A working condition global regulation hydroelectric generating set multi-source heterogeneous fault diagnosis method and system

Abstract

This invention discloses a method and system for diagnosing multi-source heterogeneous faults in hydropower units with global operating condition control, belonging to the field of intelligent monitoring and diagnosis of power units. This invention uses real-time operating conditions as the global control variable and constructs a full-process dynamic diagnostic architecture: the feature processing layer achieves adaptive filtering and cross-modal feature normalization of multi-source heterogeneous data through operating condition-domain adaptive attention and dynamic alignment of health anchors; the relationship modeling layer constructs a spatiotemporal heterogeneous graph driven by operating conditions, and combines it with a dynamic temporal topological feature map to simultaneously characterize the spatial propagation path and early degradation trend of faults in the temporal dimension; the decision optimization layer uses a two-layer Bayesian network to progressively infer the probabilities of single and compound faults, integrates a fault knowledge graph for causal verification, and outputs interpretable diagnostic results; a matching full-chain closed-loop self-iterative mechanism continuously optimizes the full-link model parameters based on field feedback. This invention effectively improves the decoupling accuracy and early warning capability of compound faults under wide-range variable operating conditions.

Description

Technical Field

[0001] This invention relates to the field of hydropower unit condition monitoring and intelligent fault diagnosis technology, specifically to a method and system for diagnosing vibration faults in hydropower units based on global control of operating conditions, applicable to online vibration monitoring and intelligent diagnosis of complex faults for all types of hydropower units under all operating conditions. Background Technology

[0002] Hydropower units, as core supporting equipment for peak shaving and frequency regulation in power systems, operate under complex dynamic conditions such as variable load, variable head, frequent start-ups and shutdowns, and low-load no-load operation. Abnormal vibration is the most common and critical fault manifestation in hydropower units. Mechanical, hydraulic, and electrical factors can all cause unit vibration, and multiple faults often occur in coupled forms. Accurate, stable, and interpretable vibration fault diagnosis is a key technological foundation for ensuring the safe and stable operation of units, reducing unplanned downtime losses, and achieving condition-based maintenance.

[0003] With the advancement of smart power plant construction, current intelligent vibration diagnosis methods for hydropower units generally adopt a hierarchical fusion architecture that integrates multi-source data fusion, attention mechanisms, graph convolutional networks, and Bayesian inference. These technologies have been publicly disclosed in numerous academic papers and patent documents. However, existing technical solutions suffer from severe homogenization, failing to fully integrate the dynamic characteristics of hydropower unit operation, and exhibiting numerous technical shortcomings in practical engineering applications. First, existing attention mechanisms mostly adopt a global fixed weight design without adaptive adjustment based on the dynamic operating characteristics of the unit. Under different operating conditions such as steady state, start-up and shutdown, variable load, and low load, the sensitivity of various fault characteristics varies significantly. Fixed weight mechanisms can easily lead to effective fault characteristics being drowned out by operating noise, resulting in missed diagnosis of early faults and misdiagnosis of typical faults.

[0004] Second, existing feature alignment methods mostly use simple extreme value normalization or global linear mapping, without considering the influence of factors such as individual differences of units, long-term aging of equipment, and drift of operating condition benchmarks. Under dynamic operating conditions, feature alignment errors are large, resulting in poor stability of diagnostic models and insufficient generalization ability across units and cycles.

[0005] Third, existing graph convolutional networks generally adopt static fixed topology structures. The graph nodes and edge weights remain unchanged for a long time after the model is trained. This makes it impossible to reflect the changes in the dynamic coupling relationship between the measurement points when the unit is operating under different conditions. It is difficult to effectively decouple the coupling characteristics of complex faults caused by multiple factors such as hydraulic, mechanical and electrical factors, resulting in low accuracy of complex fault identification.

[0006] Fourth, existing diagnostic decisions often use single-layer Bayesian networks or single deep learning models to directly output results, lacking constraints from the specific physical mechanisms of hydropower units. The models have poor interpretability and are prone to producing unreasonable diagnostic results that contradict the fault mechanisms of hydropower units, thus failing to meet the engineering implementation requirements of on-site operation and maintenance of hydropower stations.

