Ocean pipeline damage positioning method of dendritic network optimized by improved wild horse algorithm

Abstract

The invention relates to the technical field of pipeline damage detection, in particular to a marine pipeline damage positioning method of a dendritic network optimized by an improved wild horse algorithm. The method comprises the following steps: establishment of a marine pipeline damage model, acquisition of a strain data set, establishment of an improved wild horse algorithm optimized dendritic network, establishment of a pipeline damage identification experiment platform, and training and testing of an experiment data set. According to the invention, a large amount of signal data can be processed and analyzed, and the damage position can be rapidly positioned, so that the real-time monitoring and early warning of the pipeline can be realized, and potential safety risks can be timely identified and processed; a provided weight feature fusion module not only considers contribution degrees of different features, but also can retain key information in a fusion process, so that the representation capability and prediction precision of the model are improved; the state of the pipeline is analyzed from multiple angles and multiple dimensions, so that the damage position is identified more accurately, the maintenance cost is reduced, and the safety of the pipeline is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline damage detection, and in particular to a method for locating marine pipeline damage using a dendritic network optimized by an improved Mustang algorithm. Background Art

[0002] Offshore pipeline transportation is a continuous and efficient transportation method that allows for long-term stable operation and high transport efficiency. Offshore pipeline transportation avoids direct interference and damage to the transported medium caused by external factors. Furthermore, as global demand for oil and gas resources continues to rise and new discoveries of offshore oil and gas resources continue to increase, the ocean will undoubtedly become the primary source of oil and gas resources in the future. Therefore, offshore pipelines, as key components for offshore oil and gas transportation, are bound to be more widely used. However, during the manufacturing, assembly, and operation of offshore pipelines, due to material properties and external conditions, they are inevitably subject to defects and damage. These defects can weaken the pipeline's load-bearing capacity and reliability, reduce its structural strength, shorten its service life, and even pose significant safety hazards. To ensure the safe operation and extend the service life of offshore pipelines, accurately identifying the damage location is crucial. For example, the ultrasonic pipeline defect identification and location method based on an improved CapsNet network, published in Chinese Patent Publication No. CN114694051A, also has limitations. Ultrasonic detection technology has drawbacks such as high operator skill requirements and limited ultrasonic penetration. Acoustic emission testing typically produces weak signals that are easily affected by ambient noise. Furthermore, the interpretation and analysis of AE test results require specialized skills and experience, leading to a degree of subjectivity. Magnetic flux leakage testing is susceptible to the influence of material magnetism and environmental magnetic fields, such as electromagnetic fields and geomagnetic fields. These fields can interfere with the test signal, leading to misjudgments or missed detections. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a method for locating marine pipeline damage using a dendritic network optimized by an improved Mustang algorithm.

[0004] The technical solution adopted in the present invention is as follows: A marine pipeline damage location method based on a dendritic network optimized by an improved Mustang algorithm comprises the following steps: S1. Establishment of marine pipeline damage model: simulate a marine pipeline with crack damage, and dig a long and narrow groove on the surface of the marine pipeline to simulate the crack. The distance from the crack to the fixed support end is X, and the crack size is L. D H, where L is the damage length, D is the damage width, and H is the damage depth; S2. Acquisition of strain dataset: Construct a pipeline model with different damage locations, add strain sensors and acceleration sensors to the surface of the pipeline model, perform strain and acceleration analysis, and obtain strain response data and acceleration response data to obtain a strain dataset. S3. Establishment of a dendritic network optimized by the improved Mustang algorithm: Strain response data and acceleration response data are used as multi-source information to identify pipeline damage. A dendritic network optimized by the improved Mustang algorithm is constructed. The dendritic network includes a feature extraction module, a weighted feature fusion module, a multi-head attention module, a Mustang algorithm hyperparameter optimization module, and a decision module for damage location and identification. S4. Establishment of a pipeline damage identification experimental platform: The pipeline damage identification experimental platform includes a signal collector, a power amplifier, an exciter, a pipeline, a strain sensor, an acceleration sensor, a grating optical demodulator, a dynamic signal test and analysis system, and a computer. The signal collector generates an excitation signal after receiving instructions, and the power amplifier amplifies the excitation signal to control the vibration amplitude of the exciter. The exciter and pipeline are fixed to the test bench. The strain response data and acceleration data generated by pipeline vibration are measured by the strain sensor and the acceleration sensor respectively. The data are transmitted to the fiber Bragg grating demodulator via the strain sensor. The strain change data is demodulated by measuring the peak wavelength, valley wavelength, and full spectrum of the fiber Bragg grating. The demodulated signal is then transmitted to the computer. The strain and acceleration signal data are obtained through the pipeline damage identification experimental platform to obtain the experimental data set. S5. Training and testing of experimental data sets: The dendritic network optimized by the improved Wild Horse algorithm is used to perform feature processing on the experimental data sets and obtain the test accuracy.

[0005] This technical solution performs a three-dimensional simulation of an ocean pipeline and performs detailed processing on cracks at different locations on the pipeline to obtain strain datasets for different damage locations. A dendritic network optimized with an improved Mustang algorithm can perform feature extraction, weighted feature fusion, multi-head attention, and Mustang algorithm hyperparameter optimization and decision-making functions, thereby achieving damage location and identification. In a laboratory environment, a pipeline damage identification experimental platform was constructed using a test bench to realistically simulate signal data for pipeline damage identification and obtain an experimental dataset. The experimental dataset was then feature-processed using a dendritic network optimized with an improved Mustang algorithm to determine the test accuracy.

