CN-121980148-A - Marine fan vibration data missing recovery system based on multi-head output prediction mechanism

CN121980148ACN 121980148 ACN121980148 ACN 121980148ACN-121980148-A

Abstract

The invention discloses a multi-head output prediction mechanism-based offshore wind turbine vibration data missing recovery system, in particular to a deep learning method in the field of equipment monitoring and business operation and maintenance decision support, which comprises five modules including time sequence state evaluation feature extraction, cost-associated space influence analysis, multi-mechanism risk prediction and value evaluation, credible fusion and operation and maintenance decision generation and multi-objective decision model training, extracting the time sequence and the spatial characteristics of a fan vibration signal, predicting a periodic component, a transient impact component and a background vibration component, dynamically fusing and correcting a multipath prediction result, outputting a recovered complete vibration sequence, and adopting multi-objective loss function joint optimization. The invention can effectively improve the accuracy and reliability of the loss recovery of the vibration data of the offshore wind turbine, and provides direct decision support for equipment residual life assessment, preventive maintenance work order generation, spare part purchasing plan and insurance rate assessment.

Inventors

LI DONG
Wu Fanxi
WU KEDA
LIAO YIZHEN
Lin Zhanyan
MA JUNQI

Assignees

福州大学

Dates

Publication Date: 20260505
Application Date: 20260129

Claims (8)

1. Offshore wind turbine vibration data loss recovery system based on bull output prediction mechanism, its characterized in that includes: the time sequence state evaluation feature extraction module is used for performing time sequence feature learning on the fan vibration signal by adopting a deep time sequence neural network and outputting a deep time sequence feature vector for evaluating the business operation state and the health degree of the fan; The cost-associated space influence analysis module is used for integrating fault propagation and maintenance cost association based on dynamic correlation coefficients, constructing a dynamic graph structure, aggregating multiple measuring point characteristics through a graph convolution network, and generating a joint characteristic representation for fault positioning and maintenance resource scheduling optimization; The multi-mechanism risk prediction and value evaluation module is based on a physical mechanism and comprises three parallel artificial intelligent prediction sub-networks, wherein the input end of the sub-networks is coupled with the joint characteristic representation, prediction sequences of periodic components, transient impact components and background vibration components in vibration signals are respectively output, and a device residual service life evaluation report, a sudden fault risk level and environmental adaptability evaluation are generated based on the prediction results of the components; The credible fusion and operation and maintenance decision generation module calculates the time-varying confidence scores of the output sequences of all prediction sub-networks through a confidence evaluation network, dynamically fuses and refines the multipath prediction results based on a gating mechanism and a residual error learning network, and outputs a health status report with quantized confidence, wherein the report is used for generating a preventive maintenance work order, a spare part purchasing plan and a insurance rate evaluation basis; And the multi-objective decision model training module is used for jointly optimizing the neural network parameters of the modules through a multi-objective loss function, wherein the loss function comprises a loss item for restricting the data to restore the authenticity, a loss item for restricting the output smoothness and a loss item for restricting the physical compliance of each sub-network prediction result, and training a decision model based on the historical operation and maintenance data to finally output vibration restoration data and decision support information meeting the business operation and maintenance management requirements.
2. The offshore wind turbine vibration data missing recovery system based on the multi-head output prediction mechanism of claim 1, wherein the time sequence state evaluation feature extraction module comprises a short-term feature capturing unit and a long-term dependence modeling unit, the outputs of the two units are subjected to feature fusion through an adaptive weight fusion strategy, a deep time sequence feature vector is output and used for generating a wind turbine health degree evaluation report, and the report serves for optimizing a patrol inspection and maintenance period.
3. The offshore wind turbine vibration data missing recovery system based on the multi-head output prediction mechanism of claim 1 is characterized in that the cost-associated space influence analysis module specifically comprises an original signal cross-correlation value between measuring points, a physical association weight based on a mechanical connection relation, a real-time working condition correction factor and a maintenance resource scheduling scheme for fault location and optimization, wherein the association information based on optimization comprises maintenance cost association fused in a dynamic graph structure.
4. The marine fan vibration data loss recovery system based on a multi-headed output prediction mechanism of claim 1, wherein the multi-machine risk prediction and value assessment module comprises a first sub-network, a second sub-network, and a third sub-network, wherein the first sub-network is a periodic component prediction sub-network for generating a device remaining useful life assessment report associated with a preventative maintenance schedule.
5. The marine fan vibration data loss recovery system based on the multi-head output prediction mechanism of claim 4, wherein the second sub-network is a transient impact component prediction sub-network for generating a sudden failure risk level, wherein the risk level is associated with generating a high priority maintenance work order and as a safe rate assessment basis.
6. The marine fan vibration data loss recovery system based on the multi-headed output prediction mechanism of claim 4, wherein the third subnetwork is a background vibration component prediction subnetwork for generating an environmental suitability assessment, and wherein the assessment is associated with the determination of a long-term operation and maintenance strategy of the equipment.
7. The system for recovering the vibration data of the offshore wind turbine on the basis of the multi-head output prediction mechanism as set forth in claim 1, wherein the trusted fusion and operation decision generation module specifically comprises a confidence evaluation sub-network and a gating selection unit, outputs a health status report, and is additionally provided with a quantitative confidence, and is configured to be directly used for generating a preventive maintenance work order, a spare part purchasing plan and a insurance rate evaluation basis.
8. The multi-head output prediction mechanism-based offshore wind turbine vibration data loss recovery system of claim 1, wherein the multi-objective decision model training module adopts a multi-objective joint learning mechanism, a loss function comprises authenticity loss, smoothness loss and physical loss, the loss function is dynamically adjusted in the training process through a gradient weight self-adaptive mechanism, the cooperative optimization of neural network parameters of each module is realized, and vibration recovery data and decision support information meeting business operation and maintenance management requirements are output.