[0007] Fifth, existing technologies generally lack a full-link closed-loop iterative mechanism. After the diagnostic model is trained, the parameters are fixed and cannot be adaptively updated by random group aging, operating condition deviation, and fault sample accumulation. This leads to a continuous decline in diagnostic accuracy after long-term operation, making it difficult to meet the diagnostic needs of the entire life cycle of the unit. Summary of the Invention

[0008] To address the shortcomings of existing technologies, such as poor adaptability to operating conditions, weak dynamic fault feature extraction capabilities, difficulty in decoupling complex faults, insufficient interpretability of diagnostic results, and inability of models to adaptively iterate, this invention provides a method and system for diagnosing vibration faults in hydropower units based on dynamic control of unit status. By using the real-time operating status of the unit as a global control signal throughout the entire process of feature selection, feature alignment, topology modeling, and fault reasoning, this invention achieves adaptive diagnosis of hydropower units under all operating conditions, accurate decoupling of complex faults, interpretable verification of mechanisms, and full-link self-iterative optimization, significantly improving the accuracy and engineering practicality of vibration fault diagnosis in hydropower units.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a method for diagnosing vibration faults in hydropower units based on dynamic control of unit status, comprising the following steps: S1. Synchronous acquisition of multi-source heterogeneous data and real-time determination of unit operating status: Synchronously acquire unit SCADA operating parameters, multi-point vibration signals, bearing and stator temperature data, oil monitoring data, and unstructured operation and maintenance text data. Based on unit load fluctuation, head fluctuation, speed stability, and guide vane change rate, the unit operating status is divided into multiple typical operating conditions, including steady-state operation, load transition, start-stop transition, and low load. Real-time operating condition tags are bound to all acquired data. S2, Adaptive Heterogeneous Attention Weighted Filtering by Working Condition Domain: Divide multi-source heterogeneous data into multiple data subdomains, adaptively allocate attention weights to each data subdomain according to different working condition labels, and set a data quality confidence gating mechanism to automatically reduce the weight of abnormal data, thereby completing the filtering and enhancement of multi-source original features. S3. Dynamic feature alignment of health anchor points under different operating conditions: Pre-store the standard health benchmark anchor point feature vectors corresponding to each typical operating condition of the unit, and use the relative anchor point deviation normalization method to complete the unified alignment of multimodal features; use a pre-trained language model finely tuned by the domain corpus to complete word vector encoding for unstructured operation and maintenance text, and map it to the unified anchor point feature space; achieve adaptive adjustment of anchor point vectors by periodically updating health samples; S4. Working condition-driven spatiotemporal heterogeneous graph Transformer relationship modeling: Construct a heterogeneous graph network structure; dynamically update the weights of the associated edges between graph nodes based on real-time working condition parameters and feature similarity to achieve adaptive changes in the topology; embed temporal position encoding and use a spatiotemporal multi-head self-attention mechanism to mine the spatial coupling relationship of unit measurement points and the temporal evolution law of faults, and generate a dynamic temporal topology feature map. S5. Two-layer dynamic Bayesian network joint fault knowledge graph verification decision: The first layer Bayesian network calculates the independent probability of a single fault based on the topological graph features; the second layer coupled Bayesian network calls the conditional probability table for different working conditions, decouples the coupling association of composite faults and calculates the joint fault probability; the hydropower unit fault causal knowledge graph is introduced to verify the consistency of the diagnostic results, correct the fault confidence and optimize the root cause location results, and output the fault diagnosis results. S6. Full-link closed-loop adaptive iterative update: Manually verified fault samples and normal operation samples are fed back into the system to adaptively update the full-link model parameters, thereby achieving continuous optimization of the operational status changes of the diagnostic model random group.

[0010] In a preferred embodiment of the present invention, step S1 divides the unit operating status into four typical operating conditions, and the specific determination rules are as follows: Steady-state operating conditions: Unit load fluctuation ≤ ±5%, head fluctuation ≤ ±2%, and speed stable at rated speed for ≥ 10 min; Variable load transition condition: The unit load change rate is ≥2% / min and it is not during the start-up or shutdown phase; Start-up and shutdown transition conditions: the entire process of the unit speed increasing from zero to rated speed or decreasing from rated speed to zero; Low load condition: The unit operates at a load ≤ 30% of the rated load and is not in a start-stop transition state.