[0006] In addition, the marine pipeline damage location method using the dendritic network optimized by the improved Mustang algorithm proposed in the present invention may also have the following additional technical features: According to one embodiment of the present invention, in the establishment of the marine pipeline damage model in step S1, SOLIDWORKS is used to simulate a three-dimensional model of a marine pipeline with crack damage, and the crack size is 16 mm. 2mm 0.5mm, the distances from the cracks of the 13 pipes to the fixed support ends are 200mm, 203mm, 206mm, 209mm, 212mm, 215mm, 218mm, 221mm, 224mm, 227mm, 230mm and 233mm respectively.

[0007] In this technical solution, the three-dimensional model of the marine pipeline is simulated by cracks with the same spacing, so as to refine the data such as damage length, damage width, and damage depth, making the damage location more accurate.

[0008] According to one embodiment of the present invention, the acquisition of the strain data set in step S2 includes the following steps: S21. Using finite element ANSYS software, a finite element model is established, and the pipeline models with different damage locations are imported into the Transient Structural module in ANSYS software; S22. Add the material for the pipeline structure in the Engineering Data material setting component and set the material parameters according to the material properties. The Poisson's ratio of the material is 0.3 and the elastic modulus is 206 GPa. Use the Model function of Transient Structural to mesh the pipeline. S23. Set the physical environment for meshing to Mechanical environment, set the mesh size, and select the mesh generation method. S24. Add a fixed constraint at one end of the pipeline to form a cantilever beam. Apply a random excitation vertically upward and perpendicular to the pipeline axis at the other end. The maximum and minimum values ​​of the random excitation are 200N and -200N, respectively. The sensor acquisition frequency is 1000Hz, and the acquisition time is 8s. The strain response data and acceleration response data are obtained and exported to an Excel spreadsheet.

[0009] In this technical solution, based on finite element analysis, the built-in material setting components and other modules are used to set parameters for the simulated three-dimensional model of the ocean pipeline and divide it into parts to obtain a finite element model; by setting the parameters of the strain sensor and acceleration sensor, the strain response data and acceleration response data can be simulated.

[0010] According to one embodiment of the present invention, in the dendritic network optimized by the improved Wild Horse algorithm in step S3, the feature extraction module includes two dendritic networks, and the dendritic network is composed of a DD module and a linear module. The DD module is expressed as follows: (1) (2) in: and Respectively represent the input and output of the DD module; and Represent the original input and output of the DD module respectively; Indicates the Modules to The weight matrix of the modules; ◦ represents the Hadamard product; the last module is linear; Indicates the number of modules; The forward propagation of the DD module and the linear module is expressed as follows: (3) The error back propagation of the DD module and the linear module is expressed as follows: (4) (5) (6) The weight adjustment of DD is expressed as follows: (7) (8) in: and Represent the output and label of the DD module respectively; Indicates the number of training samples in each batch; based on the heuristic algorithm, Represents the learning rate that is adaptively adjusted over time.

[0011] In this technical solution, based on the strain response data and acceleration response data obtained by simulation, a dendritic network optimized by the improved Mustang algorithm is constructed. By improving the dendritic network, the learning rate of each batch of training samples is continuously improved.

[0012] According to one embodiment of the present invention, in the dendritic network optimized by the improved wild horse algorithm in step S3, the weight feature fusion module is an MSD function weight feature fusion module, and the feature fusion weight is introduced. and The importance of fusion of features from different information sources is expressed by weights, so that features with a higher correlation with the damage condition account for a higher proportion of the fused features, and the model decision performance is greatly improved. The feature fusion weight is expressed as follows: (9) (10) in: ,and , [0, 1]; The double Cartesian product is expressed as follows: (11) Where: M and N represent the feature vectors output by each sub-network respectively; the two feature vectors are fused by outer product as shown in the formula: (12) Represents the tensor product, also known as the outer product, which is an operation between two tensors of any size; (13) The feature fusion method is combined with the weight coefficient to construct the MSD function weight feature fusion module to realize multi-source information fusion. The fusion method is expressed as follows: (14) in: Represents the fused features output by the MSD function weight feature fusion module.

[0013] In this technical solution, based on the strain response data and acceleration response data obtained by simulation, a dendritic network optimized by the improved Mustang algorithm is constructed, and the fused features of each batch of training samples are continuously improved through the improvement of multi-source information fusion.

[0014] According to one embodiment of the present invention, in the dendritic network optimized by the improved wild horse algorithm in step S3, the multi-head attention module calculates the attention weight of each position in the input sequence to other positions and generates a weighted intermediate representation, calculates the attention function of a set of query vectors and packages them into a matrix Q, and the key vector and value vector are packaged into matrices K and V respectively. The output matrix is ​​expressed as follows: (13) in: Represents the dimension of the key vector; the SoftMax function is applied to the scaled dot product result; the SoftMax function converts the scaled dot product into a non-negative attention weight that sums to 1. The attention weight reflects the importance of different positions to the current query; The formula for multi-head attention is as follows: (14) (15) Among them: projection represents the parameter matrix: , , , ; h represents the number of projections, Represents the model dimension, represents the dimension of the key, Represents the dimension of the value.

