A method for identifying damage in top tension risers based on multi-source excitation fusion and its model construction method.

Abstract

This invention provides a method for identifying damage in top-tensioned risers based on multi-source excitation fusion and its model construction method, belonging to the field of machine learning and application technology. By collecting vibration response data of risers under multi-source conditions such as lift excitation and wave force excitation, a standardized multi-source excitation vibration response dataset covering multiple damage locations and degrees is constructed, enabling end-to-end intelligent diagnosis of the riser structure's health status. Based on a deep learning architecture, a one-dimensional convolutional neural network and an adaptive channel attention mechanism are introduced to establish a high-dimensional nonlinear mapping relationship between multi-source vibration signals and structural damage states. By constructing an integrated model that includes adaptive signal denoising, multi-scale deep feature extraction, and multi-task collaborative output, accurate classification of damage locations and quantitative regression of damage degrees can be achieved simultaneously under strong noise interference, providing data-driven and logically closed-loop technical support for the healthy operation and maintenance of deep-sea top-tensioned riser structures.

Description

Technical Field

[0001] This invention belongs to the field of machine learning and application technology, and in particular relates to a method for identifying top tension riser damage based on multi-source excitation fusion and its model construction method. Background Technology

[0002] In the field of deep-sea oil and gas resource development and marine engineering, top tension risers, as key components connecting subsea wellheads and floating production platforms, are constantly exposed to complex marine dynamic environments, bearing alternating loads caused by waves, currents, and platform motion. Due to their extreme slenderness ratio and complex boundary conditions, top tension risers are highly susceptible to fatigue damage, crack propagation, and even structural failure during service, which can lead not only to huge economic losses but also potentially trigger ecological disasters. Therefore, real-time and accurate structural health monitoring and damage identification of top tension risers are core tasks for ensuring the safe operation of marine oil and gas production systems. The health status of such structures is mainly reflected in changes in dynamic response characteristics, specifically involving physical phenomena such as natural frequency drift, modal distortion, and damping ratio changes. However, these macroscopic dynamic indicators do not exist in isolation but are fundamentally subject to the nonlinear mapping of microscopic damage evolution processes such as local stiffness degradation, abnormal mass distribution, and boundary constraint relaxation within the structure, and are also greatly affected by random excitations from the marine environment.

[0003] However, in the current technology system for identifying damage to riser structures, methods based on vibration signal analysis still face significant challenges. Traditional identification methods often rely on manual extraction of fixed damage features, such as frequency or modal curvature. This forward trial-and-error feature engineering heavily depends on expert experience, and the features are sensitive to environmental noise, making it difficult to maintain robustness under varying operating conditions. With the introduction of deep learning technology, although convolutional neural networks have shown advantages in automatic feature extraction, existing research mostly directly adopts the two-dimensional convolutional neural network architecture from the field of computer vision. When processing one-dimensional vibration time-series signals of riser structures, this method usually requires first converting the one-dimensional signal into a two-dimensional image through time-frequency transformation. This preprocessing not only leads to the loss of phase information and weak high-frequency features but also significantly increases computational complexity and hardware storage costs, violating the original intention of efficient end-to-end monitoring.

[0004] Furthermore, existing monitoring methods typically rely on only a single type of environmental excitation response data, such as wave force response or vortex-induced vibration response alone. This single-source data model struggles to comprehensively characterize the complex dynamic behavior of risers under multi-physics coupling, leading to the masking of damage features in specific frequency bands or modes. Simultaneously, the real marine environment is filled with broadband background noise, and existing models lack embedded adaptive signal purification mechanisms, directly resulting in decreased damage location accuracy and large quantification deviations in damage severity under low signal-to-noise ratio conditions. In summary, existing technologies cannot yet meet the practical engineering requirements for top-tension riser damage identification in terms of multi-source information fusion, high-fidelity feature extraction, noise robustness, and simultaneous quantification of location severity. Therefore, a new end-to-end damage identification paradigm is urgently needed. Summary of the Invention

[0005] To address the above problems, the first aspect of this invention provides a method for constructing a top tension riser damage identification model based on multi-source excitation fusion, comprising the following steps:

[0006] Step 1: Deploy sensor arrays in key monitoring units and key monitoring nodes of the top tension riser to collect lift-excited vibration response data and wave-force-excited vibration response data under various damage conditions;

[0007] Step 2: After segmenting and standardizing the two types of data corresponding to the same damage condition, the data are spliced ​​together to form a fused standardized vibration signal subsequence; a dataset is constructed and divided into a training set and a test set;

[0008] Step 3: Construct a top-tension riser damage identification model, including an adaptive noise reduction preprocessing module for vibration signals, which reconstructs noisy, fused, and standardized vibration signal subsequences into clean vibration signals; a one-dimensional convolutional feature extraction network for riser vibration signals, which uses a first-layer wide-kernel convolutional layer in conjunction with a cascaded residual attention feature extraction module, and prevents gradient vanishing through parallel one-dimensional convolutional main paths and identity mapping bypasses, and embeds squeeze and excitation channel attention units to adaptively recalibrate channel weights, outputting one-dimensional depth vibration features; and a damage identification module based on multi-source feature adaptive fusion and dual-task mapping, which outputs the damage location and loss degree through parallel damage location identification branches and damage degree regression branches, respectively.

[0009] Step 4: Train the model using the training set and test set to obtain the final top tension riser damage identification model.

[0010] Preferably, the damage condition includes the damage location and the damage degree. The damage location is limited to a specific location in the key monitoring unit. The damage degree is defined as the cross-sectional loss rate, which is set between 5% and 20%.

[0011] Preferably, the signal segmentation employs a sliding window technique, using a preset fixed window length and sliding over a continuous time series with a configurable sliding step size to extract a series of fixed-length signal subsequences from the data; the standardization process employs the Z-score standardization method. First, the global mean and global standard deviation of all signal subsequences for each excitation and each vibration signal channel are calculated using the training set data. Then, the original signal subsequences of the sample data are subtracted from the global mean of the corresponding excitation and channel on the training set, and then divided by the global standard deviation of the corresponding excitation and channel on the training set to obtain the standardized signal subsequences.

