A method for monitoring the quality of a marine wind power main engine room structural member

Abstract

The application discloses a kind of offshore wind power main engine cabin structural member quality monitoring method, specifically related to structural health monitoring technical field, including in structural member surface preposition flexible sensing array network, according to stress concentration and corrosion sensitivity differentiation configuration three kinds of sensors;In the interior of structural member, the acoustic waveguide network of co-curing with matrix is embedded, and the internal propagation signal acquisition is realized by coded acoustic wave excitation and reception;Under the action of environmental load, the response of multiple physical fields and acoustic signals are synchronously collected, and time-space alignment and feature extraction are carried out;Based on deep space-time graph convolution network, a structural state evolution model is constructed, the sensing data is mapped to a dynamic graph structure, and the node and edge association relationship is dynamically updated through an adaptive graph learning mechanism;Using multi-layer space-time convolution and attention fusion mechanism, three-dimensional visualization positioning and quantitative evaluation of structural member internal damage initiation, surface crack propagation and material degradation degree are realized.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, and more specifically, to a method for monitoring the quality of structural components of offshore wind turbine nacelles. Background Technology

[0002] Offshore wind turbine nacelle structural components operate in a complex and harsh marine environment for extended periods, enduring multiple environmental stresses. This can easily lead to fatigue cracks, internal damage, and material degradation, seriously threatening the safety and service life of the entire turbine. Therefore, real-time and accurate quality monitoring and early damage warning of key structural components of the nacelle have become a core technical challenge for ensuring the safe and stable operation of offshore wind farms, and are of great significance for improving operation and maintenance efficiency and reducing the total life cycle cost.

[0003] The existing technical solution involves attaching or embedding a fiber optic grating sensor array on the surface of the structural component to monitor strain and temperature distribution. Simultaneously, several piezoelectric ceramic sensors are arranged on the surface of the structural component as active acoustic emission probes to periodically excite high-frequency acoustic signals and receive their propagation response in the structure. The strain field change is obtained by analyzing the wavelength drift of the fiber optic grating, and the damage is located and assessed by combining the arrival time difference and energy attenuation of the acoustic signal.

[0004] However, in practical applications, it still has some shortcomings, such as the limited configuration and spatial resolution of the sensor network; the surface-mounted sensors are not sensitive to early internal damage and are difficult to optimize based on structural stress distribution and material degradation sensitivity; active acoustic emission and static strain monitoring are often processed independently at the data level, lacking effective spatiotemporal alignment and multiphysics deep coupling analysis methods, making it difficult to reveal the complete mechanism of damage evolution; data processing methods are mostly based on traditional signal processing and simple models, which are insufficient in identifying complex damage modes (such as internal microcrack initiation, corrosion and fatigue coupling), and the evaluation results have low visualization, making it difficult to achieve accurate three-dimensional localization and quantification of damage. The method proposed in this invention, through the fusion design of heterogeneous flexible sensor array network and built-in acoustic waveguide network, combined with deep spatiotemporal graph convolution and adaptive graph learning mechanisms, can effectively overcome the above shortcomings and achieve comprehensive and intelligent structural state monitoring and evaluation from appearance to interior and from local to global. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a method for quality monitoring of offshore wind turbine nacelle structural components, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for quality monitoring of offshore wind turbine nacelle structural components, comprising S1: pre-installing a flexible sensor array network on the surface of the nacelle structural components, wherein the flexible sensor array network is composed of heterogeneous sensor units in a non-uniform topology, and each sensor unit establishes a dynamic data path through a self-organizing protocol, wherein the sensor units include three types of sensors: strain, vibration and corrosion potential, and the arrangement density and type are differentiated according to the stress concentration coefficient and material degradation sensitivity of each region of the structural components;

[0007] S2: An acoustic waveguide network is embedded inside the structural component. The acoustic waveguide network is composed of a prefabricated waveguide array. The waveguide and the structural component body material are co-cured to form an array of acoustic wave conduction paths. The coded acoustic wave signal is generated through the excitation end, and the acoustic wave signal after propagation inside the structural component is collected at the receiving end.