[0015] In this technical solution, based on the strain response data and acceleration response data obtained from the simulation, a dendritic network optimized by the improved wild horse algorithm is constructed. By calculating the attention weight of each position in the input sequence to other positions, the multi-head attention structure composed of multiple parallel attention layers is improved. Specifically, the fused features output by the weighted feature fusion module are used as the input of the multi-head attention module. The input vector is copied into multiple parallel and independent subspaces respectively. The input vector is linearly transformed using a learnable weight matrix and mapped to different representation spaces, so that the input vector is decomposed into multiple dimensions to capture different information features in each subspace. For each subspace, scaled dot product attention is used to calculate the attention score reflecting the degree of match between the query and the key. The attention score is normalized by the SoftMax function and converted into an attention weight. The weight ensures that the contribution of each key to the final output is proportional to its degree of match with the query. After obtaining the attention weight, it is multiplied by the corresponding value vector to obtain the weighted value vector; the input information is selectively focused so that the model can more accurately focus on the part most relevant to the query; the output of each subspace is spliced ​​through the concat operation to form a unified representation; then, it is transformed through a linear layer to further integrate and refine the information, and finally the output of the multi-head attention module is obtained.

[0016] According to one embodiment of the present invention, in the dendritic network optimized by the improved Wild Horse algorithm in step S3, the Wild Horse algorithm hyperparameter optimization module considers the influencing factors of learning rate, number of layers, and batch size to optimize the hyperparameters of the dendritic neural network.

[0017] According to one embodiment of the present invention, in the dendritic network optimized by the improved Wild Horse algorithm in step S3, the decision module is composed of a fully connected layer, and the decision module makes a decision by processing the fused features and finally outputs the result.

[0018] According to one embodiment of the present invention, in the pipeline damage identification experimental platform of step S4, the data acquisition model is DH8301N, the power amplifier model is DH5871, the exciter model is DH40500, the strain sensor model is BJN-FBG-PI, the acceleration sensor model is 1C302, the demodulator model is Si255, and the dynamic signal test and analysis system model is DH8302-1.

[0019] In this technical solution, the computer issues instructions according to the set excitation parameters, the signal collector generates an excitation signal after receiving the instructions, and the power amplifier amplifies the excitation signal to a sufficient power level to drive the vibrator to generate the required vibration. At the same time, the power amplifier can control the vibration amplitude and other parameters of the vibrator to meet the needs of the experiment or research. In order to keep the relative position of the vibrator and the pipeline unchanged, the vibrator and the pipeline are fixed to the test bench. The strain response data and acceleration data generated by the pipeline vibration are measured by the strain sensor and the acceleration sensor respectively. The strain response data is transmitted to the fiber grating demodulator through the strain sensor. The strain change data is demodulated by measuring the peak wavelength, valley wavelength and full spectrum of the fiber grating. The demodulated signal is then transmitted to the computer and the signal is analyzed and stored by the ENLIGHT software. The signal of the acceleration sensor is stored in the same way.

[0020] According to one embodiment of the present invention, in the pipeline damage identification experimental platform of step S4, the positioning result indicators include precision, recall, and F1 value. Among them, precision is used to indicate the accuracy of the model in identifying a certain type of damage; recall is used to indicate the model's ability to find a certain type of damage, that is, how many actual damages of a certain type the model can find; and F1 value is used to indicate the model's performance in classifying different categories.

[0021] In this technical solution, the damaged pipe is placed on the experimental platform, and a method of fixing one end and exciting the other end is adopted to obtain new strain data obtained from the experiment under actual working conditions. This data is used as a test set, and the data in the test set is used for testing and the damage location results are summarized. Accuracy, also known as the precision rate, is used to characterize the accuracy of the model in identifying a certain type of damage. Recall, also known as the recall rate, is used to characterize the model's recall ability for a certain type of damage, that is, how many actual damages of a certain type the model can find. The recall rate is of great significance. A model with a high recall rate can more comprehensively cover a certain type of damage that actually exists, thereby reducing the possibility of missed detection. The F1 value directly reflects the performance of the model in different classifications. A high F1 value means that the model performs well in both precision and recall, and can provide an important reference for optimizing the model. The three indicators reflect the performance of the model from different perspectives.

[0022] Compared with the prior art, the present invention has the following beneficial effects: (1) In the field of pipeline damage detection, the present invention can process and analyze a large amount of signal data and quickly locate the damage location, so that the present invention can achieve real-time monitoring and early warning of pipelines and timely identify and deal with potential safety risks.

[0023] (2) The weighted feature fusion module provided by the present invention, namely the MSD function weighted feature fusion module, assigns different weights according to the importance of different features, so that the feature tensors can jointly and effectively characterize the pipeline damage status. It not only considers the contribution of different features, but also retains key information during the fusion process, thereby improving the representation ability and prediction accuracy of the model.

[0024] (3) The dendritic network optimized by the improved Mustang algorithm of the present invention can analyze the pipeline status from multiple angles and dimensions during the strain data fusion process for strain sensors and acceleration sensors, thereby more accurately identifying the damage location, which is used to reduce maintenance costs and improve pipeline safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is one of the damage schematic diagrams of the marine pipeline of the present invention.

[0026] Figure 2 This is the second damage schematic diagram of the marine pipeline of the present invention.

[0027] Figure 3 It is a schematic diagram of the pipeline sampling point positions of the present invention.

[0028] Figure 4 It is a line graph of the random excitation of the pipeline of the present invention.

[0029] Figure 5 This is a framework diagram of the dendritic network model for deep fusion of multi-source information of the present invention.

[0030] Figure 6 It is the DD module structure diagram of the present invention.

[0031] Figure 7 This is a visual architecture diagram of the DD module of the present invention.

[0032] Figure 8 It is the scaled dot product attention structure diagram of the present invention.

[0033] Figure 9 This is a multi-head attention structure diagram composed of multiple attention layers running in parallel in the present invention.