[0012] Preferably, the vibration signal adaptive noise reduction preprocessing module employs a deep convolutional network structure with symmetrical expansion and contraction paths, mainly composed of four one-dimensional convolutional layers, activation functions, and skip connections; it receives noisy fused and normalized vibration signal subsequences. As input, firstly, a shallow feature map containing high-frequency noise texture and basic waveform structure is obtained by performing preliminary feature extraction on the original noisy signal through a first-layer one-dimensional convolutional layer and activation function. Secondly, this shallow feature map is input to a second-layer one-dimensional convolutional layer and activation function for high-dimensional mapping and encoding, generating deep abstract latent features that can distinguish between random noise and structurally deterministic responses. Thirdly, the deep features are fed into a third-layer one-dimensional convolutional layer and activation function, where the shallow features output from the first layer are fused with the current deep features in the channel dimension through skip connections. This process uses deep semantics to suppress noise while supplementing signal details with shallow information, resulting in a denoised reconstructed feature map. Finally, the reconstructed feature map undergoes dimensionality reduction and smoothing processing through a fourth-layer one-dimensional convolutional layer, mapping the multi-channel features back to the original signal dimension and outputting the denoised and reconstructed vibration signal. .

[0013] Preferably, the one-dimensional convolutional feature extraction network for the riser vibration signal specifically includes a first-layer wide-kernel convolutional layer, four cascaded residual attention feature extraction modules, a flattening layer, a dropout layer, and two fully connected layers connected in sequence. First, the first-layer wide-kernel convolutional layer receives a denoised and fused normalized vibration signal subsequence and outputs a shallow, coarse-grained feature map containing long-term temporal information, which is then input to the residual attention feature extraction modules. Second, each residual attention feature extraction module receives the output of the previous layer as input and contains two parallel paths. The main path consists of two cascaded one-dimensional convolutional layers, which perform deep abstraction of the input features to output deep semantic features. The bypass path is an identity mapping shortcut connection that directly transmits the input features to retain the original residual information. Each one-dimensional convolutional layer is immediately followed by a Leaky layer. The ReLU activation function is used. Next, squeeze and excitation channel attention units are embedded in the residual module. Based on deep semantic features, global average pooling and fully connected operations are used to adaptively learn and recalibrate the weights of each feature channel, generating a weighted recalibrated feature map. This map is then element-wise added to the bypass features to obtain the residual attention features output by the current module, enabling the network to automatically focus on key feature channels that are more sensitive to damage. Finally, the high-level feature map obtained after the cascaded processing of four modules is fed into a flattening layer, transforming it into a one-dimensional vector form of deep vibration features. .

[0014] Preferably, the damage recognition module based on multi-source feature adaptive fusion and dual-task mapping specifically comprises:

[0015] First, a channel-weighted fusion layer is introduced to adaptively weight and fuse one-dimensional deep vibration features using trainable parameter vectors to obtain multi-source fusion features. Second, the multi-source fusion features are fed into a fully connected layer for nonlinear transformation to output damage semantic features.

[0016] Based on the obtained damage semantic features, a dual-branch output structure is designed. Branch 1 is the damage location identification branch, which consists of a fully connected layer and a Softmax function. The number of output neurons in the fully connected layer is equal to the total number of key monitoring units in the top tension riser. This branch receives the damage semantic features and outputs a probability distribution vector, where each element represents the predicted probability of damage to the corresponding riser unit. The final damage location is determined by taking the unit number with the highest probability value. Branch 2 is the damage degree regression branch, which is a fully connected layer with one output neuron. It directly receives the damage semantic features and outputs a continuous scalar value, i.e., the damage degree predicted by the model.

[0017] Preferably, the training objective of the vibration signal adaptive noise reduction preprocessing module is to minimize the noise-reconstructed vibration signal. With the corresponding noise-free original signal The mean square error between the two signals is used to learn a mapping function from noisy signals to clean signals;

[0018] The damage recognition module based on multi-source feature adaptive fusion and dual-task mapping adopts a multi-task learning framework. The total loss function is a weighted sum of cross-entropy loss and mean square error loss, balancing the damage location classification and damage degree regression tasks.

[0019] The one-dimensional convolutional feature extraction network for riser vibration signals receives gradient information from the output of the damage recognition module based on multi-source feature adaptive fusion and dual-task mapping through a backpropagation algorithm. It then uses an optimizer to iteratively update the weight parameters of each convolutional layer and attention unit, extracting deep discriminative features that can maximize the differentiation of different damage states.

[0020] A second aspect of the present invention provides a method for identifying top tension riser damage based on multi-source excitation fusion, comprising the following steps:

[0021] S1, based on the training dataset, construct the same frequency and sampling duration to obtain real-time lift excitation vibration response data and wave force excitation vibration response data;

[0022] S2, after signal segmentation and standardization, is spliced ​​together to form a fused standardized vibration signal subsequence;

[0023] S3, input the fused standardized vibration signal subsequence into the top tension riser damage identification model constructed by the construction method described in the first aspect;

[0024] S4, the model outputs the location and extent of damage.

[0025] A third aspect of the present invention provides a top tension riser damage identification device based on multi-source excitation fusion. The device includes at least one processor and at least one memory, the processor and the memory being coupled together. The memory stores a computer execution program for a top tension riser damage identification model constructed by the construction method described in the first aspect. When the processor executes the computer execution program stored in the memory, it causes the processor to execute a top tension riser damage identification method based on multi-source excitation fusion.

[0026] A fourth aspect of the present invention also provides a computer-readable storage medium storing a computer-executable program for a top-tension riser damage identification model constructed by the construction method described in the first aspect, wherein when the computer-executable program is executed by a processor, the processor executes a top-tension riser damage identification method based on multi-source excitation fusion.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. A signal adaptive purification and multi-source data standardization system for noisy environments was constructed: Addressing the problems of high noise interference and lack of information from single excitation sources in marine environments, this invention proposes an adaptive noise reduction preprocessing mechanism for vibration signals based on a deep convolutional autoencoder. This mechanism employs a symmetrical expansion and contraction path structure, utilizing skip connections to fuse shallow high-frequency textures and deep semantic features. Before feature extraction, the noisy signal undergoes end-to-end reconstruction and purification, significantly improving the signal-to-noise ratio. Simultaneously, a standardized fusion process for multi-source excitation data is established. Through sliding window segmentation and Z-score standardization, the lift excitation response and wave force excitation response are concatenated along the channel dimension to construct a fused standardized vibration signal subsequence. This process eliminates the dimensional differences between different excitation source data, providing a high-quality input benchmark rich in multi-physics coupling features for subsequent models.