[0008] S3: During the environmental load coupling stage, multi-physics response data are synchronously collected through a flexible sensor array network, and acoustic wave signals propagating inside the structure are obtained through an acoustic waveguide network. The two types of signals are spatiotemporally aligned to extract strain distribution characteristics, vibration mode characteristics and acoustic wave propagation characteristic parameters.

[0009] S4: Based on a deep spatiotemporal graph convolutional network, a structural state evolution model is constructed, which maps the data collected by the flexible sensor array network into a dynamic graph structure. Nodes are sensor units, and edges are the physical association strength between sensor units. At the same time, the acoustic wave propagation characteristic parameters are used as global attributes of the graph, and the association relationship between nodes and edges is dynamically updated through an adaptive graph learning mechanism.

[0010] S5: Based on the updated dynamic graph structure, local damage features and global degradation patterns of the structure are extracted through multi-layer spatiotemporal convolution. The internal propagation features obtained by the acoustic waveguide network are fused using the attention mechanism to achieve three-dimensional visualization, localization and quantitative assessment of internal damage initiation, surface crack propagation and material degradation degree of structural components.

[0011] The technical effects and advantages of this invention are as follows:

[0012] This invention proposes a collaborative monitoring architecture of "heterogeneous flexible sensor array + built-in acoustic waveguide network" to achieve integrated acquisition of multi-dimensional signals from the surface to the interior of the structure. By configuring the density and type of the sensing units differently according to the stress concentration factor and corrosion sensitivity, and by co-curing the waveguide with the substrate material, it not only significantly improves the overall spatial resolution and information richness of the monitoring network, but also enhances the early detection capability of hidden defects inside the structure (such as the initiation of microcracks), effectively solving the problems of traditional methods being insensitive to internal damage and having large monitoring blind spots.

[0013] This invention employs an intelligent fusion analysis method based on deep spatiotemporal graph convolutional networks and adaptive graph learning to map heterogeneous multi-source monitoring data into a dynamically evolving graph structure. Through an adaptive mechanism, it learns the changes in physical relationships between nodes (sensing units) in real time and combines an attention mechanism to fuse global acoustic wave propagation characteristics, achieving in-depth mining of massive high-dimensional data. It can autonomously capture the damage development chain from local stress concentration to global modal evolution, greatly improving the intelligence level and diagnostic accuracy of identifying weak damage features from complex background noise and accurately distinguishing damage types (such as fatigue cracks and corrosion).

[0014] This invention maps the features output by a deep learning model back to the three-dimensional physical space of the structural components, visually displaying the probability distribution and precise location of damage in the form of heat maps, etc. At the same time, it constructs a multi-parameter quantitative model that integrates strain, vibration, sound waves and corrosion potential, and accurately quantitatively evaluates the surface crack length, internal damage initiation index and material corrosion rate, respectively. It provides maintenance personnel with clear, intuitive and quantitative decision-making basis, and strongly supports the transformation of offshore wind power structures from "periodic maintenance" to "predictive maintenance" intelligent operation and maintenance mode. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0016] Figure 2 This is a schematic diagram illustrating the differentiated configuration of the sensors in this invention;

[0017] Figure 3 This is a schematic diagram of the acoustic waveguide network embedding of the present invention;

[0018] Figure 4 This is a schematic diagram of the adaptive graph learning mechanism of the present invention;

[0019] Figure 5 This is a schematic diagram of the three-dimensional visualization evaluation of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] As attached Figure 1The method for quality monitoring of offshore wind turbine nacelle structural components includes S1: a flexible sensor array network is pre-installed on the surface of the nacelle structural components. The flexible sensor array network is composed of heterogeneous sensor units in a non-uniform topology. Each sensor unit establishes a dynamic data path through a self-organizing protocol. The sensor units include three types of sensors: strain, vibration, and corrosion potential. The arrangement density and type are differentiated according to the stress concentration coefficient and material degradation sensitivity of each region of the structural component.

[0022] It should be specifically noted that the differentiated configuration refers to the following: based on regions with high stress concentration and fast corrosion rate, regions with medium stress concentration and medium corrosion rate, and regions with low stress concentration and low corrosion rate, three types of sensors including strain, vibration, and corrosion potential are configured; two types of sensors including strain and vibration are configured; and a combination of sensing units including a single type of strain sensor is configured, with the arrangement density decreasing sequentially.