[0011] As a preferred embodiment of the present invention, the working condition domain attention weight allocation rule in S2 includes: The weights for enhancing the effective value of vibration, dominant frequency characteristics, and temperature time series characteristics under steady-state operation conditions; the weights for enhancing vibration kurtosis, low-frequency disturbance characteristics, and guide vane change rate characteristics under variable load transition conditions and start-stop transition conditions; the weights for suppressing high-frequency hydraulic noise characteristics and amplifying weak low-frequency fault characteristics under low load conditions; and the data quality confidence gating mechanism automatically reduces the weights of data with missing rates exceeding a threshold, jumps exceeding 3σ, or drift.

[0012] As a preferred embodiment of the present invention, the normalization formula for the relative anchor point deviation in S3 is: Fnorm=(Fraw-Fanchor) / σanchor

[0013] Where Fraw is the original feature value, Fanchor is the health anchor feature value under the corresponding operating condition, and σanchor is the standard deviation of the health sample feature. This normalization eliminates the systematic error caused by individual differences and aging drift of the unit. The anchor vector is generated based on the statistical data of the unit's fault-free operation in the early stage of commissioning, and is incrementally updated using newly added normal samples with a set learning rate.

[0014] In a preferred embodiment of the present invention, the heterogeneous graph network structure in S4 is a dynamically learnable topology, including mechanical measuring point nodes, hydraulic and electrical parameter nodes, and operation and maintenance knowledge nodes; the graph edge weights are randomly grouped and dynamically refreshed in real time under different operating conditions: the association weight between the guide vane opening node and the top cover vibration node is automatically enhanced under variable load conditions; the association weight between the rotor vibration node and the bearing temperature node is automatically enhanced under steady-state conditions; the heterogeneous graph network refreshes its dynamic topology at a set time interval.

[0015] As a preferred embodiment of the present invention, S4 employs a spatiotemporal multi-head self-attention mechanism to replace the traditional static graph convolution operation, simultaneously capturing the spatial topological correlation between measurement points and the temporal degradation trend of faults within a sliding time window, thereby achieving joint feature extraction of spatial and temporal dimensions.

[0016] As a preferred embodiment of the present invention, the two-layer dynamic Bayesian network in S5 is configured with independent conditional probability tables for different operating conditions. The four types of operating conditions—steady-state operation, variable load transition, start-stop transition, and low load—each call an independent probability parameter library to avoid inference distortion of the static probability model under dynamic operating conditions.

[0017] As a preferred embodiment of the present invention, the fault causal knowledge graph in S5 incorporates the fault coupling mechanism rules of hydropower units, which are used to verify the causal matching relationship between hydraulic excitation, rotor imbalance, bearing wear, guide vane jamming, and abnormal electromagnetic vibration. For diagnostic results that do not conform to the physical mechanism, the confidence level is automatically lowered and the associated faults are re-searched. The diagnostic results include fault type, fault location, and verified confidence level, and the maintenance priority is calculated based on the urgency of the fault type and the confidence level.

[0018] As a preferred embodiment of the present invention, the full-link closed-loop adaptive iterative update in S6 includes: synchronously updating the working condition attention weight matrix, health anchor feature vector, heterogeneous graph topology edge weight, and Bayesian network conditional probability table; after the model is updated, backtesting is performed using a historical fault sample set to ensure that the diagnostic accuracy after the update is not lower than the level before the update.

[0019] This invention also provides a hydropower unit vibration fault diagnosis system based on dynamic control of unit status, for implementing the above method, comprising: The data acquisition and operating condition determination module is used to synchronously collect multi-source heterogeneous data and determine the unit's operating condition in real time, and to bind operating condition tags to the collected data. The working condition domain feature filtering module is used to adaptively allocate attention weights for each data subdomain based on the working condition label, and to filter and enhance the original features through a data quality gating mechanism. The health anchor feature alignment module is used to achieve multimodal feature normalization alignment based on health anchors under different working conditions, and to complete cross-modal feature space mapping. The spatiotemporal heterogeneous graph modeling module is used to construct dynamic topological heterogeneous graph networks and mine spatial coupling and temporal evolution features through a spatiotemporal multi-head self-attention mechanism. The two-layer Bayesian decision module is used to calculate the probability of single and compound faults through a two-layer dynamic Bayesian network, and to complete mechanism verification and output diagnostic results by combining the fault knowledge graph. The closed-loop iterative update module is used to update the model parameters across the entire process based on the verified sample data, thereby enabling adaptive optimization of the diagnostic system.