[0029] 2. A lossless feature extraction architecture based on a one-dimensional residual attention network was designed: To overcome the feature loss and computational redundancy problems existing in traditional two-dimensional convolutional networks when processing one-dimensional signals, this invention designs a one-dimensional custom convolutional neural network feature extractor specifically for riser monitoring. This architecture abandons the traditional downsampling pooling layer to prevent the loss of signal abrupt features caused by minor damage during dimensionality reduction; it innovatively uses a wide-kernel convolution in the first layer to capture long-term coarse-grained information, followed by cascading multiple residual attention feature extraction modules. Within each module, gradient vanishing is mitigated through parallel one-dimensional convolutional main paths and identity mapping bypasses, and squeezed and activated channel attention units are embedded to adaptively recalibrate feature channel weights. This design achieves a direct mapping from the original time series to high-order damage-sensitive features, ensuring that the model can automatically focus on key signal segments containing damage information, significantly improving the purity and efficiency of feature extraction;

[0030] 3. A multi-source feature adaptive weighted fusion and dual-task collaborative recognition model was constructed. To address the problem of simultaneously solving damage localization and quantitative assessment tasks, this invention proposes a damage recognition model based on multi-source feature adaptive fusion and dual-task mapping. This model first introduces a trainable channel-weighted fusion layer, which automatically learns the contribution of different excitation source features to the recognition task based on backpropagation gradients, achieving intelligent weighted fusion of multi-source information. Subsequently, a fully connected layer maps the fused features to a high-dimensional damage semantic space, and branches the data into two parallel branches: damage location recognition and damage severity regression. The location branch outputs the probability distribution vector of the damage unit, and the severity branch outputs the scalar prediction value of the cross-sectional loss rate. By constructing a weighted total loss function composed of cross-entropy loss and mean squared error loss, the model is driven to perform joint optimization within a multi-task learning framework, achieving a logical closed loop in both damage localization and damage quantification dimensions, thus improving the comprehensiveness and accuracy of the diagnostic results. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the technical route of the present invention.

[0032] Figure 2 A schematic diagram showing the deployment locations of twenty key monitoring units and twenty-one key monitoring nodes of the connecting units.

[0033] Figure 3 This is a diagram of the one-dimensional convolutional feature extraction network for riser vibration signals.

[0034] Figure 4 This is a structural diagram of a damage recognition module based on multi-source feature adaptive fusion and dual-task mapping.

[0035] Figure 5 Figure (1) shows the damage identification results under lift excitation in an embodiment of the present invention.

[0036] Figure 6 Figure (2) shows the damage identification results under lift excitation in an embodiment of the present invention.

[0037] Figure 7 Figure (3) shows the damage identification results under lift excitation in an embodiment of the present invention.

[0038] Figure 8 Figure (1) shows the damage identification results under wave force excitation in an embodiment of the present invention.

[0039] Figure 9 Figure (2) shows the damage identification results under wave force excitation in an embodiment of the present invention.

[0040] Figure 10 Figure (3) shows the damage identification results under wave force excitation in an embodiment of the present invention.

[0041] Figure 11 Figure (1) shows the damage identification results under the fusion of different excitation force response data in this embodiment of the invention.

[0042] Figure 12 Figure (2) shows the damage identification results under the fusion of different excitation force response data in this embodiment of the invention.

[0043] Figure 13 Figure (3) shows the damage identification results under the fusion of different excitation force response data in this embodiment of the invention.

[0044] Figure 14 Figure (1) shows the damage identification results at 70dB in this embodiment of the invention.

[0045] Figure 15 Figure (2) shows the damage identification results at 70dB in this embodiment of the invention.

[0046] Figure 16 Figure (3) shows the damage identification results at 70dB in this embodiment of the invention.

[0047] Figure 17 Figure (1) shows the damage identification results at 50dB in this embodiment of the invention.

[0048] Figure 18 Figure (2) shows the damage identification results at 50dB in this embodiment of the invention.

[0049] Figure 19 Figure (3) shows the damage identification results at 50dB in this embodiment of the invention.

[0050] Figure 20 Figure (1) shows the damage recognition result of the CNN-LSTM model in this embodiment of the invention.

[0051] Figure 21 Figure (2) shows the damage recognition results of the CNN-LSTM model in this embodiment of the invention.

[0052] Figure 22 Figure (3) shows the damage recognition results of the CNN-LSTM model in this embodiment of the invention. Detailed Implementation

[0053] The present invention proposes a top tension riser damage identification method based on multi-source excitation fusion, the overall technical route flowchart of which is shown below. Figure 1 As shown, the specific steps are as follows:

[0054] Step 1: Construct a multi-source excitation vibration response dataset for the top tension riser; based on the finite element modal analysis results, deploy an accelerometer array in the key monitoring units and key monitoring nodes of the top tension riser, and independently collect vibration response data under lift excitation and wave force excitation conditions using a high-fidelity numerical simulation platform; pre-set multiple damage conditions covering different damage locations and degrees, and accurately label the collected vibration response data with damage location and damage degree labels.

[0055] Step 2: Standardize and preprocess the multi-source excitation vibration response dataset; use the sliding window technique to segment the lift excitation vibration response data and wave force excitation vibration response data to extract fixed-length signal subsequences; use the Z-score standardization method to process the signal subsequences to eliminate differences in data dimensions and baseline offset; concatenate the standardized lift excitation subsequences and standardized wave force excitation subsequences corresponding to the same damage condition in the channel dimension to construct a fused standardized vibration signal subsequence, and divide it into training set and test set.

[0056] Step 3: Construct a top-tension riser damage identification model based on multi-source excitation fusion. The model integrates a vibration signal adaptive denoising preprocessing module, a riser vibration signal one-dimensional convolutional feature extraction network, and a damage identification module based on multi-source feature adaptive fusion and dual-task mapping. The vibration signal adaptive denoising preprocessing module adopts a symmetrical deep convolutional encoding and decoding structure, and uses skip connections to fuse shallow high-frequency texture and deep semantic features to reconstruct the noisy fused standardized vibration signal subsequence into a clean vibration signal. The riser vibration signal one-dimensional convolutional feature extraction network adopts a first-layer wide-kernel convolutional layer combined with a cascaded residual attention feature extraction module. It prevents gradient vanishing through parallel one-dimensional convolutional main paths and identity mapping bypasses, and embeds squeeze and excitation channel attention units to adaptively recalibrate channel weights, outputting one-dimensional deep vibration features. The damage identification module based on multi-source feature adaptive fusion and dual-task mapping introduces a trainable channel weighted fusion layer to adaptively weight the deep vibration features with multi-source information, and outputs the damage unit probability distribution vector and cross-sectional loss rate prediction value through parallel damage location identification branches and damage degree regression branches, respectively.