[0023] It should be further explained that, considering the complex load characteristics of the main engine compartment structural components, three types of customized sensing units were selected to construct a heterogeneous array: the strain sensing unit uses a polyimide-based flexible strain gauge doped with carbon nanotubes (model: CNT-PI-100), with a range of -2000με to +2000με, a resolution of 0.1με, and a temperature drift ≤0.05με / ℃; the vibration sensing unit uses a MEMS piezoelectric flexible vibration sensor (model: MEMS-PZT-08), with a frequency response range of 5Hz to 10kHz, a sensitivity of 100mV / g, and a mass of 0.2g; the corrosion potential sensing unit is a customized Ag / AgCl reference electrode integrated flexible corrosion sensor (model: Flex-EC-05), with a measurement range of -1.5V to +0.5V, an accuracy of 1mV, and a response time <100ms.

[0024] As attached Figure 2 As shown, the implementation of non-uniform topology and differentiated layout involves the following steps: The stress distribution and material degradation-sensitive areas of the structural components are obtained through finite element simulation. The specific implementation process and core parameters are as follows: A 1:1 three-dimensional solid model of the main engine compartment structural components is established using ABAQUS software. The model mesh uses eight-node linear reduced integral elements (C3D8R), with mesh refinement in critical areas (element size 2mm×2mm×2mm) and element size 5mm×5mm×5mm in non-critical areas, ensuring a mesh quality factor ≥0.8 (to avoid distorted meshes affecting calculation accuracy). The material properties of the structural components are set according to the actual parameters of Q690 steel: elastic modulus... Poisson's ratio Yield strength fracture toughness Material degradation parameters under corrosive environments were obtained through preliminary experiments. The elastic modulus attenuation coefficient of the corrosive region was set to have a linear relationship with corrosion time, i.e. , Corrosion time, in years; This is the initial elastic modulus.

[0025] The load and boundary conditions were applied according to the actual working scenario of the main engine room: the rated load included an axial tensile force of 500 kN (applied as a concentrated force to the center of the top flange of the structural component) and a torque of 10 kN·m (applied as a torque coupling constraint to the top flange). The load was applied in a step-by-step manner, with each step applying 10% of the rated load and holding the load for 10 seconds to avoid the influence of impact loads. The corrosive environment boundary conditions were determined using an electrochemical-mechanical coupled analysis module, with the corrosive medium set as a 5% salt spray solution, a temperature of 35℃, a relative humidity of 90%, and an electrode potential of -0.8V (simulating the marine atmospheric corrosion potential). The corrosion current density was calculated using Faraday's law. (T is temperature, unit: °C); The simulation analysis type adopts a coupled analysis of "static general analysis + corrosion diffusion analysis", and the analysis step is set to 100 incremental steps. Each incremental step corresponds to 100 hours of actual time. The total simulation time covers 1 / 10 of the design life of the structural component (approximately 10,000 hours).

[0026] The core calculation parameters and formulas are as follows: The stress concentration factor (SCF) is calculated using the maximum principal stress method, and the formula is: ,in The maximum principal stress in the region is expressed in MPa. Let be the nominal stress of a uniform cross section of a structural member under the same load. , For axial load, The cross-sectional area is represented by the SCF value. A larger SCF value indicates a higher degree of stress concentration in the region, making it more prone to fatigue crack initiation. The material corrosion rate CR is calculated based on the corrosion depth, using the formula: (Unit: mm / year), where The corrosion depth of the region at the end of the simulation (unit: m). To simulate corrosion time (unit: years), regions with a CR value ≥ 0.1 mm / year were identified as high corrosion rate areas. Following the simulation, SCF and CR contour maps of the entire structural component were output. Data extraction was performed using Origin software to obtain the SCF and CR values ​​for each grid cell, providing accurate data support for subsequent differentiated sensor array deployment.

