Offshore platform flow pipeline sensor layout optimization method, system, and storage medium

Abstract

The present invention relates to an optimization method, system and storage medium for sensor placement in a flow pipeline of an offshore platform, wherein the optimization method includes: obtaining parameter information of the flow pipeline of the offshore platform, constructing a finite element simulation model, obtaining strain modal data, obtaining strain modal data, and constructing a training data set; performing noise simulation, data interpolation, and linear normalization on the training data set; performing time-frequency analysis on the continuous sequence data set using wavelet packet analysis to obtain a damage feature information set of the strain modal data of the sensor corresponding to the flow pipeline of the offshore platform; establishing a damage feature-damage position mapping relationship set through a convolutional neural network; and optimizing and adjusting the sensor placement position using an improved whale optimization algorithm. Compared with the prior art, the present invention couples the sensor placement scheme with pipeline damage identification, effectively improving the sensor placement efficiency and the accuracy and precision of damage identification, and improving the unmanned and intelligent level of health monitoring of active offshore platforms.

Description

Technical Field

[0001] The present invention relates to the field of machinery and marine engineering, and in particular to a method, system and storage medium for optimizing the placement of sensors in a flow pipeline of an offshore platform. Background Art

[0002] On offshore platforms, flow pipelines play an important role. They are responsible for transporting fluid media such as water, oil, and natural gas. Due to the particularity of the working environment of marine engineering, facing the high temperature, high humidity and corrosiveness of the marine environment, flow pipelines are very likely to be damaged, which in turn causes leakage of pipeline contents, threatening the normal service of marine engineering and the lives of workers. At present, in order to address this problem, sensors can be deployed on the surface of the pipeline to collect data such as pipeline acceleration, stress strain, and vibration frequency. After modal identification, the data is compared with the data when the pipeline is not damaged. This method has two defects. First, the sensor deployment scheme is mostly based on experience. Second, on the active platform, data in the undamaged state can no longer be collected. This brings great difficulties to the accuracy and precision of damage identification of offshore platform flow pipelines.

[0003] CN114857504A discloses a pipeline safety monitoring method based on distributed optical fiber sensors and deep learning, including the following steps: S1, collecting optical fiber vibration signals at various positions of the pipeline through a distributed optical fiber signal acquisition device; S2, performing data cleaning, wavelet packet denoising and normalization on the collected optical fiber vibration signals, and dividing the processed data into a training set, a test set and a prediction set; S3, constructing a convolutional neural network model, inputting the training set and test set data for learning and training, inputting the prediction set data to evaluate the performance of the model, and then using the convolutional neural network as a feature extractor and the support vector machine as a classification selector to establish a convolutional neural network and support vector machine joint model; S4, inputting the collected real-time data into the convolutional neural network and support vector machine joint model for identification, and classifying the type according to the output digital label to achieve pipeline safety monitoring. Although the scheme uses deep learning to improve the accuracy of pipeline safety detection, it still lacks technical guidance for the current difficulties in optimizing the layout of sensors in the current offshore platform flow pipeline, and R&D personnel are still required to further solve this problem. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method, system and storage medium for optimizing the deployment of sensors in the flow pipeline of an offshore platform. The present invention couples the sensor deployment scheme with pipeline damage identification, effectively improves the sensor deployment efficiency and the accuracy and precision of damage identification, and improves the unmanned and intelligent level of health monitoring of existing offshore platforms.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A first aspect of the present invention provides a method for optimizing the placement of sensors in a flow pipeline of an offshore platform, comprising the following steps:

[0007] S1: Obtain parameter information of the flow pipeline of the offshore platform, build a finite element simulation model, build pipeline damage information in the finite element simulation model, set the data acquisition starting point, obtain strain modal data on the axial path of the pipeline, obtain strain modal data at each sensor position based on the original sensor layout plan, and use the pipeline damage position as a label for the data to construct a training data set;

[0008] S2: performing noise simulation, data interpolation, and linear normalization on the training data set to obtain a preprocessed continuous sequence data set;

[0009] S3: performing time-frequency analysis on the continuous sequence data set using wavelet packet analysis to obtain a damage feature information set of the strain modal data of the sensor corresponding to the flow pipeline of the offshore platform;

[0010] S4: Based on the damage feature information set, establishing a damage feature-damage location mapping relationship set through a convolutional neural network;

[0011] S5: Based on the damage feature-damage location mapping relationship set and using an improved whale optimization algorithm, the sensor deployment position is optimized and adjusted.