[0057] Step 4: Perform end-to-end joint training of the top tension riser damage identification model using the training set; optimize the model parameters using the Adam optimizer, and drive the model to simultaneously minimize the damage location classification error and damage degree regression error through the total loss function composed of the weighted sum of the cross-entropy loss function and the mean square error loss function; use the trained model to process the real-time collected top tension riser vibration signal, and directly output the damage location number and damage degree value through automatic noise reduction, feature extraction and multi-task mapping to complete the structural health monitoring task.

[0058] The invention will be further described below with reference to specific embodiments.

[0059] S1. Construction and Preprocessing of Multi-Source Excitation Vibration Response Dataset

[0060] First, based on the finite element modal analysis results, an accelerometer array was deployed at key monitoring units and nodes on the top-tension riser structure. Then, a high-fidelity numerical simulation platform was used to simulate lift-excitation and wave-force-excitation conditions, independently collecting vibration response data while ensuring the same sampling frequency and duration. Next, multiple damage conditions were pre-defined, including damage location and degree, and corresponding lift-excitation and wave-force-excitation vibration response data were collected under each condition. Then, the data samples were precisely labeled, including damage location and degree labels, and integrated to construct a multi-source excitation vibration response dataset. Finally, the lift-excitation and wave-force-excitation vibration response data were segmented and standardized respectively to obtain standardized lift-excitation and wave-force-excitation subsequences. Then, the standardized lift-excitation and wave-force-excitation subsequences under the same damage condition were concatenated along the channel dimension to form a fused standardized vibration signal subsequence.

[0061] Structural health monitoring of top-tension risers requires comprehensive capture of their dynamic response characteristics. Therefore, this embodiment selects twenty key monitoring units on the riser structure, sensitive to low- to high-order main vibration modes and capable of reflecting local stress concentration characteristics, along with twenty-one key monitoring nodes connecting these units, as sensor deployment locations based on the finite element modal analysis results. Figure 2 As shown. High-sensitivity accelerometer arrays are deployed at these selected key locations to dynamically acquire vibration response signals of the riser under various typical excitations. The specific construction process includes the following two steps:

[0062] S1-1 Multi-Source Excitation Vibration Response Data Acquisition: Two typical marine environmental excitation conditions with different physical meanings were simulated using a high-fidelity numerical simulation platform, and data were acquired independently. First, under the lift excitation condition, the vibration responses of all key units and nodes of the riser were collected, resulting in a lift excitation vibration response dataset. Second, under the wave force excitation condition, the vibration responses of all key units and nodes of the riser were collected, resulting in a wave force excitation vibration response dataset. To ensure strict alignment and comparability of the data in the time dimension, the data acquisition under both excitation conditions used the exact same sampling frequency (100 sample points per second) and the same sampling duration (100 seconds).

[0063] S1-2 Damage Condition Setting and Data Labeling: To simulate actual damage, damage conditions are pre-set. Damage conditions are constructed by systematically combining two dimensions: damage location and damage severity. The damage location is limited to a specific location among the twenty key monitoring units in this embodiment; the damage severity is defined as the cross-sectional loss rate, ranging from 5% to 20%. Under each set damage condition, the steps described in S1-1 are executed to simultaneously acquire the corresponding lift-excited vibration response data and wave-excited vibration response data. Subsequently, each acquired data sample is precisely labeled. The labeling includes two core pieces of information: the first is the damage location label, represented by a specific unit number; the second is the damage severity label, represented by the percentage of cross-sectional loss. Finally, the data from both excitations, the damage labels, and the corresponding condition information are integrated to construct a unified multi-source excitation vibration response dataset. A sample in this dataset consists of a pair of data from the same condition. Each data point is a three-dimensional tensor, its dimensions determined by the time step, the number of sampling points, and the number of channels. A single data point has twenty-one channels, corresponding to twenty-one key nodes.

[0064] S1-3 Vibration Signal Preprocessing and Standardization: For each independent set of lift-excited vibration response data and wave-excited vibration response data in the multi-source excitation vibration response dataset constructed in S1-2, the same standardization preprocessing procedure is performed. The first step is signal segmentation, using a sliding window technique with a preset fixed window length and a configurable sliding step size across the continuous time series, thereby extracting a series of fixed-length signal subsequences from each set of data. The second step is data standardization, using the Z-score standardization method. Specifically, using only the training set data, the global mean and global standard deviation of all signal subsequences for each excitation and each vibration signal channel are calculated. For any sample data to be processed, the standardization process is achieved by subtracting the global mean of the corresponding excitation and channel on the training set from the original signal subsequence, and then dividing by the global standard deviation of the corresponding excitation and channel on the training set, thus obtaining the standardized signal subsequence. This step eliminates dimensional differences and baseline shifts between data from different excitation sources and different channels. After this step, standardized lift excitation subsequences and standardized wave-excited subsequences are obtained respectively. Finally, when constructing each training or test sample, the standardized lift excitation subsequence and the standardized wave force excitation subsequence corresponding to the same damage condition are concatenated along the channel dimension to form a fused standardized vibration signal subsequence with forty-two channels. A custom data loading class is used to encapsulate this fused standardized vibration signal subsequence, the corresponding damage location label, and the damage degree label. The dataset is then divided according to a predefined ratio, constructing an efficient data pipeline that supports batch loading for model training.

[0065] S2. Design of an adaptive noise reduction preprocessing module for vibration signals based on convolutional neural networks.

[0066] To improve the reliability and robustness of the overall damage identification system under real-world high-noise marine conditions, a dedicated noise reduction preprocessing stage is added before the signal enters the core feature extraction network. This module aims to perform front-end purification of vibration signals containing environmental noise, thereby improving the accuracy and stability of subsequent feature extraction and damage identification.

[0067] S2-1 Noise Environment Simulation and Impact Assessment: To realistically simulate broadband noise interference in the marine environment, Gaussian white noise with a specific signal-to-noise ratio is actively injected into the vibration response signals of the test set constructed in S1, thereby generating noisy vibration signals for assessment. By comparing the difference in recognition performance of the model on the original signal and the noisy vibration signal, the impact of noise on the subsequent one-dimensional convolutional feature extraction network for riser vibration signals is quantitatively assessed.