Description

Marine fan vibration data missing recovery system based on multi-head output prediction mechanism Technical Field The invention relates to the technical field of deep learning in equipment monitoring and business operation and maintenance decision support, in particular to a marine fan vibration data loss recovery system based on a multi-head output prediction mechanism. Background The safe and stable running of the offshore wind turbine is highly dependent on continuous monitoring of the structural vibration state of the offshore wind turbine, time sequence data acquired by the vibration sensor is a key basis for assessing the health state of the wind turbine and identifying early faults, however, the severe running environment on the sea often causes local loss of vibration signals, and the accuracy of subsequent data analysis and diagnosis is seriously affected. In the prior art, a deep learning method based on the combination of a graph neural network and time sequence prediction is used for carrying out time sequence modeling on a multi-measuring-point vibration signal, extracting depth characteristics, constructing a measuring-point spatial relation graph, utilizing a graph convolution network to aggregate spatial characteristics, reconstructing a missing part through a unified time sequence prediction network, and carrying out end-to-end training in a process of depending on a known segment to aim at realizing a minimum reconstruction error. However, when the method is actually used, the method still has some defects, such as difficulty in capturing physical components with obvious differences in vibration signals at the same time by adopting a single prediction network, distortion of recovered signals under complex working conditions, failure to fully consider prior knowledge of physical connection relations, working condition parameters and the like when multi-measuring-point information is fused, weak physical interpretability of spatial modeling, lack of a mechanism for evaluating dynamic confidence of multi-path prediction results and correcting refined residual errors, and insufficient robustness of recovery when the signal-to-noise ratio of data is low or a missing section is long. Disclosure of Invention In order to overcome the defects in the prior art, the embodiment of the invention provides a multi-head output prediction mechanism-based offshore wind turbine vibration data missing recovery system, which solves the problems in the prior art through the following scheme. In order to achieve the aim, the invention provides the technical scheme that the offshore wind turbine vibration data missing recovery system based on the multi-head output prediction mechanism comprises a time sequence state evaluation feature extraction module, a time sequence feature analysis module and a time sequence analysis module, wherein the time sequence feature extraction module is used for performing time sequence feature learning on a wind turbine vibration signal and outputting a time sequence feature vector for evaluating the commercial running state and the health degree of the wind turbine; The cost-associated space influence analysis module is used for integrating fault propagation and maintenance cost association based on dynamic correlation coefficients, constructing a dynamic graph structure, aggregating multiple measuring point characteristics through a graph convolution network, and generating a joint characteristic representation for fault positioning and maintenance resource scheduling optimization; The multi-mechanism risk prediction and value evaluation module is based on a physical mechanism and comprises three parallel artificial intelligent prediction sub-networks, wherein the input end of the sub-networks is coupled with the joint characteristic representation, prediction sequences of periodic components, transient impact components and background vibration components in vibration signals are respectively output, and a device residual service life evaluation report, a sudden fault risk level and environmental adaptability evaluation are generated based on the prediction results of the components; The credible fusion and operation and maintenance decision generation module calculates the time-varying confidence scores of the output sequences of all prediction sub-networks through a confidence evaluation network, dynamically fuses and refines the multipath prediction results based on a gating mechanism and a residual error learning network, and outputs a health status report with quantized confidence, wherein the report is used for generating a preventive maintenance work order, a spare part purchasing plan and a insurance rate evaluation basis; And the multi-objective decision model training module is used for jointly optimizing the neural network parameters of the modules through a multi-objective loss function, wherein the loss function comprises a loss item for restricting the data to resto