[0012] Furthermore, in S1, the parameter information of the offshore platform fluid delivery pipeline includes length, outer diameter, inner diameter, fluid, material, and surrounding working environment parameters.

[0013] Further, in S1, the process of constructing pipeline damage information in the finite element simulation model includes:

[0014] Pipeline damage construction: Pipeline damage with a length of L mm, a width of W mm, and a depth of H mm is obtained by rotary cutting on the inner wall of the pipeline simulation model;

[0015] Obtaining the damage position mark: Obtain the distance X mm between one end point of the damage and the end of the pipe corresponding to the end point, and use the X mm as the damage position mark.

[0016] Furthermore, in S1, the process of constructing the training data set includes:

[0017] The Y mm at one end of the offshore platform flow pipeline simulation model is set as the data collection initial point, and the (L-30) mm at one end of the offshore platform flow pipeline simulation model is set as the data collection end point. The strain modal data on the pipeline axial path is obtained by finite element calculation. Then, according to the sensor layout plan, the strain modal data at each sensor layout position is taken out, which is a single training data, and the pipeline damage position Xmm is used as the label of the data.

[0018] Furthermore, in S2, during the noise simulation, white noise is added to the simulated strain modal data to simulate the generation of noise in actual engineering;

[0019] Among them, random numbers based on standard normal distribution are used to generate white noise:

[0020] x noise =x+normal(0,1) (1)

[0021] Among them, normal(0,1) is the standard normal distribution, x is the input data, x noise is the data after adding noise;

[0022] In the data interpolation process, three data interpolations are performed based on the original data features;

[0023] In the linear normalization process, the range of each order strain mode is mapped to [0, 1], and the dimension of the real sensor strain modal data of any order is unified. The normalization formula is as follows:

[0024]

[0025] Among them, min(x) is the minimum function and max(x) is the maximum function.

[0026] Further, in S3, the process of wavelet packet analysis includes:

[0027] The number of iterations for each data in the continuous sequence data set is num iter The binary decomposition of iter When the decomposition is performed, the frequency domain is divided into nodes, each node corresponds to a frequency band, and each frequency band contains the frequency domain feature information of the original signal within this range. The wavelet packet analysis process is shown in the following formula:

[0028]

[0029] In the formula, p (i，j) is the frequency domain coefficient of the jth node in the i-th iteration, H and G are a set of wavelet basis functions;

[0030] The result of the wavelet packet analysis is the coefficient vector p of all nodes in the last iteration. (last，j) , defined as follows:

[0031] p (last，j) =[fre1,Fre2,...,fre len ]

[0032] Among them, last is the last iteration, j is the jth node, Fre is the frequency feature, and len is the number of frequency features;

[0033] Based on the coefficient vector p of all nodes (last，j) A damage feature information set of the strain modal data of sensors corresponding to the flow pipeline of the offshore platform is constructed.

[0034] Furthermore, in S4, in the process of establishing the damage feature-damage location mapping relationship set through the convolutional neural network:

[0035] The coefficient stacking method is used to stack a series of two-dimensional matrices composed of damage features into a three-dimensional space. The height of the stacking is the number of channels input into the convolutional neural network.

[0036] During the stacking process, set the number of stacking layers to fold. The specific process is as follows:

[0037]

[0038] in, represents the frequency characteristics of all nodes after the last iteration in wavelet packet analysis, Represented as a length of num res , a two-dimensional matrix with a width of len, represents the folded frequency feature matrix, Represents a height of fold and length of (num res / fold), a three-dimensional matrix with a width of len,

[0039] → represents a folding process, which is implemented by calling the open source library numpy.reshape function. The specific process is:

[0040] The convolutional neural network is divided into an input layer, a hidden layer and an output layer;

[0041] The output layer is a fully connected layer and a LogSoftmax layer, which are used to obtain the probability of damage at each position of the pipeline. The hidden layer structure is three layers, and each layer is composed of convolution layer-BN layer-activation layer-pooling layer-regularization. The convolution kernel size is (5, 5), the pooling layer size is 2, the activation layer function is the ReLU() function, the regularization ratio is 0.1, the optimizer uses the RMSprop function, and the damage function uses the cross entropy CrossEntrop function. The initial learning rate is set to 0.05, and the learning rate decreases by 75% every 5 iterations. The maximum number of iterations is set to 200. The damage feature-damage position mapping relationship set is established through the convolutional neural network.