[0068] S2-2 Adaptive Noise Reduction Preprocessing Module for Vibration Signals: An independent adaptive noise reduction preprocessing module for noisy vibration signals is designed. This module employs a deep convolutional network structure with symmetrical expansion and contraction paths, primarily consisting of four one-dimensional convolutional layers, activation functions, and skip connections. The network receives noisy, fused, and normalized vibration signal subsequences. As input, firstly, a shallow feature map containing high-frequency noise texture and basic waveform structure is obtained by performing preliminary feature extraction on the original noisy signal through a first-layer one-dimensional convolutional layer and activation function. Secondly, this shallow feature map is input to a second-layer one-dimensional convolutional layer and activation function for high-dimensional mapping and encoding, generating deep abstract latent features that can distinguish between random noise and structural deterministic responses. Thirdly, the deep features are fed into a third-layer one-dimensional convolutional layer and activation function, where the shallow features output from the first layer are fused with the current deep features in the channel dimension through skip connections. This process uses deep semantics to suppress noise while supplementing signal details with shallow information, resulting in a denoised reconstructed feature map. Finally, the reconstructed feature map undergoes dimensionality reduction and smoothing processing through a fourth-layer one-dimensional convolutional layer, mapping the multi-channel features back to the original signal dimension, thereby outputting the denoised and reconstructed vibration signal. The training objective of the vibration signal adaptive denoising preprocessing module is to minimize the denoised and reconstructed vibration signal. With the corresponding noise-free original signal The mean square error between the two signals is used to learn a mapping function from noisy signals to clean signals. After training, the solidified vibration signal adaptive denoising preprocessing module is deployed before the input of the subsequent riser vibration signal one-dimensional convolutional feature extraction network. When facing actual noise interference or conducting noise robustness tests, the acquired or simulated noisy vibration signal will first pass through this denoising module for processing. The output denoised and fused normalized vibration signal subsequence is then used as valid input to the subsequent riser vibration signal one-dimensional convolutional feature extraction network.

[0069] S3. Design of a multi-source signal deep feature extraction network based on a one-dimensional custom convolutional neural network.

[0070] Design a one-dimensional convolutional feature extraction network specifically for processing one-dimensional vibration signals of risers under top tension, as the core feature extractor, such as... Figure 3 As shown, this network has a customized structure that has been specifically optimized. It consists of a first wide-kernel convolutional layer, four cascaded residual attention feature extraction modules, a flattening layer, a dropout layer, and two fully connected layers. First, the first wide-kernel convolutional layer receives a denoised and fused normalized vibration signal subsequence, outputting a shallow, coarse-grained feature map containing long-term temporal information, which is then input to the residual attention feature extraction modules. Second, each residual attention feature extraction module receives the output of the previous layer as input. Internally, it contains two parallel paths. The main path consists of two cascaded one-dimensional convolutional layers, which perform deep abstraction of the input features to output deep semantic features. The bypass path is an identity mapping shortcut connection that directly transmits the input features to preserve the original residual information, effectively mitigating gradient vanishing and accelerating model convergence. Each one-dimensional convolutional layer is immediately followed by a Leaky layer. The ReLU activation function is used to introduce a nonlinear transformation. Next, squeeze and excitation channel attention units are embedded in the residual module. Based on deep semantic features, global average pooling and fully connected operations are used to adaptively learn and recalibrate the weights of each feature channel, generating a weighted recalibrated feature map. This map is then element-wise added to the bypass features to obtain the residual attention features output by the current module, enabling the network to automatically focus on key feature channels that are more sensitive to damage. Finally, the high-level feature map obtained after the cascaded processing of four modules is fed into a flattening layer, transforming it into a one-dimensional vector form of deep vibration features. The one-dimensional convolutional feature extraction network for riser vibration signals avoids the use of traditional downsampling pooling layers. Its design aims to prevent the loss of subtle signal features due to any possible downsampling operations. The wide-kernel convolutional and shallow residual modules at the front of the network are primarily responsible for capturing subtle amplitude abrupt changes or waveform distortions caused by local damage. The deep residual module at the rear of the network combines and abstracts these local features into high-level global features related to the damage location pattern and damage intensity. (Deep vibration features) This is the comprehensive feature representation of the riser vibration signal, which is the final output of the one-dimensional convolutional feature extraction network and contains rich damage information. It will be directly fed into the subsequent damage identification model based on multi-source feature adaptive fusion and dual-task mapping.

[0071] S4. Construction of a damage recognition module based on multi-source feature adaptive fusion and dual-task mapping

[0072] Module structure as follows Figure 4 As shown, a channel-weighted fusion layer is first introduced, which uses trainable parameter vectors to analyze the depth vibration features. Adaptive weighted fusion is performed to obtain multi-source fusion features. Then, the multi-source fusion features are fed into a fully connected layer for nonlinear transformation to output damage semantic features. Next, a dual-branch output structure is designed based on the damage semantic features. The damage location identification branch outputs a 20-dimensional probability distribution vector, and the damage degree regression branch outputs the predicted damage degree. Finally, a multi-task learning framework is adopted, and the total loss function is a weighted sum of cross-entropy loss and mean squared error loss to balance the damage location classification and damage degree regression tasks.

[0073] This module receives depth vibration characteristics from S3. Through an innovative fusion mechanism and dual-branch decoding architecture, it achieves simultaneous high-precision identification of damage location classification and damage degree regression.

[0074] S4-1 Multi-Source Feature Adaptive Fusion and Advanced Feature Mapping: Extracting Deep Vibration Features from One-Dimensional Convolutional Features of Riser Vibration Signals via Network Output The input is processed in this module. First, an innovative channel-weighted fusion layer is introduced. This layer contains a trainable parameter vector with the same dimension as the input feature channels, i.e., the channel fusion weight coefficient vector. This vector is learned end-to-end through the model's backpropagation process, and the value of each element quantifies the importance or contribution of the corresponding input channel's feature to completing the current damage identification task. The feature fusion operation is defined as a weighted summation along the feature channel dimension, resulting in a multi-source fusion feature after adaptive weighted fusion. This design enables intelligent and dynamic fusion of vibration signals from different excitation sources at the feature level, allowing the model to adaptively focus on the signal source that responds most significantly to a specific damage pattern. Subsequently, this multi-source fusion feature is fed into a fully connected layer for further nonlinear transformation and high-dimensional space mapping, outputting more refined damage semantic features.