[0034] Figure 10 It is a data collection flow chart of the present invention.

[0035] Figure 11 It is a broken line graph of damage location using the accuracy model of the present invention.

[0036] Figure 12 It is a line graph of damage positioning using the recall rate model of the present invention.

[0037] Figure 13 It is a line graph of damage positioning using the F1 value model of the present invention.

[0038] Figure 14 It is a box diagram of damage location using the accuracy model of the present invention.

[0039] Figure 15 It is a box plot of damage localization using the recall model of the present invention.

[0040] Figure 16 It is a box plot of damage positioning using the F1 value model of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0042] Example 1 like Figure 1 As shown, this embodiment provides a method for locating marine pipeline damage using a dendritic network optimized by an improved Mustang algorithm, comprising the following steps: S1. Establishment of marine pipeline damage model: A three-dimensional model of the structure is established in SOLIDWORKS according to the simulation requirements. In the present invention, the model is 13 marine pipelines with crack damage. The overall structural parameters of the pipelines are shown in Table 1.

[0043] Table 1 Parameters of marine pipelines

[0044] Pipeline damage features are established on the surface of the pipeline. Cracks are simulated by digging elongated grooves on the surface of the pipeline. The distance from the crack to the fixed support end is X. The sizes are 200mm, 203mm, 206mm, 209mm, 212mm, 215mm, 218mm, 221mm, 224mm, 227mm, 230mm, and 233mm respectively. There are 13 pipelines with different damage locations. The crack size of the 13 pipelines is L D H, where L is the damage length (damage arc length), which is 16mm, D is the damage width, which is 2mm, and H is the damage depth, which is 0.5mm. Figure 1 and Figure 2 shown.

[0045] S2. Acquisition of strain data set: Use finite element ANSYS software to build a finite element model, import the newly built 13 pipeline models with different damage locations into the Transient Structural module in ANSYS software, add the material of the pipeline structure in the material setting component Engineering Data, set the material parameters according to the material properties, set the material Poisson's ratio to 0.3, and the elastic modulus to 206GPa, and use the Model function of Transient Structural to mesh the pipeline. Set the physical environment of meshing to Mechanical environment, set the mesh size, and select the mesh generation method. Add strain sensors and acceleration sensors to the pipeline surface. In order to conduct a comprehensive detection of damage on the pipeline surface, evenly arrange the sensors. The positions of the sensors are as follows: Figure 3 A fixed constraint is added at one end of the pipe to form a cantilever beam, and a random excitation is applied vertically upward and perpendicular to the pipe axis at the other end. The maximum and minimum values ​​of the random excitation are 200N and -200N respectively. The random excitation is as follows: Figure 4 As shown, the sensor acquisition frequency is 1000 Hz and the acquisition time is 8 seconds. Strain and acceleration analysis was performed on the 13 pipes mentioned above. The strain response data and acceleration response data were obtained and exported to Excel tables to obtain a strain dataset. The data was divided into a training set and a validation set.

[0046] S3. Establishment of a dendritic network optimized with an improved Mustang algorithm: In practical engineering applications, any overlooked or misjudged safety hazards increase the risk of pipeline rupture or leakage. Further improving identification accuracy is essential to meet the requirements of practical engineering applications. External loads in practical engineering applications not only cause strain in the pipeline but also cause micro-displacements, which in turn generate acceleration. Acceleration can serve as another important indicator of pipeline damage. Using strain response data and acceleration response data as multi-source information for pipeline damage identification is feasible. This paper proposes a dendritic network optimized with an improved Mustang algorithm (IWHO-MFDD) for damage location and identification, achieving better results in marine pipeline damage identification.

[0047] The improved Wild Horse algorithm optimized dendritic network (IWHO-MFDD) mainly consists of five modules. The first module is the feature extraction module, which is composed of two sub-networks based on the dendritic module. Each sub-network accepts a type of input data and extracts features from the data according to the principle of the dendritic network. After the feature extraction is completed, the module outputs two sets of feature vectors based on strain response data and acceleration response data respectively. The two sets of feature vectors contain different information about structural damage. The second module is the weighted feature fusion module, which can set the fusion weight according to the richness of the features contained in different data to ensure the richness and high correlation of the fused features to the greatest extent. The third module is the multi-head attention module, through which the input features are further fused, and different attention information is captured from different subspaces through multiple heads, thereby improving the expressive ability of the model. The fourth module is the Wild Horse algorithm optimization hyperparameter module. The fifth module is the decision module, which is composed of a fully connected layer. The module makes decisions by processing the fused features. In addition, the model also includes an input layer and an output layer. The structural framework of the model is shown in the figure below. Figure 5 shown.

[0048] The feature extraction module mainly includes two dendritic networks (DD), which are composed of DD modules and linear modules. The DD module is relatively simple, and its structure is as follows: Figure 6 shown.

[0049] The DD module is expressed as follows: (1) (2) in: and Respectively represent the input and output of the DD module; and Represent the original input and output of the DD module respectively; Indicates the Modules to The weight matrix of the modules; ◦ represents the Hadamard product; the last module is linear; Indicates the number of modules; The DD module uses a DD learning rule based on error back propagation. The learning method is simple and facilitates the application of DD in different fields. The forward propagation of the DD module and the linear module is expressed as follows: (3) The error back propagation of the DD module and the linear module is expressed as follows: (4) (5) (6) The weight adjustment of DD is expressed as follows: (7) (8) in: and Represent the output and label of the DD module respectively; Indicates the number of training samples in each batch; based on the heuristic algorithm, Represents the learning rate that is adaptively adjusted over time, and can also be fixed at a smaller value. The visual principle diagram of the DD module is as follows Figure 7 shown.