[0027] Based on simulation results, a "three-level differentiated" layout strategy is adopted: Level 1 region (SCF≥3.0 and CR≥0.1mm / year): a unit combination of "3 strain + 2 vibration + 1 corrosion" is used, with a layout density of 2 units / cm². Level 1 region is a high-risk damage area, requiring multi-dimensional synchronous monitoring of strain concentration, modal abrupt changes, and corrosion initiation; Level 2 region (1.5≤SCF<3.0 and 0.05≤CR<0.1mm / year): a unit combination of "2 strain + 1 vibration" is used, with a layout density of 1 unit / cm². Level 2 region is mainly characterized by fatigue damage, with a weaker impact from corrosion. The corrosion sensor configuration is simplified to reduce costs; Level 3 region (SCF<1.5 and CR<0.05mm / year): a single strain sensor unit is deployed, with a layout density of 0.5 units / cm². Level 3 region has extremely low damage risk and only basic strain monitoring is required.

[0028] The self-organizing protocol of the array network adopts the improved LEACH protocol. The improved LEACH protocol is based on the traditional low-power adaptive clustering hierarchical protocol LEACH and is optimized for the heterogeneity of sensor arrays and monitoring scenarios. The core uses "energy consumption weight + area priority" to dynamically elect cluster heads (10-second rotation cycle). The priority of high-energy-consuming sensor units (such as vibration sensors) is reduced by 30%, and the cluster head density in the first-level monitoring area is increased to 1.5 times that of the second-level area. Within the cluster, a "heterogeneous data differential compression" strategy is adopted. Strain / vibration data is compressed using wavelet threshold (compression ratio 10:1), and corrosion potential data is compressed using differential encoding (compression ratio 5:1). Then, it is transmitted to the aggregation node through short-distance multi-hop routing to solve the problems of uneven energy consumption and transmission redundancy caused by the random cluster selection of traditional LEACH.

[0029] As attached Figure 3 As shown, S2: An acoustic waveguide network is embedded inside the structural component. The acoustic waveguide network is composed of a prefabricated waveguide array. The waveguide and the structural component body material are co-cured to form an array of acoustic wave conduction paths. The coded acoustic wave signal is generated through the excitation end, and the acoustic wave signal after propagation inside the structural component is collected at the receiving end.

[0030] It should be noted that the waveguide has a periodic microgroove structure on its inner wall, with a groove depth of 0.1 mm and a groove spacing of 0.5 mm.

[0031] The co-curing molding process includes the following steps: pre-setting waveguide positioning fixtures in the structural component mold, laying the structural component substrate and waveguide, applying epoxy resin adhesive, hot-pressing curing under set temperature and pressure conditions, and confirming the integrity of the waveguide and the body by ultrasonic testing after cooling.

[0032] It should be further noted that the waveguide is made of carbon fiber reinforced epoxy resin composite material (CF / EP), with a cross-sectional dimension of 2mm × 5mm (thickness × width). The length is designed in segments according to the structural component dimensions (maximum length 1.5m per segment). The inner wall is processed with periodic microgrooves (0.1mm depth, 0.5mm spacing) using laser etching to enhance the directional conduction of sound waves. Selection criteria: Acoustic impedance of CF / EP material (3.2 × 10⁻⁶). 6 kg / (m²∙s) and Q690 steel (4.5×10) commonly used in main engine compartment structural components 6 With a high matching degree of kg / (m²∙s) and a sound wave transmittance of ≥85%, it can solve the problem of large sound attenuation in traditional metal waveguides.

[0033] Co-curing molding process steps: Pre-set waveguide positioning fixtures in the structural component mold to ensure that the relative positional deviation between the waveguide and the structural component body is ≤±0.2mm; lay Q690 steel substrate and waveguide, apply epoxy resin adhesive (model: E-51), adhesive layer thickness 0.1mm; use autoclave for curing, curing temperature 120℃, pressure 0.8MPa, heat preservation time 2h; after cooling to room temperature, remove the fixtures, and verify the integrity of the connection between the waveguide and the body by ultrasonic testing (frequency 5MHz), the connection defect rate is ≤0.5%.