[0042] Furthermore, in S4, the process of optimizing and adjusting the sensor layout positions includes:

[0043] The damage identification and positioning accuracy is used as the objective function, and the number and spacing of sensors are used as optimization variables, as shown below:

[0044] var = [num, inter]

[0045] Among them, uar is the optimization variable, num is the number of sensors deployed, and inter is the spacing between sensors;

[0046] Afterwards, the whale algorithm is used to find the optimization variables that maximize the accuracy of damage identification and location, including the following steps:

[0047] Initialize the global variables of the whale algorithm. After entering the iteration, the specific evolutionary strategy executed by the current whale individual is determined according to a uniform random number r. When r≤0.5, the impact search strategy is executed, and when r>0.5, the spiral search strategy is executed.

[0048] In the direct search strategy, the whale individuals to be directly searched are determined according to the following formula;

[0049]

[0050]

[0051] in, is a random number ranging from 0 to 1;

[0052] when When , a whale individual is randomly selected from the population for search. The specific formula is as follows:

[0053]

[0054]

[0055]

[0056] in, For randomly selected whale individuals;

[0057] when When , the current global best individual is selected as the object to execute the spiral search strategy. The specific formula is as follows:

[0058]

[0059]

[0060]

[0061] in, is the best individual in the world;

[0062] The spiral search strategy is as follows:

[0063]

[0064]

[0065] After the current whale individual has executed the evolutionary strategy, it checks whether its offspring have crossed the boundary and repairs the variables that have crossed the boundary;

[0066] When the number of sensors deployed exceeds the variable range, the following process is used to repair the number of sensors that have exceeded the range:

[0067] First, the specific patching strategy executed by the current whale individual is determined by a random decimal p. When p≤0.5, the number of sensors deployed by the current best individual is selected to replace the current number of sensors deployed that cross the boundary. When p>0.5, the number of sensors deployed by the global best individual is selected to replace the current number of sensors deployed that cross the boundary.

[0068] After that, the number of sensors deployed by a whale individual in the population is randomly selected. When the number of sensors deployed by the random individual is equal to the number of sensors deployed by the global best individual, the number of sensors deployed by the offspring of the current iteration individual is randomly perturbed. The specific formula is as follows:

[0069]

[0070]

[0071] else:num=num+round(normal(0,1))

[0072] in, is the best individual in the world, is the current best individual, round() is the rounding function, and normal(0, 1) is the standard normal distribution;

[0073] Finally, after each whale individual executes the evolutionary strategy, it executes the global elitism strategy, comparing the offspring of the current iteration individual with their parents, and with the best individual in history and the best individual in the current group, so as to keep the best individual used in each whale iteration instant, thereby accelerating the convergence speed of the group;

[0074] When the iteration is finished, the optimization variables of the best individual in history are used as the optimal sensor deployment parameters.

[0075] The second aspect of the present invention provides an offshore platform flow pipeline sensor layout optimization system, including a simulation and data construction acquisition module, a data post-processing module, a feature analysis module, a convolution calculation module, and an optimization calculation module, wherein specifically:

[0076] The simulation and data construction acquisition module obtains parameter information of the offshore platform flow pipeline, constructs a finite element simulation model, constructs pipeline damage information in the finite element simulation model, sets a data acquisition starting point, obtains strain modal data on the pipeline axial path, obtains strain modal data at each sensor position based on the original sensor layout plan, and uses the pipeline damage position as a label for the data to construct a training data set;

[0077] The data post-processing module performs noise simulation, data interpolation, and linear normalization on the training data set to obtain a pre-processed continuous sequence data set;

[0078] The feature analysis module uses wavelet packet analysis to perform time-frequency analysis on the continuous sequence data set to obtain a damage feature information set of the strain modal data of the sensor corresponding to the offshore platform flow pipeline;

[0079] The convolution calculation module establishes a damage feature-damage position mapping relationship set through a convolutional neural network based on the damage feature information set;

[0080] The optimization calculation module optimizes and adjusts the sensor layout position based on the damage feature-damage position mapping relationship set and adopts an improved whale optimization algorithm.