[0020] This invention uses the real-time operating condition of the unit as a global control variable throughout the entire diagnostic process, and has the following significant technical advantages compared with existing technologies: (1) It pioneered a global full-link control mechanism for operating conditions, which uses the dynamic operating status of the unit as the top-level control signal to run through the entire process of attention allocation, feature alignment, graph topology modeling, and probabilistic reasoning. This completely solves the industry pain point that traditional fixed models cannot adapt to dynamic operating conditions such as variable load, start-up and shutdown, and low load, and significantly reduces the false alarm rate and false alarm rate. (2) By adopting the working condition domain heterogeneous attention combined with the data quality gating mechanism, feature weighting rules are customized for different working conditions of hydropower units to achieve adaptive enhancement of fault features and adaptive suppression of noise. The feature extraction accuracy is significantly better than the existing fixed weight scheme. (3) A dynamic alignment mechanism for health anchor points under different operating conditions is proposed. The traditional fixed normalization method is abandoned. The relative anchor point deviation normalization is adopted to effectively eliminate the systematic errors caused by individual differences of units, equipment aging, and operating condition drift. The characteristic stability and cross-operating condition generalization ability are significantly improved. (4) The use of spatiotemporal heterogeneous graph Transformer dynamic topology modeling to replace the traditional static graph convolutional network can update the topological relationship between equipment measurement points in real time under random working conditions, accurately capture the spatiotemporal coupling characteristics of compound faults, and greatly improve the diagnostic accuracy under multiple fault coupling conditions. (5) A two-layer dynamic Bayesian network combined with a fault knowledge graph is used for mechanism verification to achieve deep integration of data-driven and physical mechanisms, solve the problem of poor interpretability of pure deep learning models and easy occurrence of diagnostic results that violate engineering mechanisms, and the diagnostic results have strong engineering applicability; (6) It has the capability of full-link closed-loop self-iteration, and can continuously optimize the full-link model parameters according to changes in the unit's operating status and the accumulation of fault samples, so as to ensure the stability of diagnostic accuracy throughout the unit's entire life cycle. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the hydropower unit vibration fault diagnosis method provided in this embodiment; Figure 2 This embodiment provides a logic flowchart for classifying and determining the four types of operating conditions of the generating unit. Figure 3 This is a flowchart of the adaptive heterogeneous attention feature processing for different operating conditions provided in this embodiment; Figure 4 This is a flowchart of the spatiotemporal heterogeneous graph Transformer dynamic topology modeling process provided in this embodiment; Figure 5 The flowchart for the two-layer dynamic Bayesian network combined with knowledge graph verification decision-making provided in this embodiment is shown. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment uses a large mixed-flow hydropower unit as an application object to illustrate the technical solution in detail. The technical solution is also applicable to vibration fault diagnosis of various types of hydropower units such as axial flow, through flow, and impulse flow.

[0023] like Figure 1 As shown, the hydropower unit vibration fault diagnosis method based on dynamic control of unit status proposed in this invention includes six core steps. It achieves full-process adaptive intelligent diagnosis through global control of unit operating conditions. The specific implementation methods of each step are as follows: Step 1: Multi-source heterogeneous data acquisition and accurate operating condition determination. Simultaneously acquire X, Y, and Z-axis vibration data of key components such as the upper frame, lower frame, top cover, stator, and rotor of the unit, with the vibration signal sampling frequency set to 1kHz; simultaneously acquire SCADA operating parameters such as active power, operating head, unit speed, guide vane opening, and excitation current; temperature data of components such as the upper guide bearing, lower guide bearing, thrust bearing, and stator winding; oil monitoring indicators such as oil particle size and oil viscosity; and unstructured maintenance text data such as historical maintenance records, defect logs, and operation logs.

