uniform fixed length isolation window. Instead, the load change amplitude and guide vane opening change rate are input into the grading model of the transition isolation window to identify whether the data in the transition isolation window has entered a stable transition state. Thus, when the output is a stable transition isolation window, the objectified state data outside the window can be used as a candidate set of stable samples to increase the number of available healthy samples. When the output is an unstable transition isolation window, the data in the window is prohibited from entering the healthy sample, and the data outside the window is processed for verification and the operating condition transition anomaly is indicated, which helps to avoid the pollution of the healthy benchmark by transition state data that has not been fully decayed.

[0068] The above-disclosed embodiments are merely some examples of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A condition-based maintenance system for hydropower units based on multi-source situational awareness, characterized in that, include: The multi-source data access module is used to acquire multi-source status data, operating condition data and maintenance-related data during the operation of hydropower units. Based on the multi-source status data, a status data object is established, and the multi-source status data is marked with data quality to obtain objectified status data with quality marking. The operating condition boundary identification module is used to identify the operating condition switching boundary of the hydropower unit based on the operating condition data, and to set different response stabilization waiting times for the measurement point response characteristics corresponding to different types of state variables. The health baseline purification module is used to establish a transition isolation window that matches the response characteristics of the measurement point for the working condition switching boundary. Outside the transition isolation window, objectified state data is selected to form a stable sample candidate set, and the stable sample candidate set is purified to generate a health baseline associated with the corresponding quality label, working condition range, corresponding component and corresponding measurement point type. Within the transition isolation window, data within the transition isolation window is prohibited from entering the health sample. The evidence classification module is used to classify evidence based on the real-time status value corresponding to the health benchmark to generate an evidence set, including sample deviation, gradual variable, mutation amount and coupling anomaly amount; The evidence chain mapping module is used to form the sample deviation, gradual variable, mutation amount and coupling anomaly amount into evidence objects and construct an abnormal evidence chain, and determine the condition maintenance decision result based on the abnormal evidence chain; The feedback update module is used to obtain the maintenance execution results after the condition maintenance decision is executed, and to update the operating condition health benchmark, transition isolation window and abnormal evidence discrimination rules based on the maintenance execution results.

2. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 1, characterized in that, The specific process for obtaining objectified state data with quality identifiers is as follows: The system connects to the data interface of the hydropower unit in the online monitoring system to acquire multi-source status data, operating condition data, and maintenance-related data. The multi-source state data is processed with a unified time reference, measurement point encoding and unit normalization, and a state data object is established based on the processed multi-source state data. The process of establishing a state data object based on the processed multi-source state data is as follows: Read the basic information of the corresponding measuring point according to the measuring point code; After the multi-source status data has been processed with a unified time reference, it is matched with the measurement point code and the basic information of the measurement point to determine the measurement point type and the component to which each multi-source status data belongs. Based on the unit-normalized measurement point data, extract the measurement point values ​​within the corresponding sampling time or statistical time slice, and bind the measurement point values ​​with the corresponding timestamp, sampling frequency, measurement point type, and the component to which they belong; Determine whether the bound multi-source status data meets the data quality anomaly type. If the corresponding multi-source status data meets any data quality anomaly type, record the corresponding multi-source status data in the quality governance record and prohibit it from entering the healthy sample. If the corresponding multi-source status data does not meet any data quality anomaly type, record the timestamp, component, and quality label corresponding to the multi-source status data. Combine the measurement point value, measurement point type, component, sampling frequency, timestamp, and data quality label to form a status data object, and treat the multi-source status data as objectified status data. The multi-source state data includes one or more of the following: vibration, sway, pressure pulsation, air gap, magnetic flux density, unit noise, generator rotor temperature, collector ring carbon brush temperature, turbine top cover bolt stress, and generator stator partial discharge.

3. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 1, characterized in that, The specific process for identifying the operating condition switching boundary of the hydropower unit based on the operating condition data is as follows: Acquire operating data and, based on changes in the hydropower unit's head, load, guide vane opening, and rotational speed, identify the operating condition switching boundaries during the hydropower unit's operation, including: When the speed of the hydropower unit enters the preset grid-connected speed range from the zero speed range, it is determined that there is a start-up boundary. If the speed of the hydropower unit does not enter the preset grid-connected speed range from the zero speed range, it is determined that there is no start-up boundary. When the speed of the hydropower unit drops from the preset grid-connected speed range to the preset low speed range or zero speed range, it is determined that there is a shutdown boundary. If the speed of the hydropower unit does not drop from the preset grid-connected speed range to the preset low speed range or zero speed range, it is determined that there is no shutdown boundary. When the load of the hydropower unit changes by more than the preset load change threshold within a set time, it is determined that there is a load change boundary; otherwise, it is determined that there is no load change boundary. If the duration of continuous change in the guide vane opening of the hydropower unit exceeds the preset duration, it is determined that there is a guide vane opening adjustment boundary; otherwise, it is determined that there is no guide vane opening adjustment boundary. When the unit's operating point enters or leaves the preset vibration-sensitive zone, it is determined that it has crossed the boundary of the vibration-sensitive zone; otherwise, it is determined that it has not crossed the boundary of the vibration-sensitive zone.

4. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 3, characterized in that, The specific process for setting different response stabilization waiting times for measurement points corresponding to different types of state variables is as follows: The fluctuation decay time of the same measuring point before and after a preset number of operating condition changes is statistically analyzed, and the time required for the state change amplitude to recover to the preset stable fluctuation range is taken as the response stabilization waiting time corresponding to the measuring point.

5. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 4, characterized in that, The specific process for establishing a transition isolation window that matches the response characteristics of the measurement point at the operating condition switching boundary is as follows: Based on the operating condition switching boundary, the response stabilization waiting time of the corresponding measuring point, the preparation time before the boundary, and the stabilization confirmation time, a transition isolation window for the corresponding measuring point is generated. Objectized state data located within the transition isolation window are marked as operating condition transition data, and operating condition transition data are prohibited from entering healthy samples. Objectized state data located outside the transition isolation window are used as a stable sample candidate set.

6. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 5, characterized in that, The specific process for generating health benchmarks associated with corresponding quality labels, operating condition ranges, corresponding components, and corresponding measurement point types is as follows: The maintenance test samples, start-up and shutdown samples, load shedding samples, strong excitation samples, mutation samples and continuous unidirectional drift samples in the stable sample candidate set are removed to obtain the purified samples. The purified sample is input into a preset working condition health generation model to generate a health benchmark, which includes a health center, health fluctuation range, number of samples, sample time span, and sample confidence level. Evidence is categorized based on the real-time status value corresponding to the health benchmark to generate an evidence set, including: the process of generating the sample deviation is as follows: the difference between the real-time status value and the health center within the working condition interval is calculated, and the absolute value of the result is taken as the ratio of the value within the health fluctuation range and the maximum value in the denominator protection value, which is used as the sample deviation. The gradual variable is represented by the slope of the state change of the measuring point within the current time window; The mutation amount is represented by the instantaneous change amplitude of the measurement point at adjacent sampling times or within a short time window; The coupling anomaly value represents the difference between the correlation coefficient of two measurement points and the preset correlation coefficient.

7. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 6, characterized in that, The process of forming the sample deviation, gradual variable, mutation amount, and coupled anomaly amount as evidence objects and constructing an abnormal evidence chain, and determining the condition-based maintenance decision result based on the abnormal evidence chain, is as follows: The sample deviation, gradual variable, mutation amount, and coupling anomaly amount are each encapsulated as evidence objects; The system invokes a pre-defined fault knowledge base to map evidence objects to measurement point anomalies, component statuses, candidate faults, and maintenance objects, thus constructing an anomaly evidence chain. The specific process is as follows: The evidence objects of real-time generated sample deviation, gradual variable, mutation amount and coupling anomaly amount are stored in a unified format; Within a preset time window, evidence objects are clustered and grouped according to their component affiliation. Each evidence group is sequentially matched against the matching rules in the preset fault knowledge base. If a match is successful, the successfully matched evidence group is associated as an abnormal evidence chain. The abnormal evidence chain is input into the preset maintenance decision result generation model, outputs the condition maintenance decision result, and binds the condition maintenance decision result to the specific component of the corresponding hydropower unit. If a match is unsuccessful, the evidence object in the evidence group is marked as unmatched evidence, and an audit alarm is pushed.

8. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 7, characterized in that, The process of updating the operating condition health benchmark, transition isolation window, and abnormal evidence discrimination rules based on the maintenance execution results is as follows: Obtain the maintenance execution results after the maintenance decision is executed according to the condition maintenance decision. The maintenance execution results include confirmed fault, no fault found, normal operation restored after maintenance, and fault found but no early warning was given. When the maintenance execution result shows a confirmed fault, the multi-source status data before, during and after the fault is triggered will be written into the fault sample library. When the inspection result shows that no fault was found, the corresponding evidence object is written into the false alarm record and a prompt to reduce the confidence of the evidence object is sent. When the maintenance execution result shows that the system has returned to normal after maintenance, the corresponding multi-source status data is added to the health sample library as a highly reliable health sample, and the corresponding feedback update target is updated. When the maintenance execution results show that a fault was found but no early warning was given, the status data before the fault was retrieved.

9. The hydropower unit condition-based maintenance system based on multi-source situational awareness as described in claim 5, characterized in that, The establishment of a transition isolation window that matches the response characteristics of the measuring point at the operating condition switching boundary also includes dynamically adjusting the response stabilization waiting time based on the load change magnitude and the guide vane opening change rate. The specific process is as follows: Monitor the load change amplitude and guide vane opening change rate. When the load change amplitude exceeds the preset load change amplitude threshold or the guide vane opening change rate exceeds the preset guide vane opening change rate threshold, input the corresponding load change amplitude and guide vane opening change rate into the transition isolation window's hierarchical model to divide the transition isolation window into sub-intervals, including stable transition isolation windows and unstable transition isolation windows. When the output is a stable transition isolation window, the objectified state data outside the stable transition isolation window is used as a candidate set of stable samples. When the output is an unstable transition isolation window, the objectified state data within the window is prohibited from entering the healthy sample, and the objectified state data outside the window is used as data to be verified, and an abnormal working condition transition prompt is sent.

10. A condition-based maintenance method for hydropower units based on multi-source situational awareness, characterized in that: include: Acquire multi-source status data, operating condition data and maintenance-related data during the operation of hydropower units, establish status data objects based on multi-source status data, perform data quality identification on multi-source status data, and obtain objectified status data with quality identification. Based on the operating data, the operating condition switching boundary of the hydropower unit is identified, and different response stabilization waiting times are set for the measurement point response characteristics corresponding to different types of state variables. A transition isolation window matching the response characteristics of the measuring point is established for the working condition switching boundary. Outside the transition isolation window, objectified state data is selected to form a stable sample candidate set, and the stable sample candidate set is purified to generate a health benchmark associated with the corresponding quality label, working condition range, corresponding component and corresponding measuring point type. Within the transition isolation window, data within the transition isolation window is prohibited from entering the health sample. Evidence is categorized based on real-time status values ​​corresponding to health benchmarks to generate an evidence set, including sample deviation, gradual variable, mutation amount, and coupling anomaly amount; The sample deviation, gradual variable, mutation amount and coupling anomaly amount are formed as evidence objects and an abnormal evidence chain is constructed. The condition maintenance decision result is determined based on the abnormal evidence chain. Obtain the maintenance execution results after the condition-based maintenance decision is executed, and update the operating condition health benchmark, transition isolation window, and abnormal evidence discrimination rules based on the maintenance execution results.

Images