[0075] S4-2 Damage Location and Severity Synchronous Output Module: Based on the obtained damage semantic features, a dual-branch output structure is designed to solve the damage localization and quantitative assessment problems in parallel. Branch 1 is the damage location identification branch, which consists of a fully connected layer and a Softmax function. The number of output neurons in its fully connected layer is equal to the total number of key monitoring units in the top tension riser, which is twenty. This branch receives damage semantic features and outputs a twenty-dimensional probability distribution vector, where each element represents the predicted probability of damage in the corresponding riser unit. The final damage location is determined by taking the unit number with the highest probability value. Branch 2 is the damage severity regression branch, which is also a fully connected layer, but its output neuron count is one. It directly receives damage semantic features and outputs a continuous scalar value, i.e., the damage severity predicted by the model. The entire damage recognition model based on multi-source feature adaptive fusion and dual-task mapping is jointly optimized using a multi-task learning framework. The total loss function is obtained by weighted summation of the losses of the two tasks, specifically the sum of the cross-entropy loss of location classification and the mean squared error loss of degree regression. Two adjustable hyperparameters are introduced to balance the relative importance of the two recognition tasks in the overall training objective of the model.

[0076] S5. Execution of damage identification task based on end-to-end deep learning model

[0077] First, the one-dimensional convolutional feature extraction network for riser vibration signals constructed in S2 to S4, the vibration signal adaptive denoising preprocessing module based on convolutional neural networks, and the damage recognition module based on multi-source feature adaptive fusion and dual-task mapping are integrated into a complete end-to-end deep learning model. Second, the Adam optimizer is used to train the model, and an early stopping mechanism is introduced to optimize performance. After training, the core task of this deep learning model is to perform damage recognition in various real-world scenarios, i.e., simultaneously outputting the damage location and degree. Third, to comprehensively evaluate the performance and robustness of this deep learning model in performing the damage recognition task, multi-scenario verification experiments were designed. Verification scenarios include: recognition performance under a single type of excitation (lift only or wave force only); recognition performance under multi-source excitation fusion to verify the effectiveness of the data fusion strategy; recognition performance under different intensities of Gaussian white noise interference, and comparative analysis of the enhancement effect of the denoising preprocessing module; and comparison with the classic CNN-LSTM hybrid model in this field to demonstrate the advantages of the proposed network structure in capturing local damage features of risers. Finally, the trained and validated deep learning model was encapsulated and integrated into the structural health monitoring system software platform for top-tensioned risers. In practical applications, the system receives real-time signals from an array of accelerometers at twenty-one key monitoring nodes, automatically performs signal standardization, fusion, and noise reduction preprocessing, and then drives the model to perform forward inference, generating a structured diagnostic report containing the specific damage unit number (one of the twenty key monitoring units) and the percentage of cross-sectional loss, thus completing the automated damage identification task from data to diagnosis.

[0078] S6. Experimental Verification and Analysis

[0079] In the damage recognition model based on multi-source feature adaptive fusion and dual-task mapping proposed in this invention, the Leaky ReLU function is used as the activation function in the convolutional and fully connected layers of the one-dimensional convolutional feature extraction network for riser vibration signals to enhance nonlinear expressive power while avoiding the gradient vanishing problem. Four loss functions—MSE, MAE, RMSE, and R2—are selected as effective metrics to measure the difference between the model's predicted values ​​and the true values. All of these are suitable for regression tasks, and this paper uses R2 as the main reference metric; a value closer to 1 indicates better convergence. For the optimizer, the Adam optimizer is used, which can dynamically adjust the learning rate during training to improve the model's convergence speed and stability. A learning rate of 0.001 is set, a moderate value that ensures model training efficiency while avoiding oscillations caused by an excessively large learning rate or slow convergence caused by an excessively small learning rate. The training iterations are set to 100 times, allowing the model to fully learn the data features through multiple iterations. Simultaneously, an early stopping mechanism (based on the minimum loss on the validation set) is used to prevent overfitting, and the model performance is evaluated using a test set after training.

[0080] Damage recognition performance and fusion effects under different environmental stimuli:

[0081] Nine different damage scenarios were randomly set (different damage locations and damage degrees; for ease of comparison, all scenarios below are set to this damage mode). Figure 5 , Figure 6 , Figure 7 The figure shows the predicted damage location and severity of riser under nine different damage conditions under lift excitation. From this, we can conclude that:

[0082] (1) For all damaged units, the accuracy of the predicted damage location is 100%. This means that under lift excitation, the model can accurately identify the damage location of different damaged units, and the predicted damage location is consistent with the actual damage location;

[0083] (2) Regarding the identification of damage severity, the overall average identification accuracy for the nine different damage conditions was 91.91%. This result indicates that the model has certain errors in predicting damage severity, and the prediction accuracy for some damage severity conditions has a large deviation, such as damage unit 12 (actual accuracy 7%, prediction accuracy 4.47%, accuracy 63.86%) and damage unit 16 (actual accuracy 7%, prediction accuracy 7.89%, accuracy 87.29%).

[0084] Using this model to process the riser vibration response under wave force excitation, the predicted results for the location and extent of riser damage under nine different damage conditions are as follows: Figure 8 , Figure 9 , Figure 10 As shown in the figure, we can conclude that:

[0085] (1) The model has a 100% accuracy in identifying all damage locations, which means that the model can accurately identify the different damage unit locations of the riser under wave force excitation, and the predicted damage locations are consistent with the actual damage locations.

[0086] (2) Regarding the identification of damage severity, the overall average identification accuracy for the nine different damage conditions was 88.93%. This result indicates that the model has certain errors in predicting damage severity, and the prediction accuracy of some damage severity conditions has large deviations, such as damage unit 9 (actual accuracy 7%, predicted accuracy 9.46%, accuracy 64.86%) and damage unit 16 (actual accuracy 7%, predicted accuracy 5.76%, accuracy 82.29%).

[0087] This model was used to process the riser vibration response under combined lift and wave force excitation. The predicted damage locations and degrees of riser damage under nine different damage scenarios are as follows: Figure 11 , Figure 12 , Figure 13 As shown in the figure, we can conclude that:

[0088] (1) The accuracy of the damage location prediction remained 100% in all cases, showing that the model still maintains a high degree of accuracy in damage localization;

[0089] (2) The fusion of lift and wave force excitation response data significantly improved the accuracy of damage degree identification, with an average overall damage degree identification accuracy of 93.63%, higher than 91.91% under lift excitation and 88.93% under wave force excitation. This indicates that the fusion of riser response data effectively improved the model's ability to predict damage degree;

[0090] (3) Although the accuracy of identifying some damage levels still has some errors (such as damage units 12 and 9), overall, the data fusion method has improved the stability of the model in identifying damage to a certain extent. This research result provides strong theoretical support for further optimization of the model, especially for model application under complex working conditions.