[0034] The excitation end of the acoustic excitation uses a piezoelectric ceramic exciter (model: PZT-40) with an output power of 10W and an excitation frequency range of 20kHz to 200kHz, generating a binary phase-shift keying (BPSK) encoded acoustic signal (64-bit code length, 1MHz code rate). The selection criteria are: the encoded acoustic signal has strong anti-interference characteristics, effectively distinguishing between structural background noise and damage scattering signals, solving the problem of traditional single-frequency acoustic waves being susceptible to interference. The receiving end of the acoustic excitation uses a fiber optic grating acoustic sensor (model: FBG-AC-20) with a sensitivity of 5pm / Pa and a sampling rate of 1MHz, arranged at the end of the waveguide (one receiver per waveguide). Data acquisition uses a synchronous acquisition card (model: NI-9234), synchronized with the acquisition clock of the flexible sensor array network via GPS, with a time synchronization error ≤ ±1μs.

[0035] S3: During the environmental load coupling stage, multi-physics response data are synchronously acquired through a flexible sensor array network, and acoustic wave signals propagating inside the structure are obtained through an acoustic waveguide network. The two types of signals are spatiotemporally aligned to extract strain distribution characteristics, vibration mode characteristics, and acoustic wave propagation characteristic parameters.

[0036] It should be specifically explained that the spatiotemporal alignment refers to establishing a three-dimensional coordinate system with the feature points of the structural components as the origin, acquiring the three-dimensional coordinates of each sensing unit and the endpoint of the waveguide through a high-precision coordinate acquisition device, forming a coordinate mapping table, and realizing the spatial matching of the positions of the sensing units, waveguides and structures.

[0037] The steps for achieving spatiotemporal alignment in the three-dimensional coordinate system include establishing a three-dimensional rectangular coordinate system with preset feature points on the main cabin structural components as the origin; associating and mapping each sensing unit in the flexible sensing array network and each waveguide in the acoustic waveguide network with specific spatial coordinates in the three-dimensional rectangular coordinate system; and using the timestamp provided by GPS timing as a reference, performing time synchronization processing on multi-physics response data and acoustic signals with spatial coordinate mapping relationships.

[0038] It should be further explained that during the environmental load coupling phase, the actual working loads of the simulated main engine compartment structural components are collected, and the coupling loads are applied through an electro-hydraulic servo loading system (model: MTS-810). Specifically, these include alternating stress loads with amplitudes ranging from 0 to 500 MPa, a frequency of 10 Hz, and a cycle count of 10. 6 The vibration load ranged from 0 to 5g, with a frequency of 50 to 200Hz, and was random. The corrosive environment consisted of a 5% salt spray concentration, a temperature of 35℃, and a relative humidity of 90%. The data acquisition duration was 1 / 10 of the fatigue life of the structural component (approximately 100 hours). The sampling frequencies were 1kHz for strain and vibration signals, 10Hz for corrosion potential signals, and 1MHz for acoustic signals.

[0039] Data storage is preprocessed locally on edge computing nodes before being uploaded to the cloud database. The storage capacity of the edge nodes is 1TB, and data transmission uses a 5G communication module (rate ≥1Gbps) to ensure data real-time performance.

[0040] Spatiotemporal alignment is implemented based on GPS time stamps to align the data of the flexible sensor array with the acoustic wave data. Spatial alignment is achieved by establishing a three-dimensional coordinate system for the structural components. Specifically, the center of the bottom flange of the structural component is selected as the origin O, the X-axis is along the axis of the main beam of the structural component to the top, the Y-axis is along the radial direction of the flange to the right, and the Z-axis is perpendicular to the flange plane upward, establishing a right-hand rectangular coordinate system (coordinate unit: mm). The three-dimensional coordinates of the center of each flexible sensor unit, both ends of the waveguide, and key structural feature points (such as the center of the welded joint) are collected by a laser tracker (measurement accuracy ±0.02mm). The coordinates are entered into the system to form a coordinate mapping table, realizing a one-to-one correspondence between "sensor unit-waveguide-structure position", with a coordinate matching error ≤ ±0.1mm.

[0041] The specific steps of feature extraction include strain distribution characteristics: Empirical Mode Decomposition (EMD) is applied to the time-domain signal of each strain sensing unit, decomposing it into 8 intrinsic mode functions (IMFs), and the first 3 IMF components are selected to calculate the strain mean. Standard deviation With peak factor, =Peak value / Effective value, forming a 3D strain characteristic vector. Selection criteria: During the damage initiation stage, the average strain decreases by 5% to 10%, while the peak factor increases by more than 20%. The first three IMF components contain more than 90% of the strain energy.