[0024] according to Figure 2The process shown uses real-time computer data to define the unit's operating status characteristics, accurately classifying the unit's operating status into four typical conditions: steady-state operation, variable load transition, start-stop transition, and low load. Each set of collected data is then labeled with a corresponding real-time operating condition tag, providing a foundation for subsequent adaptive control of the entire process. The specific threshold values ​​for the four operating conditions have been verified and optimized using actual operating data from multiple power plants, ensuring the accuracy of the operating condition classification and its applicability to engineering projects.

[0025] Step 2: Heterogeneous attention-weighted filtering based on operating conditions. The collected multi-source heterogeneous data is divided into three data subdomains: time-series numerical data, operation and maintenance text, and operating condition status.

[0026] according to Figure 3 The weight allocation mechanism shown dynamically adjusts the attention weights of each subdomain and specific features within each subdomain for different operating conditions: Under steady-state conditions, the weighting coefficients of slowly varying characteristics such as vibration dominance frequency, effective vibration value, and bearing temperature should be emphasized. Weighting coefficients for rapidly changing impact characteristics such as enhanced vibration kurtosis, low-frequency wavelet packet disturbance energy, and guide vane change rate under variable load transition and start-stop transition conditions; Weighting of active suppression of high-frequency hydraulic noise characteristics under low-load conditions, and weighting of amplification of weak low-frequency fault characteristics such as tailrace vortex bands.

[0027] At the same time, a data quality confidence gating mechanism is activated. For abnormal data with a missing rate greater than 10%, numerical jumps exceeding 3σ, or slow drift, the weight is automatically reduced to 30% of the original weight. This completes the adaptive purification and enhancement of the original features, avoiding interference from low-quality data with the diagnostic results.

[0028] Step 3: Dynamic feature alignment of health anchor points under different operating conditions. Based on the fault-free normal operation data of the unit during the first 3 months of commissioning, the unit health benchmark anchor point feature vectors corresponding to the four operating conditions are statistically generated, and a health anchor point feature library for different operating conditions is constructed.

[0029] A relative anchor point deviation normalization method is used to normalize and align all numerical features. The normalization formula is Fnorm=(Fraw-Fanchor) / σanchor, where Fraw is the original feature value, Fanchor is the health anchor point feature value under the corresponding operating condition, and σanchor is the feature standard deviation of the healthy sample. This normalization method uses the unit's own health status as a benchmark, which can effectively eliminate systematic errors caused by individual unit differences, equipment aging, and differences in operating condition benchmarks.

[0030] For unstructured operation and maintenance text data, a BERT model fine-tuned with hydropower unit terminology is used to encode word vectors. These vectors are then mapped through a fully connected layer to a 128-dimensional anchor feature space that is consistent with the numerical features, achieving a unified representation of cross-modal features. Daily, health anchor vector parameters are updated with a small learning rate increment of 0.01 using manually confirmed normal operation samples to adaptively adapt to the slow aging process of the units.

[0031] Step 4: Working Condition-Driven Spatiotemporal Heterogeneous Graph Transformer Modeling. Construct a heterogeneous graph network structure containing three types of nodes: the first type is mechanical measurement point nodes, corresponding to various vibration, temperature, and other sensor measurement points; the second type is hydraulic and electrical parameter nodes, corresponding to operating parameters such as load, head, guide vane opening, and excitation current; the third type is maintenance knowledge nodes, corresponding to textual semantic features such as historical defects and maintenance records.

[0032] according to Figure 4 The process shown first calculates the initial correlation weights between nodes based on real-time operating parameters, and then corrects the edge weights using characteristic cosine similarity: under variable load conditions, the edge weight between "guide vane opening - top cover vibration" is automatically increased to twice the baseline value, strengthening the representation of the hydraulic factor transmission path; under steady-state conditions, the edge weight between "rotor vibration - thrust bearing temperature" is automatically increased to 1.8 times the baseline value, strengthening the representation of the mechanical factor transmission path. The above weight amplification coefficients are determined based on the statistical analysis of the transmission sensitivity of corresponding correlation paths in historical unit failure cases, and different units can be calibrated on-site based on actual operating data.