[0091] Damage identification performance and noise reduction method selection under noisy conditions:

[0092] Figure 14 , Figure 15 , Figure 16 The figure shows the predicted damage location and damage extent of risers under nine different damage conditions when 70dB noise is applied. The following conclusions can be drawn:

[0093] (1) Under noise interference, the accuracy of damage location prediction decreased, indicating that noise has a certain impact on the identification of damage location. Although the model can accurately locate most damages, some damage locations still cannot be located well.

[0094] (2) The overall accuracy of damage identification decreased, averaging 72.85%. There were some large prediction errors for the degree of damage, especially for damaged units 3, 10 and 15, with damage identification accuracies of 68.05%, 65.11% and 66.57% respectively, which showed that noise had a significant impact on the model's prediction performance;

[0095] (3) The accuracy of identifying the damage location and damage degree of the damage unit 20 is 0%, which may be due to excessive noise causing the model to be unable to identify the damage at all.

[0096] Figure 17 , Figure 18 , Figure 19 The figure shows the predicted damage location and damage extent of risers under nine different damage conditions when 50dB noise is applied. The following conclusions can be drawn:

[0097] (1) Under 50dB noise interference, the overall accuracy of damage identification decreased, with an average of 63.72%. The damage identification error of some damage units was relatively large, such as damage unit 1 (identification accuracy of 65%) and damage unit 10 (identification accuracy of 63.68%), which indicates that the high noise interference had a significant impact on the model's damage identification.

[0098] (2) The accuracy of the identification of damage location and damage degree of damage unit 16 and damage unit 20 is 0%, which also indicates that the model’s ability to predict the degree of damage is greatly challenged in a high-noise environment. In practical applications, noise puts forward higher requirements for the effectiveness of damage identification model.

[0099] Table 1 shows a comparison of the loss function values ​​under different noise conditions. The changes in R² values ​​indicate that increased noise significantly reduces the model's fit, especially under high noise levels of 50 dB, where the model's predictive ability drops significantly. The model performs best under noise-free conditions.

[0100] Table 1 Loss function values ​​under different noise conditions

[0101]

[0102] To further reduce the impact of noise on the model's damage recognition performance, this invention uses a vibration signal adaptive noise reduction preprocessing module based on convolutional neural networks for data noise reduction.

[0103] Recognition performance of CNN and CNN-LSTM models:

[0104] CNN-LSTM (Convolutional Neural Network - Long Short-Term Memory) is a deep learning model that combines the advantages of Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs). It aims to simultaneously handle the spatial features and temporal dependencies of data, and is widely used to process complex tasks with spatiotemporal characteristics, such as video analysis, speech recognition, and time series prediction. The core idea of ​​this model is to extract the spatial features of the input data using CNNs, and then capture the temporal dependencies within these features using LSTMs, thereby enhancing the modeling ability of spatiotemporal data.

[0105] Depend on Figure 20 , Figure 21 , Figure 22 We can conclude that:

[0106] (1) The accuracy of CNN-LSTM in damage location identification remains at 100%, indicating that its performance in locating damage locations is still quite good. However, a misclassification occurred when locating damage unit 12, which may be because the features of damage unit 12 are different from other damage units, making it difficult for the model to correctly distinguish the damage.

[0107] (2) Regarding the accuracy of damage identification, the CNN-LSTM model has an average damage identification accuracy of 84.17%, which is lower than the 93.63% of the damage identification model based on multi-source feature adaptive fusion and dual-task mapping. In particular, for damage unit 12, its damage identification accuracy is only 32.29%.

[0108] (3) For the vibration response of the riser in this paper, processing local damage features may be more important than capturing the time-series dependencies in the signal features. Therefore, a damage identification model based on multi-source feature adaptive fusion and dual-task mapping may be more suitable for the needs of this task.

[0109] This invention also provides a top-tension riser damage identification device based on multi-source excitation fusion. The device includes at least one processor and at least one memory, as well as a communication interface and an internal bus. The memory stores a computer-executable program. The memory also stores a computer-executable program for a top-tension riser damage identification model constructed using the method described above. When the processor executes the computer-executable program stored in the memory, it can execute a top-tension riser damage identification method based on multi-source excitation fusion. The internal bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, the bus in this application is not limited to only one bus or one type of bus. The memory may include high-speed RAM, and may also include non-volatile memory (NVM), such as at least one disk storage device, or a USB flash drive, portable hard drive, read-only memory, disk, or optical disk, etc.

[0110] The device may be provided as a terminal, server, or other type of device. In an exemplary embodiment, the electronic device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0111] The present invention also provides a computer-readable storage medium storing a computer-executable program of a top-tension riser damage identification model constructed by the construction method described above. When the computer-executable program is executed by a processor, the processor can execute a top-tension riser damage identification method based on multi-source excitation fusion.