[0042] Vibration modal characteristics: Perform Fast Fourier Transform (FFT) on the vibration signal to extract the first 5 natural frequencies. With modal damping ratio Then, the instantaneous frequency standard deviation of the vibration signal is calculated using the Hilbert-Huang Transform (HHT). This forms an 11-dimensional vibration characteristic vector. Selection criteria: Structural crack propagation leads to a 3% to 5% decrease in natural frequency and a 15% to 20% increase in modal damping ratio; the instantaneous frequency standard deviation reflects vibration instability.

[0043] Sound wave propagation characteristics: Correlation decoding is performed on the received coded sound wave signal to extract the sound wave propagation time. Amplitude attenuation rate With phase offset Then, the energy entropy of 8 frequency bands is extracted by wavelet packet decomposition. This forms an 11-dimensional acoustic feature vector. The selection criteria are: internal structural damage leads to an increase in acoustic wave propagation time (0.2 μs per 1 mm crack), an increase in amplitude attenuation rate of 10% to 15%, and energy entropy reflects the complexity of acoustic wave scattering.

[0044] S4: Based on a deep spatiotemporal graph convolutional network, a structural state evolution model is constructed, which maps the data collected by the flexible sensor array network into a dynamic graph structure. Nodes are sensor units, and edges are the physical association strength between sensor units. At the same time, the acoustic wave propagation characteristic parameters are used as global attributes of the graph, and the association relationship between nodes and edges is dynamically updated through an adaptive graph learning mechanism.

[0045] As attached Figure 4 As shown, it should be specifically explained that the adaptive graph learning mechanism is implemented as follows: calculate the feature similarity matrix that reflects the correlation of node signals; calculate the global attention factor based on the acoustic phase offset and adjust the similarity matrix; normalize the adjusted similarity matrix to obtain the updated edge weights, and dynamically delete low-weight edges or add high-weight auxiliary edges according to a preset threshold.

[0046] It should be further explained that the dynamic graph structure mapping method uses the N sensing units of the flexible sensing array as nodes to construct a dynamic graph. The specific meaning is as follows: node set Each node For each sensing unit, the node set Nodes consisting of multiple corresponding sensing units The system consists of a node feature vector that fuses sensor data and spatial information. The element strain feature vector and vibration feature vector are concatenated (dimension 3 + 11 = 14), and the node spatial coordinates (x, y, z) are added to form a 17-dimensional initial feature.

[0047] Edge set E: When the spatial distance between two sensing units is ≤5cm, the edge connection is determined and established based on the stress transmission range of the structural components, and the edge weight is determined. The initial value is ， The spatial distance between the two nodes.

[0048] The adjacency matrix A(t) consists of edge weights and includes the amplitude attenuation rate in the sound wave propagation characteristics. It is incorporated into the adjacency matrix as a global attribute, that is... This enables the initial fusion of multi-source data.

[0049] The mapping tool uses the PyTorch Geometric framework. It associates sensor data with spatial coordinates through a custom dataset loading function. The core is to use a pre-defined unique identifier of the sensor unit (such as ID number) as a bridge to match and bind the pre-processed sensor feature data of the edge nodes with the corresponding three-dimensional spatial coordinates (x, y, z) of the sensor unit collected and recorded by the laser tracker. This forms a structured dataset that associates "feature data-spatial location", which can be directly called when modeling dynamic graphs to construct node features with spatial information and realize real-time updates of dynamic graphs (update cycle 1 second).

[0050] Design an adaptive graph learning module based on an attention mechanism to dynamically update the relationships between nodes and edges. Specific steps include: calculating the node feature similarity matrix S. Reflecting the signal correlation between two nodes; introducing a global attention factor. Due to the phase shift of the sound wave calculate, Adjust the similarity matrix to The updated edge weights are obtained by normalization using the Softmax function. ;when When <0.1, delete the edge; when When the value is greater than 0.8, the newly added auxiliary edge reflects the change in stress transmission path caused by damage.