[0033] The temporal position encoding with an embedded 10-second sliding window is used to replace the traditional static graph convolution operation with an 8-head spatiotemporal self-attention mechanism. It captures the dynamic coupling relationship between each measurement point in the spatial dimension and explores the temporal degradation and evolution law of the fault in the temporal dimension. The dynamic topological feature map is refreshed every 10 seconds to achieve spatiotemporal joint feature extraction.

[0034] Step 5: A two-layer dynamic Bayesian network combined with a knowledge graph validates the decision. For example... Figure 5 As shown, diagnostic decision-making is divided into two levels and incorporates mechanism verification: The first layer is a single-fault Bayesian network. The input is a spatiotemporal topological feature map, and the output is the independent occurrence probability of six typical single faults: rotor imbalance, rotor misalignment, bearing wear, guide vane jamming, hydraulic excitation, and electromagnetic vibration anomaly. The second layer is a coupled fault Bayesian network, which calls the independent conditional probability table under the corresponding working condition to decouple the coupling relationship of typical composite faults such as guide vane jamming + hydraulic vibration and bearing wear + rotor misalignment, and calculates the joint occurrence probability of various composite faults.

[0035] The two-layer Bayesian network is configured with independent conditional probability table parameters for each of the four operating conditions, avoiding inference distortion of the static probability model under different operating conditions. The preliminary diagnostic results output by the Bayesian network are input into the built-in knowledge graph of the causal relationship of hydropower unit faults for mechanism consistency verification: for example, if the preliminary diagnosis is rotor imbalance, but the corresponding bearing temperature does not rise abnormally, the confidence of the fault is automatically reduced by 30%, and the associated faults are re-searched; if the diagnostic results are consistent with the fault causal mechanism, the confidence is maintained.

[0036] The maintenance priority is calculated based on the urgency of the fault type and the confidence level after verification. The urgency level is divided into three levels: Level I (requires immediate shutdown and handling), Level II (requires maintenance to be arranged within 72 hours), and Level III (can be handled according to the plan). The priority ranking rules are as follows: Level I faults are ranked from high to low confidence level, Level II faults take precedence over Level III faults, and faults of the same level are ranked from high to low confidence level.

[0037] The final output is a structured diagnostic result that includes the fault type, fault location, post-verification confidence level, and corresponding maintenance priority, providing interpretable decision-making basis for operation and maintenance personnel.

[0038] Step Six: End-to-End Closed-Loop Adaptive Iterative Update. Establish a manual review and feedback mechanism for diagnostic results, periodically feeding back fault samples confirmed on-site by maintenance personnel and newly added normal operating samples into the diagnostic system. The system automatically performs an end-to-end parameter update weekly: optimizing the operating condition domain attention weight matrix based on new samples, fine-tuning the health anchor feature vector, updating the heterogeneous graph network topology edge weights, and correcting the Bayesian network conditional probability table parameters. After each model parameter update, backtesting is performed using a historical fault sample database to ensure that the diagnostic accuracy of the updated model is not lower than the level before the update. This achieves long-term adaptive optimization of the diagnostic model in response to random group aging, operating condition deviations, and fault sample accumulation, ensuring the stability of diagnostic accuracy throughout the unit's entire lifecycle.

[0039] Accordingly, this embodiment also provides a hydropower unit vibration fault diagnosis system based on dynamic control of unit status, including: The data acquisition and operating condition determination module is used to synchronously collect multi-source heterogeneous data and determine the unit's operating condition in real time, and to bind operating condition tags to the collected data. The working condition domain feature filtering module is used to adaptively allocate attention weights for each data subdomain based on the working condition label, and to filter and enhance the original features through a data quality gating mechanism. The health anchor feature alignment module is used to achieve multimodal feature normalization alignment based on health anchors under different working conditions, and to complete cross-modal feature space mapping. The spatiotemporal heterogeneous graph modeling module is used to construct dynamic topological heterogeneous graph networks and mine spatial coupling and temporal evolution features through a spatiotemporal multi-head self-attention mechanism. The two-layer Bayesian decision module is used to calculate the probability of single and compound faults through a two-layer dynamic Bayesian network, and to complete mechanism verification and output diagnostic results by combining the fault knowledge graph. The closed-loop iterative update module is used to update the model parameters across the entire process based on the verified sample data, thereby enabling adaptive optimization of the diagnostic system.