[0050] The weighted feature fusion module can be categorized as either early or late fusion, depending on where feature fusion is performed within the model. Obviously, in the weighted feature fusion module, feature fusion occurs before the decision module, making it considered early fusion. Early fusion primarily employs two fusion methods: serial feature fusion (concat) and parallel fusion (add). Serial feature fusion is the most straightforward fusion method. In deep learning, when two or more features need to be fused, these feature vectors are concatenated along a certain dimension (usually the feature dimension) to form a longer feature vector. The parallel strategy simultaneously considers multiple information sources or features and combines them in some manner (usually element-wise addition). Both of these methods are essentially still at the initial concatenation stage of joint feature representation. These feature fusion methods result in features with varying degrees of correlation with the output being fused with equal weights, which fails to fully utilize the available features.

[0051] This paper proposes a new weighted feature fusion module, namely the MSD function weighted feature fusion module. This fusion method is an improved weighted feature fusion method based on the SoftMax function. Compared with the two feature tensor concatenation methods mentioned above, it is fundamentally different. This module can assign different weights to different features based on their importance, so that the feature tensors can jointly and effectively represent the pipeline damage status. This method not only considers the contribution of different features but also retains key information during the fusion process, improving the model's representation capabilities and prediction accuracy. The following will introduce the weighted feature fusion module from the perspectives of weight and fusion.

[0052] The MSD function weight feature fusion module introduces feature fusion weight and The importance of fusing features from different information sources is represented by weights, so that features with a higher correlation with the damage condition account for a higher proportion of the fused features, and the model decision-making performance is greatly improved.

[0053] The feature fusion weight is expressed as follows: (9) (10) in: ,and , [0, 1]; The double Cartesian product is expressed as follows: (11) Where: M and N represent the feature vectors output by each sub-network respectively; the two feature vectors are fused by outer product as shown in the formula: (12) Represents the tensor product, also known as the outer product, which is an operation between two tensors of any size; (13) The feature fusion method is combined with the weight coefficient to construct the MSD function weight feature fusion module to realize multi-source information fusion. The fusion method is expressed as follows: (14) in: Represents the fused features output by the MSD function weight feature fusion module.

[0054] like Figure 8 As shown, the multi-head attention module. The attention mechanism is a concept in deep learning that simulates how humans focus on different parts of information when processing information. In this mechanism, the model assigns a weight to each part of the input data, indicating how much attention the model should pay to that part for the current task. By weighting different parts of the input data, the model can better process task-relevant information while ignoring irrelevant information.

[0055] like Figure 9 As shown in the figure, scaled dot product attention is a core component of the self-attention mechanism, ensuring both efficiency and stability when calculating attention weights. The query, key, and value vectors corresponding to each position in the self-attention mechanism are obtained by linearly transforming the input data. The dot product between the query and all keys is calculated. The dot product operation measures the similarity between the query and each key, and the larger the value, the higher the similarity. However, directly using the dot product result may lead to gradient vanishing or gradient exploding problems, especially when the dot product result is large. To avoid this, a scaling factor is introduced, which is usually a fixed value ,in is the dimension of the key vector. The dot product result of each key is divided by the scaling factor to obtain the scaled dot product. The SoftMax function is applied to the scaled dot product result. The SoftMax function converts the scaled dot products into non-negative attention weights that sum to 1. These weights reflect the importance of different positions to the current query. Finally, these attention weights are applied to the corresponding value vectors. Specifically, each value vector is multiplied by its corresponding attention weight, and then all weighted value vectors are added together to obtain a weighted feature vector. This intermediate representation fuses the information of all positions and is weighted according to the attention weights, thereby highlighting the most relevant parts to the query. Through the above process, the scaled dot product attention mechanism can efficiently calculate the attention weight of each position in the input sequence to other positions and generate a weighted intermediate representation, providing key information for subsequent model processing.

[0056] Calculate the attention function of a set of query vectors and pack them into matrix Q, key vectors and value vectors into matrices K and V respectively. The output matrix is ​​shown in the formula: (13) in: Represents the dimension of the key vector; the SoftMax function is applied to the scaled dot product result; the SoftMax function converts the scaled dot product into a non-negative attention weight that sums to 1. The attention weight reflects the importance of different positions to the current query; The scaled dot product attention structure is as follows Figure 8 shown.

[0057] The keys, values, and queries in the attention mechanism are all of the same dimension as the model, so the model directly uses the information of the input data embedding dimension to perform attention calculations. But now, we have found a better way: instead of using only one dimension for attention calculations, we project the query, key, and value information multiple times into different dimensions (respectively 、 and This allows the model to focus on different parts of the data from multiple perspectives and dimensions, and for each projection, the model performs attention calculations in parallel. This means the model can focus on multiple aspects of the data simultaneously, improving information processing efficiency. The results of all parallel calculations are then merged and projected again to produce the final output.

[0058] The formula for multi-head attention is as follows: (14) (15) Among them: projection represents the parameter matrix: , , , ; h represents the number of projections, Represents the model dimension, represents the dimension of the key, Represents the dimension of the value.