[0081] A third aspect of the present invention provides a storage medium comprising computer executable instructions, which, when executed by a computer processor, are used to execute the above-mentioned method for optimizing the deployment of sensors in a flow pipeline of an offshore platform.

[0082] Compared with the prior art, the present invention has the following technical advantages:

[0083] 1) This application proposes a new method to patch out-of-bounds variables. The traditional variable patching method is to randomly select or replace the variable with the boundary of the variable interval. In the problem of optimizing the sensor layout of the offshore platform flow pipeline, the traditional variable patching method is not applicable to the number of sensors and the spacing of sensors. If the number of sensors and the spacing of sensors are reset to a certain boundary, this is undoubtedly not in line with the actual engineering needs; if the number of sensors and the spacing of sensors are randomly reset, this will affect the continued optimization of the existing optimization scheme. Therefore, the present invention fits the characteristics of the optimization problem of sensor layout of the offshore platform flow pipeline, and proposes a new variable resetting method. This method aims at the characteristics of the obvious local search characteristics of the number of sensors and the spacing of sensors in the optimization problem, so that the number of sensors and the spacing of sensors are randomly searched within a certain interval around the historical best individual that meets the requirements according to a certain strategy, thereby improving the optimization effect of sensor layout.

[0084] 2) This application aims to solve the problem of slow convergence of the original whale algorithm, and proposes a new elitist method. This method extends the comparison of the current individual to the best individual in history. By instantly comparing the current whale individual with the best individual in history, it ensures that each iterative individual can instantly learn the optimization information carried by the best individual in history. Compared with the traditional elitist method, this method can effectively improve the convergence speed and optimization efficiency of the optimization algorithm, and help the algorithm save optimization time when the population size is set to be large.

[0085] 3) This application aims to solve the problem that the interval between the number of sensors and the spacing between them is too large and the numerical types are different, which affects the optimization effect. The present invention proposes a variable isolation strategy method to isolate the number of sensors and the spacing between them during the iteration process. The two do not interfere with each other during the execution of the evolutionary strategy, and the position information of the number of sensors and the spacing between them is maintained to be passed to the next generation of whale individuals. If the number of sensors and the spacing between them interact through the evolutionary strategy, the number of sensors and the spacing between them will quickly reach the optimization boundary, and the optimization information of the previous generation cannot be well passed to the subsequent whale individuals.

[0086] 4) The present application proposes a wavelet packet analysis folding technology that can effectively improve the convolutional neural network. The traditional method is to input the results of wavelet packet analysis into the convolutional neural network as a two-dimensional matrix. Under this method, the convolutional neural network can only sense the frequency feature information on adjacent frequencies and adjacent coordinates, resulting in a smaller receptive field of the convolutional neural network. Therefore, the present invention proposes a wavelet packet analysis folding technology, which folds the frequency characteristics of all nodes of the wavelet packet analysis into a three-dimensional matrix in equal proportion, so that the convolutional neural network can sense the frequency characteristics of more nodes, improve the receptive field of the convolutional neural network, and help the damage function of the convolutional neural network to decrease faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 Schematic diagram of the wavelet packet analysis principle in an embodiment of the present invention. DETAILED DESCRIPTION

[0088] In specific implementation, the method for optimizing the layout of sensors in the flow pipeline of an offshore platform of the present invention adopts the following technical scheme, including the following steps:

[0089] Step 1: Model simulation

[0090] For the offshore platform flow pipeline where sensors need to be deployed, the parameters such as length, outer diameter, inner diameter, fluid, material, and surrounding working environment are collected, and a simulation model is constructed in the finite element software. The pipeline damage with a length of Lmm, a width of Wmm, and a depth of Hmm is obtained by rotary cutting on the inner wall of the pipeline simulation model. The position of the damage is marked by the distance Xmm between the left end of the damage and the left end of the pipeline. Since the front and rear ends of the pipeline are set as fixed supports in the finite element calculation model, the strain modal calculation results of the front and rear ends will produce large vibrations, which are inconsistent with the actual situation. Therefore, 30mm at the left end of the offshore platform flow pipeline simulation model is set as the data acquisition starting point, and (L-30)mm at the left end of the offshore platform flow pipeline simulation model is set as the data acquisition end point. The strain modal data on the axial path of the pipeline is calculated by the finite element calculation software. Then, according to the sensor layout plan, the strain modal data at each sensor layout position is taken out, which is a single training data. The pipeline damage position Xmm is used as the label of the data.