[0112] Specifically, a system, apparatus, or device may be provided equipped with a readable storage medium on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system, apparatus, or device reads and executes the instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any of the embodiments described above, therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0113] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0114] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for constructing a top-tension riser damage identification model based on multi-source excitation fusion, characterized in that, Includes the following steps: Step 1: Deploy sensor arrays in key monitoring units and key monitoring nodes of the top tension riser to collect lift-excited vibration response data and wave-force-excited vibration response data under various damage conditions; Step 2: After segmenting and standardizing the two types of data corresponding to the same damage condition, the data are spliced ​​together to form a fused standardized vibration signal subsequence; a dataset is constructed and divided into a training set and a test set; Step 3: Construct a top-tension riser damage identification model, including an adaptive noise reduction preprocessing module for vibration signals, which reconstructs the noisy fused and standardized vibration signal subsequence into a clean vibration signal; and a one-dimensional convolutional feature extraction network for riser vibration signals, which uses a first-layer wide-kernel convolutional layer in conjunction with a cascaded residual attention feature extraction module, and prevents gradient vanishing through parallel one-dimensional convolutional main paths and identity mapping bypasses, and embeds squeeze and excitation channel attention units to adaptively recalibrate channel weights, outputting one-dimensional depth vibration features; It includes a damage recognition module based on multi-source feature adaptive fusion and dual-task mapping, which outputs the damage location and loss degree through parallel damage location recognition branch and damage degree regression branch, respectively. The vibration signal adaptive noise reduction preprocessing module employs a deep convolutional network structure with symmetrical expansion and contraction paths, mainly consisting of four one-dimensional convolutional layers, activation functions, and skip connections; it receives noisy, fused, and normalized vibration signal subsequences. As input, firstly, the original noisy signal is subjected to preliminary feature extraction through the first one-dimensional convolutional layer and activation function to obtain a shallow feature map containing high-frequency noise texture and basic waveform structure; secondly, the shallow feature map is input to the second one-dimensional convolutional layer and activation function for high-dimensional mapping and encoding to generate deep abstract latent features that can distinguish between random noise and structural deterministic response. again, The deep features are fed into the third one-dimensional convolutional layer and activation function. Here, skip connections are used to fuse the shallow features output from the first layer with the current deep features in the channel dimension. This process uses deep semantics to suppress noise while supplementing signal details with shallow information, resulting in a denoised reconstructed feature map. Finally, the reconstructed feature map is passed through the fourth one-dimensional convolutional layer for dimensionality reduction and smoothing, mapping the multi-channel features back to the original signal dimension and outputting the denoised and reconstructed vibration signal. ; The one-dimensional convolutional feature extraction network for the riser vibration signal specifically includes a first-layer wide-kernel convolutional layer, four cascaded residual attention feature extraction modules, a flattening layer, a dropout layer, and two fully connected layers connected in sequence. First, the first-layer wide-kernel convolutional layer receives the denoised and fused normalized vibration signal subsequence and outputs a shallow coarse-grained feature map containing long temporal information, which is then input into the residual attention feature extraction module. Secondly, each residual attention feature extraction module receives the output of the previous layer as input. Internally, it contains two parallel paths. The main path consists of two cascaded one-dimensional convolutional layers, which perform deep abstraction on the input features to output deep semantic features. The bypass path is an identity mapping shortcut connection that directly transmits the input features to preserve the original residual information. A Leaky ReLU activation function is immediately connected after each one-dimensional convolutional layer. Thirdly, squeeze and excitation channel attention units are embedded in the residual module. Based on deep semantic features, global average pooling and fully connected operations are used to adaptively learn and recalibrate the weights of each feature channel, generating a weighted recalibrated feature map. This map is then element-wise added to the bypass features to obtain the residual attention feature output by the current module, enabling the network to automatically focus on key feature channels that are more sensitive to damage. Finally, the high-level feature map obtained after the cascaded processing of the four modules is fed into a flattening layer, transforming it into a one-dimensional vector form of deep vibration features. ; The damage recognition module based on multi-source feature adaptive fusion and dual-task mapping is specifically as follows: First, a channel-weighted fusion layer is introduced to adaptively weight and fuse one-dimensional deep vibration features using trainable parameter vectors to obtain multi-source fusion features. Second, the multi-source fusion features are fed into a fully connected layer for nonlinear transformation to output damage semantic features. Based on the obtained damage semantic features, a dual-branch output structure is designed. Branch 1 is the damage location identification branch, which consists of a fully connected layer and a Softmax function. The number of output neurons in the fully connected layer is equal to the total number of key monitoring units in the top tension riser. This branch receives the damage semantic features and outputs a probability distribution vector, where each element represents the predicted probability of damage to the corresponding riser unit. The final damage location is determined by taking the unit number with the highest probability value. Branch 2 is the damage severity regression branch, which is a fully connected layer with one output neuron. It directly receives the damage semantic features and outputs a continuous scalar value, i.e., the damage severity predicted by the model. Step 4: Train the model using the training set and test set to obtain the final top tension riser damage identification model.

2. The method for constructing a top tension riser damage identification model based on multi-source excitation fusion as described in claim 1, characterized in that: The damage condition includes the location of the damage and the degree of damage. The location of the damage is limited to a specific location in the key monitoring unit. The degree of damage is defined as the cross-sectional loss rate, which is set between 5% and 20%.

3. The method for constructing a top tension riser damage identification model based on multi-source excitation fusion as described in claim 1, characterized in that: The signal segmentation employs a sliding window technique, using a preset fixed window length and a configurable sliding step size to slide across the continuous time series, extracting a series of fixed-length signal subsequences from the data. The standardization process uses the Z-score standardization method. First, the global mean and global standard deviation of all signal subsequences for each excitation and each vibration signal channel are calculated using the training set data. Then, the original signal subsequences of the sample data are subtracted from the global mean of the corresponding excitation and channel on the training set, and then divided by the global standard deviation of the corresponding excitation and channel on the training set to obtain the standardized signal subsequences.

4. The method for constructing a top tension riser damage identification model based on multi-source excitation fusion as described in claim 1, characterized in that: The training objective of the vibration signal adaptive noise reduction preprocessing module is to minimize the noise-reconstructed vibration signal. With the corresponding noise-free original signal The mean square error between the two signals is used to learn a mapping function from noisy signals to clean signals; The damage recognition module based on multi-source feature adaptive fusion and dual-task mapping adopts a multi-task learning framework. The total loss function is a weighted sum of cross-entropy loss and mean square error loss, balancing the damage location classification and damage degree regression tasks. The one-dimensional convolutional feature extraction network for riser vibration signals receives gradient information from the output of the damage recognition module based on multi-source feature adaptive fusion and dual-task mapping through a backpropagation algorithm. It then uses an optimizer to iteratively update the weight parameters of each convolutional layer and attention unit, extracting deep discriminative features that can maximize the differentiation of different damage states.

5. A method for identifying damage to top-tension risers based on multi-source excitation fusion, characterized in that, The process includes the following: S1, based on the training dataset, construct the same frequency and sampling duration to obtain real-time lift excitation vibration response data and wave force excitation vibration response data; S2, after signal segmentation and standardization, is spliced ​​together to form a fused standardized vibration signal subsequence; S3, input the fused standardized vibration signal subsequence into the top tension riser damage identification model constructed by the construction method as described in any one of claims 1 to 4; S4, the model outputs the location and extent of damage.

6. A top tension riser damage identification device based on multi-source excitation fusion, characterized in that: The device includes at least one processor and at least one memory, the processor and the memory being coupled together; the memory stores a computer execution program for a top tension riser damage identification model constructed by the construction method as described in any one of claims 1 to 4; when the processor executes the computer execution program stored in the memory, the processor executes a top tension riser damage identification method based on multi-source excitation fusion.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer executable program for a top tension riser damage identification model constructed by the construction method as described in any one of claims 1 to 4. When the computer executable program is executed by a processor, the processor executes a top tension riser damage identification method based on multi-source excitation fusion.

Images