[0059] The fused features output by the weighted feature fusion module serve as the input to the multi-head attention module. The input vector is replicated into multiple parallel and independent subspaces. A learnable weight matrix is ​​used to linearly transform the input vector, mapping it to different representation spaces. This decomposes the input vector into multiple dimensions, capturing different information features in each subspace. For each subspace, scaled dot-product attention is used to calculate an attention score reflecting the degree of match between the query and the key. The attention scores are normalized using a softmax function and converted into attention weights. These weights ensure that each key's contribution to the final output is proportional to its match with the query. Once the attention weights are obtained, they are multiplied by the corresponding value vector to produce a weighted value vector. This step is key to selectively focusing on the input information, enabling the model to more accurately focus on the most relevant parts of the query. The outputs of each subspace are concatenated using a concat operation to form a unified representation. This representation is then transformed through a linear layer to further integrate and refine the information, ultimately resulting in the output of the multi-head attention module.

[0060] Then, the hyperparameters of the neural network are optimized. The learning rate, number of layers, and batch size are all factors that affect the neural network model. Based on the dendritic neural network, the present invention uses the improved Wild Horse Optimization (IWHO) algorithm to optimize the hyperparameters of the dendritic neural network.

[0061] The decision module consists of a fully connected layer, which processes the fused features to make decisions and finally outputs the results.

[0062] S4. Establishment of a Pipeline Damage Identification Experimental Platform: The pipeline damage identification experimental platform primarily consists of a signal acquisition device, power amplifier, vibrator, pipeline, strain sensor, accelerometer, grating light demodulator, dynamic signal testing and analysis system, and a computer equipped with signal acquisition and analysis software and ENLIGHT. The models of each device are listed in Table 2.

[0063] Table 2 Pipeline damage identification experimental system equipment information

[0064] like Figure 10As shown in the figure, the experimental process is as follows: the computer software issues commands according to the set excitation parameters. The signal acquisition device generates an excitation signal after receiving the commands. The power amplifier amplifies the excitation signal to a sufficient power level to drive the vibrator to produce the desired vibration. Simultaneously, the power amplifier can control parameters such as the vibration amplitude of the vibrator to meet the needs of the experiment or research. To maintain the relative position of the vibrator and pipeline, the vibrator and pipeline are fixed to the test bench. The strain response data and acceleration data generated by the pipeline vibration are measured by the strain sensor and accelerometer, respectively. The strain response data is transmitted to the fiber Bragg grating (FBG) demodulator via the strain sensor. The strain change data is demodulated by measuring the peak wavelength, valley wavelength, and full spectrum of the FBG. The demodulated signal is then transmitted to the computer for analysis and storage using ENLIGHT software. The accelerometer signal is stored in the same manner.

[0065] The signal data of strain and acceleration are obtained through experiments, and the data are imported into a table to obtain a test set.

[0066] S5. Training and testing of experimental data sets: The trained neural network model, i.e., the improved Wild Horse algorithm optimized dendritic network (IWHO-MFDD) model, is used to perform feature processing on the experimental data sets and obtain the test accuracy.

[0067] Example 2 Based on Example 1, the present invention provides a specific implementation case of marine pipeline damage location based on a dendritic network optimized by an improved Mustang algorithm.

[0068] S1. Schematic diagram of damage to marine pipelines is shown in the attached figure. Figure 1 and Figure 2 As shown. Figure 1 and Figure 2 The damaged pipeline model in the paper is put into ANSYS software for strain analysis to obtain the strain and acceleration data of the pipeline.

[0069] S2. Process the strain and acceleration data of the 13 pipes obtained in S1 and input them into the improved Wild Horse algorithm-optimized dendritic network model, namely the IWHO-MFDD model, for training. The neural network model IWHO-MFDD in this invention is developed based on Python 3.8 and Pytorch 1.10 framework. The experimental platform hardware configuration: 4GB NVIDIA GeForce RTX 1650 GPU, Intel Core Ryzen i5-10400 CPU, 16GB memory. The structural framework of the improved Wild Horse algorithm-optimized dendritic network model is as follows: Figure 3 shown.

[0070] S3. Use the training set to train the improved Wild Horse algorithm optimized dendritic network (IWHO-MFDD) model and five other network models (DD, WNN, TCNN, RNN, LSTM), and adjust parameters such as learning rate and number of iterations according to the loss curve during the iteration process of the validation set.

[0071] S4. Using the constructed experimental platform, the damaged pipe was placed on the platform. Using a fixed-end and stimulated-end method, new strain data was obtained from the experiment under actual working conditions. This data was used as the test set for testing and the damage localization results were summarized. The main indicators of the localization results include precision, recall, and F1 score. Precision, also known as the recall rate, measures the model's accuracy in identifying a specific type of damage. Recall, also known as the recall rate, measures the model's ability to detect a specific type of damage, that is, how many actual damages of that type the model can identify. Recall is very important. A model with a high recall rate can more comprehensively cover actual damage types, thereby reducing the possibility of missed detections. The F1 score directly reflects the model's performance in classifying different categories. A high F1 score indicates that the model performs well in both precision and recall, providing important reference for model optimization. These three indicators reflect the model's performance from different perspectives.

[0072] Table 3 Accuracy of damage location of 6 models

[0073] First, Table 3 shows that IWHO-MFDD achieves the highest results among the six models in terms of maximum, minimum, and average positioning accuracy, with the highest accuracy reaching 100%. This indicates that IWHO-MFDD has the best positioning accuracy, while WNN has the lowest maximum, minimum, and average positioning accuracy, indicating the lowest positioning accuracy. Recall, also known as recall, is crucial for marine pipeline damage identification. Table 4 shows that IWHO-MFDD achieves the highest maximum, minimum, and average positioning recall, with the maximum value reaching 100%. This indicates that the model can ensure the timely detection and resolution of potential marine pipeline security threats, while WNN has the lowest maximum, minimum, and average recall rates. The F1 value, an important metric that reconciles precision and recall, provides a more comprehensive understanding of model performance. Table 5 shows that IWHO-MFDD achieves the highest maximum, minimum, and average F1 value among the six models, while WNN has the lowest.