[0051] S5: Based on the updated dynamic graph structure, local damage features and global degradation patterns of the structure are extracted through multi-layer spatiotemporal convolution. The internal propagation features obtained by the acoustic waveguide network are fused using the attention mechanism to achieve three-dimensional visualization, localization and quantitative assessment of internal damage initiation, surface crack propagation and material degradation degree of structural components.

[0052] It should be specifically explained that the three-dimensional visualization positioning and quantitative assessment includes mapping the features of graph convolution nodes to the three-dimensional coordinate system of the structural component, displaying the damage probability distribution through a heat map, and locating the damage location by combining the change in sound wave propagation time; constructing a crack length assessment model based on the changes in strain peak factor and vibration natural frequency; constructing an internal damage initiation index based on the sound wave amplitude attenuation rate and energy entropy; and calculating the material corrosion rate based on the changes in corrosion potential and mean strain.

[0053] As attached Figure 5 As shown, it is necessary to further explain that a dual-branch network structure of "graph convolution + temporal convolution" is constructed: the graph convolution layer adopts 3 layers of graph convolution. The first layer takes the dynamic graph G(t) as input and outputs 64-dimensional node features to capture local damage features; the second layer introduces residual connections and outputs 128-dimensional node features; the third layer obtains 64-dimensional global features through global pooling to reflect the overall degradation mode of the structure.

[0054] The temporal convolution branch uses two layers of 1D convolution (kernel size 5, stride 1) to process the sound wave propagation time series and output 64-dimensional temporal features; then, the time-dependent features are extracted through a bidirectional LSTM layer (64 hidden units) to obtain 64-dimensional temporal fusion features.

[0055] Attention fusion layer: Calculates the feature attention weights of the graph convolutional branch and the temporal convolutional branch (based on feature entropy values, the higher the entropy value, the greater the weight), and weights them to obtain the final 128-dimensional features.

[0056] The network was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The loss function was a combination of cross-entropy loss (for classification tasks) and mean squared error loss (for regression tasks). The training dataset included 100 sets of structural component data with different damage states (no damage, microcracks, macrocracks, and corrosion). The model accuracy was ≥95%.

[0057] 3D visualization localization includes damage localization: mapping the node features output by the graph convolution branch to the 3D coordinate system of the structural component, visualizing the strain distribution and damage probability through a heat map, with red areas indicating high-probability damage areas, and a probability ≥80% being considered damage, combined with the sound wave propagation time difference. , Given the propagation time without damage, calculate the location of the damage, with a positioning error ≤ ±2mm.

[0058] Quantitative evaluation indicators and methods include surface crack propagation quantification: based on strain peak factor. With the natural frequency of vibration Establish a regression model for crack length. , , The reference value is taken when there is no damage, and the measurement error is ≤ ±0.1mm;

[0059] Quantification of Internal Damage Initiation: Based on Acoustic Wave Amplitude Attenuation Rate With energy entropy Establish damage initiation index A value of D ≥ 1.2 indicates the presence of internal micro-damage (size ≥ 0.1 mm), with an accuracy rate ≥ 92%.

[0060] Quantification of material degradation: based on corrosion potential with mean strain Calculate the corrosion rate Quantization error ≤ ±0.5%.

[0061] The visualization output uses the Unity3D engine to build a 3D model of the structural components, and updates information on damage location, crack length and corrosion rate in real time. It supports user interaction and query, providing an intuitive basis for structural component maintenance decisions.