[0040] The specific implementation methods of each module correspond one-to-one with the above-described steps, and will not be repeated here.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing multi-source heterogeneous faults in hydropower units with global control under operating conditions, characterized in that, Includes the following steps: S1. Synchronous acquisition of multi-source heterogeneous data and real-time determination of unit operating status: Synchronously acquire unit SCADA operating parameters, multi-point vibration signals, bearing and stator temperature data, oil monitoring data, and unstructured operation and maintenance text data. Based on unit load fluctuation, head fluctuation, speed stability, and guide vane change rate, the unit operating status is divided into multiple typical operating conditions, including steady-state operation, load transition, start-stop transition, and low load. Real-time operating condition tags are bound to all acquired data. S2, Adaptive Heterogeneous Attention Weighted Filtering by Working Condition Domain: Divide multi-source heterogeneous data into multiple data subdomains, adaptively allocate attention weights to each data subdomain according to different working condition labels, and set a data quality confidence gating mechanism to automatically reduce the weight of abnormal data, thereby completing the filtering and enhancement of multi-source original features. S3. Dynamic feature alignment of health anchor points under different operating conditions: Pre-store the feature vectors of the unit's standard health benchmark anchor points corresponding to each typical operating condition, and use the relative anchor point deviation normalization method to complete the unified alignment of multimodal features; use a pre-trained language model finely tuned by the domain corpus to complete word vector encoding for unstructured operation and maintenance text, and map it to the unified anchor point feature space. The anchor vector is adaptively adjusted by periodically updating healthy samples; the multi-source features after attention filtering are fused with the normalized features after anchor alignment to construct a unified multimodal feature tensor as the input for the next relationship modeling layer. S4. Working condition-driven spatiotemporal heterogeneous graph Transformer relationship modeling: Construct a heterogeneous graph network structure; dynamically update the weights of the associated edges between graph nodes based on real-time working condition parameters and feature similarity to achieve adaptive changes in the topology; embed temporal position encoding and use a spatiotemporal multi-head self-attention mechanism to mine the spatial coupling relationship of unit measurement points and the temporal evolution law of faults, and generate a dynamic temporal topology feature map. S5. Two-layer dynamic Bayesian network joint fault knowledge graph verification decision: The first layer Bayesian network calculates the independent probability of a single fault based on the topological graph features; the second layer coupled Bayesian network calls the conditional probability table for different working conditions, decouples the coupling association of composite faults and calculates the joint fault probability; the hydropower unit fault causal knowledge graph is introduced to verify the consistency of the diagnostic results, correct the fault confidence and optimize the root cause location results, and output the fault diagnosis results. S6. Full-link closed-loop adaptive iterative update: Manually verified fault samples and normal operation samples are fed back into the system to adaptively update the full-link model parameters, thereby achieving continuous optimization of the operational status changes of the diagnostic model random group.

2. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 1, characterized in that, In step S1, the unit's operating status is divided into four typical operating conditions, and the specific judgment rules are as follows: Steady-state operating conditions: Unit load fluctuation ≤ ±5%, head fluctuation ≤ ±2%, and speed stable at rated speed for ≥ 10 min; Variable load transition condition: The unit load change rate is ≥2% / min and it is not during the start-up or shutdown phase; Start-up and shutdown transition conditions: the entire process of the unit speed increasing from zero to rated speed or decreasing from rated speed to zero; Low load condition: The unit's operating load is ≤30% of the rated load and it is not in a start-stop transition state; the above condition judgment thresholds are pre-calibrated based on the unit's rated parameters and actual on-site operating conditions, and on-site maintenance personnel are allowed to set and adjust them.

3. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 2, characterized in that, The attention weight allocation rules for working condition domains in S2 include: The attention score coefficient of each data subdomain is dynamically calculated based on the operating condition category. The higher the attention score, the greater the contribution of the subdomain features in subsequent processing. Specifically, under steady-state operating conditions, higher attention weights are assigned to the effective vibration value, dominant frequency features, and temperature time series features. Under variable load transition and start-stop transition operating conditions, higher attention weights are assigned to the vibration kurtosis, low-frequency disturbance features, and guide vane change rate features. Under low load operating conditions, lower attention weights are assigned to the high-frequency hydraulic noise features, while the weight of weak low-frequency fault features is amplified. The data quality confidence gating mechanism automatically reduces the weight of data with missing rates exceeding the threshold, jumps exceeding 3σ, or drift.

4. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 1, characterized in that, The normalization formula for the relative anchor point deviation in S3 is: Fnorm=(Fraw-Fanchor) / σanchor Where Fraw is the original feature value, Fanchor is the health anchor feature value under the corresponding operating condition, and σanchor is the standard deviation of the health sample feature. This normalization eliminates the systematic error caused by individual differences and aging drift of the unit. The anchor vector is generated based on the statistical data of the unit's fault-free operation in the early stage of commissioning, and is incrementally updated using newly added normal samples with a set learning rate.

5. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 1, characterized in that: The heterogeneous graph network structure in S4 is a dynamically learnable topology, including mechanical measuring point nodes, hydraulic and electrical parameter nodes, and operation and maintenance knowledge nodes. The operation and maintenance knowledge nodes participate in graph calculation using semantic embedding vectors of operation and maintenance text. The graph edge weights are dynamically refreshed in real time under different operating conditions: under variable load conditions, the association weight between the guide vane opening node and the top cover vibration node is automatically enhanced; under steady-state conditions, the association weight between the rotor vibration node and the bearing temperature node is automatically enhanced. The heterogeneous graph network refreshes its dynamic topology at set time intervals.

6. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 1, characterized in that: The S4 mechanism uses a spatiotemporal multi-head self-attention mechanism to replace the traditional static graph convolution operation. Within the sliding time window, it simultaneously captures the spatial topological relationship between measurement points and the temporal degradation trend of faults, thereby achieving joint feature extraction of spatial and temporal dimensions.

7. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 1, characterized in that, The S5 dual-layer dynamic Bayesian network is configured with independent conditional probability tables for different operating conditions. The four types of operating conditions—steady-state operation, variable load transition, start-stop transition, and low load—call independent probability parameter libraries respectively, avoiding inference distortion of the static probability model under dynamic operating conditions.

8. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 1, characterized in that, The fault causal knowledge graph in S5 incorporates rules for the coupling mechanism of hydropower unit faults. It is used to verify the causal matching relationship between hydraulic excitation, rotor imbalance, bearing wear, guide vane jamming, and abnormal electromagnetic vibration. For diagnostic results that do not conform to the physical mechanism, the confidence level is automatically lowered and the associated faults are re-searched. The diagnostic results include fault type, fault location, and verified confidence level. The maintenance priority is calculated based on the urgency of the fault type and the confidence level.

9. The method for diagnosing multi-source heterogeneous faults in hydropower units with global control of operating conditions according to claim 1, characterized in that, The full-link closed-loop adaptive iterative update in S6 includes: synchronously updating the working condition attention weight matrix, health anchor feature vector, heterogeneous graph topology edge weights, and Bayesian network conditional probability table; after the model is updated, it is backtested and verified using a historical fault sample set, with the updated diagnostic accuracy not being lower than the level before the update as the condition for accepting the update, and if it is lower, it is rolled back to the parameter version before the update.

10. A multi-source heterogeneous fault diagnosis system for hydropower units with global operating condition control, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition and operating condition determination module is used to synchronously collect multi-source heterogeneous data and determine the unit's operating condition in real time, and to bind operating condition tags to the collected data. The working condition domain feature filtering module is used to adaptively allocate attention weights for each data subdomain based on the working condition label, and to filter and enhance the original features through a data quality gating mechanism. The health anchor feature alignment module is used to achieve multimodal feature normalization alignment based on health anchors under different working conditions, and to complete cross-modal feature space mapping. The spatiotemporal heterogeneous graph modeling module is used to construct dynamic topological heterogeneous graph networks and mine spatial coupling and temporal evolution features through a spatiotemporal multi-head self-attention mechanism. The two-layer Bayesian decision module is used to calculate the probability of single and compound faults through a two-layer dynamic Bayesian network, and to complete mechanism verification and output diagnostic results by combining the fault knowledge graph. The closed-loop iterative update module is used to update the model parameters across the entire process based on the verified sample data, thereby enabling adaptive optimization of the diagnostic system.

Images