[0074] Table 4 Recall rates of 6 models for damage location

[0075] Table 5 F1 values ​​of 5 models for damage location

[0076] In order to show the damage location effect more intuitively, each type of location result is presented in the form of a line graph, such as Figures 11 to 13 As shown. From the line graph, it can be seen that no matter in terms of precision, recall or F1 value, there is a clear stratification between the results of each model. The result value of IWHO-MFDD is the highest, indicating that its damage location performance is the best, which is also consistent with the data in Tables 3, 4 and 5. In addition, the positioning results of each model change with the damage conditions, indicating that the positioning accuracy of the model varies with the damage condition category. This is related to the random division of the data set. For example, if there are more samples of a certain type of damage condition in the test set, there will be more samples with recognition errors. Precision, recall and F1 value are used to comprehensively measure the performance of the model to eliminate the impact of random division of the data set as much as possible. Figures 11 to 13 It can be seen that the fluctuation range of each discount is certain, which means that the location difference of different damages by the same model is also certain, which to a certain extent reflects the stability of the damage location performance of the model.

[0077] By drawing box plots of the six models, we can observe the differences and judge the stability of the models. Figures 14 to 16 There are no other numerical points besides the upper and lower whiskers in the box plot, indicating that there are no abnormal data in the prediction accuracy of the six models, which further proves the reliability of the damage location dataset based on multi-source information. Figures 14 to 16 From the median, we can see that the median position of IWHO-MFDD box plot is the highest, indicating that its overall level of positioning performance is the highest. Conversely, the median position of WNN is the lowest, indicating that its overall level of damage localization performance is the lowest. Figures 14 to 16 The distances between the upper and lower quartiles in the data are the smallest for WNN, WNN, and IWHO-MFDD, indicating that the data dispersion of WNN and IWHO-MFDD is the smallest. This, in turn, indicates that WNN and IWHO-MFDD have the most stable damage localization performance. This indicates that IWHO-MFDD has significantly improved performance stability compared to DD. Therefore, in terms of comprehensive positioning accuracy and positioning stability, IWHO-MFDD has a significant advantage.

[0078] Although the present invention is described in detail with reference to the accompanying drawings and in combination with preferred embodiments, the present invention is not limited thereto. Without departing from the spirit and essence of the present invention, a person of ordinary skill in the art may make various equivalent modifications or substitutions to the embodiments of the present invention, and such modifications or substitutions shall be within the scope of the present invention. Any person skilled in the art who can easily conceive of changes or substitutions within the technical scope disclosed in the present invention shall be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope of protection of the claims.

Claims

1. A marine pipeline damage location method based on a dendritic network optimized by an improved Mustang algorithm, characterized in that: The steps include: S1. Establishment of marine pipeline damage model: simulate a marine pipeline with crack damage, and dig a long and narrow groove on the surface of the marine pipeline to simulate the crack. The distance from the crack to the fixed support end is X, and the crack size is L. D H, where L is the damage length, D is the damage width, and H is the damage depth; S2. Acquisition of strain dataset: Construct a pipeline model with different damage locations, add strain sensors and acceleration sensors to the surface of the pipeline model, perform strain and acceleration analysis, and obtain strain response data and acceleration response data to obtain a strain dataset. S3. Establishment of a dendritic network optimized by the improved Mustang algorithm: Strain response data and acceleration response data are used as multi-source information to identify pipeline damage. A dendritic network optimized by the improved Mustang algorithm is constructed. The dendritic network includes a feature extraction module, a weighted feature fusion module, a multi-head attention module, a Mustang algorithm hyperparameter optimization module, and a decision module for damage location and identification. S4. Establishment of a pipeline damage identification experimental platform: The pipeline damage identification experimental platform includes a signal collector, a power amplifier, an exciter, a pipeline, a strain sensor, an acceleration sensor, a grating optical demodulator, a dynamic signal test and analysis system, and a computer. The signal collector generates an excitation signal after receiving instructions, and the power amplifier amplifies the excitation signal to control the vibration amplitude of the exciter. The exciter and pipeline are fixed to the test bench. The strain response data and acceleration data generated by pipeline vibration are measured by the strain sensor and the acceleration sensor respectively. The data are transmitted to the fiber Bragg grating demodulator via the strain sensor. The strain change data is demodulated by measuring the peak wavelength, valley wavelength, and full spectrum of the fiber Bragg grating. The demodulated signal is then transmitted to the computer. The strain and acceleration signal data are obtained through the pipeline damage identification experimental platform to obtain the experimental data set. S5. Training and testing of experimental data sets: The dendritic network optimized by the improved Wild Horse algorithm is used to perform feature processing on the experimental data sets and obtain the test accuracy.

2. The marine pipeline damage location method based on the dendritic network optimized by the improved Mustang algorithm according to claim 1 is characterized in that: In the establishment of the marine pipeline damage model in step S1, SOLIDWORKS is used to simulate a three-dimensional model of a marine pipeline with crack damage, and the crack size is 16 mm. 2mm 0.5mm, the distances from the cracks of the 13 pipes to the fixed support ends are 200mm, 203mm, 206mm, 209mm, 212mm, 215mm, 218mm, 221mm, 224mm, 227mm, 230mm and 233mm respectively.