[0062] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments of this disclosure. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0063] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for quality monitoring of offshore wind turbine nacelle structural components, characterized in that, include: S1: A flexible sensor array network is pre-installed on the surface of the main cabin structural components. The flexible sensor array network is composed of heterogeneous sensor units in a non-uniform topology. Each sensor unit establishes a dynamic data path through a self-organizing protocol. The sensor units include three types of sensors: strain, vibration, and corrosion potential. The arrangement density and type are configured differently according to the stress concentration coefficient and material degradation sensitivity of each region of the structural components. S2: An acoustic waveguide network is embedded inside the structural component. The acoustic waveguide network is composed of a prefabricated waveguide array. The waveguide and the structural component body material are co-cured to form an array of acoustic wave conduction paths. The coded acoustic wave signal is generated through the excitation end, and the acoustic wave signal after propagation inside the structural component is collected at the receiving end. S3: During the environmental load coupling stage, multi-physics response data are synchronously collected through a flexible sensor array network, and acoustic wave signals propagating inside the structure are obtained through an acoustic waveguide network. The two types of signals are spatiotemporally aligned to extract strain distribution characteristics, vibration mode characteristics and acoustic wave propagation characteristic parameters. S4: Based on a deep spatiotemporal graph convolutional network, a structural state evolution model is constructed, which maps the data collected by the flexible sensor array network into a dynamic graph structure. Nodes are sensor units, and edges are the physical association strength between sensor units. At the same time, the acoustic wave propagation characteristic parameters are used as global attributes of the graph, and the association relationship between nodes and edges is dynamically updated through an adaptive graph learning mechanism. The adaptive graph learning mechanism is implemented by calculating a feature similarity matrix that reflects the correlation of node signals. The global attention factor is calculated based on the acoustic wave phase offset, and the similarity matrix is ​​adjusted. The adjusted similarity matrix is ​​normalized to obtain the updated edge weights. Low-weight edges are dynamically deleted or high-weight auxiliary edges are added according to a preset threshold. S5: Based on the updated dynamic graph structure, local damage features and global degradation patterns of the structure are extracted through multi-layer spatiotemporal convolution. The internal propagation features obtained by the acoustic waveguide network are fused using the attention mechanism to achieve three-dimensional visualization, localization and quantitative assessment of internal damage initiation, surface crack propagation and material degradation degree of structural components. The three-dimensional visualization positioning and quantitative assessment includes mapping graph convolution node features to the three-dimensional coordinate system of the structural component, displaying the damage probability distribution through a heat map, and locating the damage location by combining the change in sound wave propagation time; constructing a crack length assessment model based on the changes in strain peak factor and vibration natural frequency; constructing an internal damage initiation index based on the sound wave amplitude attenuation rate and energy entropy; and calculating the material corrosion rate based on the changes in corrosion potential and mean strain.

2. The method for quality monitoring of offshore wind turbine nacelle structural components according to claim 1, characterized in that: The differentiated configuration specifically refers to the following: based on regions with high stress concentration and fast corrosion rate, regions with medium stress concentration and medium corrosion rate, and regions with low stress concentration and low corrosion rate, three types of sensors including strain, vibration, and corrosion potential are configured; two types of sensors including strain and vibration are configured; and a combination of sensing units including a single type of strain sensor is configured, with the arrangement density decreasing sequentially.

3. The method for quality monitoring of offshore wind turbine nacelle structural components according to claim 1, characterized in that: The waveguide has a periodic microgroove structure on its inner wall, with a groove depth of 0.1 mm and a groove spacing of 0.5 mm.

4. The method for quality monitoring of offshore wind turbine nacelle structural components according to claim 1, characterized in that: The co-curing molding process includes the following steps: pre-setting waveguide positioning fixtures in the structural component mold, laying the structural component substrate and waveguide, applying epoxy resin adhesive, hot-pressing curing under set temperature and pressure conditions, and confirming the integrity of the waveguide and the body by ultrasonic testing after cooling.

5. The method for quality monitoring of offshore wind turbine nacelle structural components according to claim 1, characterized in that: The spatiotemporal alignment specifically involves establishing a three-dimensional coordinate system with the feature points of the structural components as the origin, acquiring the three-dimensional coordinates of each sensing unit and the endpoint of the waveguide through a high-precision coordinate acquisition device, forming a coordinate mapping table, and realizing spatial matching of the positions of the sensing units, waveguides, and structures.

6. A method for quality monitoring of offshore wind turbine nacelle structural components according to claim 5, characterized in that: The steps for achieving spatiotemporal alignment in the three-dimensional coordinate system include establishing a three-dimensional rectangular coordinate system with preset feature points on the main cabin structural components as the origin; associating and mapping each sensing unit in the flexible sensing array network and each waveguide in the acoustic waveguide network with specific spatial coordinates in the three-dimensional rectangular coordinate system; and using the timestamp provided by GPS timing as a reference, performing time synchronization processing on multi-physics response data and acoustic signals with spatial coordinate mapping relationships.

Images