3. The marine pipeline damage location method based on the dendritic network optimized by the improved Mustang algorithm according to claim 1 is characterized in that: The acquisition of the strain data set in step S2 includes the following steps: S21. Using finite element ANSYS software, a finite element model is established, and the pipeline models with different damage locations are imported into the Transient Structural module in ANSYS software; S22. Add the material for the pipeline structure in the Engineering Data material setting component and set the material parameters according to the material properties. The Poisson's ratio of the material is 0.3 and the elastic modulus is 206 GPa. Use the Model function of Transient Structural to mesh the pipeline. S23. Set the physical environment for meshing to Mechanical environment, set the mesh size, and select the mesh generation method. S24. Add a fixed constraint at one end of the pipeline to form a cantilever beam. Apply a random excitation vertically upward and perpendicular to the pipeline axis at the other end. The maximum and minimum values ​​of the random excitation are 200N and -200N, respectively. The sensor acquisition frequency is 1000Hz, and the acquisition time is 8s. The strain response data and acceleration response data are obtained and exported to an Excel spreadsheet.

4. The marine pipeline damage location method based on the dendritic network optimized by the improved Mustang algorithm according to claim 3, characterized in that: In the dendritic network optimized by the improved Wild Horse algorithm in step S3, the feature extraction module includes two dendritic networks, and the dendritic network is composed of a DD module and a linear module. The DD module is expressed as follows: （1） （2） in: and Respectively represent the input and output of the DD module; and Represent the original input and output of the DD module respectively; Indicates the Modules to The weight matrix of the modules; ◦ represents the Hadamard product; the last module is linear; Indicates the number of modules; The forward propagation of the DD module and the linear module is expressed as follows: （3） The error back propagation of the DD module and the linear module is expressed as follows: （4） （5） （6） The weight adjustment of DD is expressed as follows: （7） （8） in: and Represent the output and label of the DD module respectively; Indicates the number of training samples in each batch; based on the heuristic algorithm, Represents the learning rate that is adaptively adjusted over time.

5. The marine pipeline damage location method based on the dendritic network optimized by the improved Mustang algorithm according to claim 4, characterized in that: In the dendritic network optimized by the improved wild horse algorithm in step S3, the weight feature fusion module is an MSD function weight feature fusion module, and the feature fusion weight is introduced. and The importance of fusion of features from different information sources is expressed by weights, so that features with a higher correlation with the damage condition account for a higher proportion of the fused features, and the model decision performance is greatly improved. The feature fusion weight is expressed as follows: （9） （10） in: ,and , [0, 1]; The double Cartesian product is expressed as follows: （11） Where: M and N represent the feature vectors output by each sub-network respectively; the two feature vectors are fused by outer product as shown in the formula: （12） Represents the tensor product, also known as the outer product, which is an operation between two tensors of any size; （13） The feature fusion method is combined with the weight coefficient to construct the MSD function weight feature fusion module to realize multi-source information fusion. The fusion method is expressed as follows: （14） in: Represents the fused features output by the MSD function weight feature fusion module.

6. The marine pipeline damage location method using a dendritic network optimized by an improved Mustang algorithm according to claim 5, characterized in that: In the dendritic network optimized by the improved wild horse algorithm in step S3, the multi-head attention module calculates the attention weight of each position in the input sequence to other positions and generates a weighted intermediate representation. It calculates the attention function of a set of query vectors and packages them into a matrix Q. The key vector and value vector are packaged into matrices K and V respectively. The output matrix is ​​expressed as follows: （13） in: Represents the dimension of the key vector; the SoftMax function is applied to the scaled dot product result; the SoftMax function converts the scaled dot product into a non-negative attention weight that sums to 1. The attention weight reflects the importance of different positions to the current query; The formula for multi-head attention is as follows: （14） （15） Among them: projection represents the parameter matrix: , , , ; h represents the number of projections, Represents the model dimension, represents the dimension of the key, Represents the dimension of the value.

7. The marine pipeline damage location method based on the dendritic network optimized by the improved Mustang algorithm according to claim 6 is characterized in that: In the dendritic network optimized by the improved Wild Horse algorithm in step S3, the Wild Horse algorithm hyperparameter optimization module considers the influencing factors of learning rate, number of layers, and batch size to optimize the hyperparameters of the dendritic neural network.

8. The marine pipeline damage location method based on the dendritic network optimized by the improved Mustang algorithm according to claim 7 is characterized in that: In the dendritic network optimized by the improved Wild Horse algorithm in step S3, the decision module is composed of a fully connected layer. The decision module makes a decision by processing the fused features and finally outputs the result.

9. The marine pipeline damage location method based on the dendritic network optimized by the improved Mustang algorithm according to claim 1, characterized in that: In the pipeline damage identification experimental platform of step S4, the data collector model is DH8301N, the power amplifier model is DH5871, the exciter model is DH40500, the strain sensor model is BJN-FBG-PI, the acceleration sensor model is 1C302, the demodulator model is Si255, and the dynamic signal test and analysis system model is DH8302-1.

10. The marine pipeline damage location method based on dendritic network optimized by improved wild horse algorithm according to claim 1, characterized in that: In the pipeline damage identification experimental platform of step S4, the positioning result indicators include precision, recall, and F1 value. Precision is used to indicate the accuracy of the model in identifying a certain type of damage; recall is used to indicate the model's ability to find a certain type of damage, that is, how many actual damages of a certain type the model can find; and F1 value is used to indicate the model's performance in classifying different categories.

Images