[0091] Step 2: Data preprocessing

[0092] (1) Noise simulation. Considering that the data collected by the sensor in actual engineering will be interfered by noise, it is necessary to add white noise to the sensor simulation strain modal data to simulate the generation of noise in actual engineering. The present invention uses random numbers based on standard normal distribution to simulate the generation of white noise. The formula is as follows:

[0093] x noise =x+normal(0,1) (1)

[0094] Among them, normal(0,1) is the standard normal distribution, x is the input data, x noise This is the data after adding noise. It is worth noting that the random noise of each discrete data point of the sensor simulation strain modal data here is different.

[0095] (2) Cubic spline interpolation. Since the number of sensors is usually small, the data length of the training set is short. Too short data will make it difficult to extract damage features, which will affect the accuracy of damage recognition of the neural network. Therefore, cubic spline interpolation is used here to interpolate data while ensuring the original data features as much as possible.

[0096] Since the cubic spline interpolation method is a relatively mature and widely open source interpolation method, and this method is not the main innovation of the present invention, its specific mathematical process is not described here. The specific process uses the interpolate class in the Python.Scipy library.

[0097] (3) Linear normalization. Since the ranges of strain modes of different orders are different, it is necessary to map the ranges of strain modes of different orders to [0, 1] to unify the dimensions of strain modal data of real sensors of any order. The normalization formula is as follows:

[0098]

[0099] Among them, min(x) is the minimum function and max(x) is the maximum function.

[0100] Step 3: Wavelet packet analysis

[0101] The data after noise simulation-cubic spline interpolation-normalization is regarded as a set of continuous sequence data, and the wavelet packet analysis is used to perform time-frequency analysis. Wavelet packet analysis is an iterative process. In each iteration, the wavelet packet analysis will convert the input data into low-frequency coefficients and high-frequency coefficients of the input data. The current low-frequency coefficients and high-frequency coefficients are the input data for the next wavelet packet analysis. Therefore, the number of iterations of the wavelet packet analysis needs to be determined according to the specific implementation plan, and the number of results of the wavelet packet analysis is determined by the following formula:

[0102]

[0103] Among them, num res is the number of wavelet packet analysis results, num iter is the number of wavelet packet iterations.

[0104] The iterative diagram of wavelet packet is as follows Figure 1 shown. Figure 1 Taking a 4-iteration wavelet packet analysis as an example, it can be clearly seen that the essence of wavelet packet transform is to iterate the original signal for num times. iter The binary decomposition of iter When the decomposition is performed, the frequency domain is divided into nodes, each node corresponds to a frequency band, and each frequency band contains the frequency domain feature information of the original signal within this range. The wavelet packet analysis process is shown in the following formula:

[0105]

[0106] In the formula, p (i，j) is the frequency domain coefficient of the jth node in the ith iteration, H and G are a set of wavelet basis functions. The selection of wavelet basis functions needs to be determined according to the specific implementation case.

[0107] The result of wavelet packet analysis is the coefficient vector p of all nodes in the last iteration. (last，j) , defined as follows:

[0108] p (last，j) =[fre1,fre2,...,fre len ]

[0109] Among them, last is the last iteration, j is the jth node, fre is the frequency feature, and len is the number of frequency features. After obtaining the wavelet packet analysis results, they need to be stacked, and the number of stacking layers is set to fold. The specific process is as follows:

[0110]

[0111] in, represents the frequency characteristics of all nodes after the last iteration, Indicates that this is a length of num res , a two-dimensional matrix with a width of len; represents the folded frequency feature matrix, Represents a height of fold and length of (num res / fold), a three-dimensional matrix with a width of len. → indicates that this is a folding process, which is implemented by calling the open source library numpy.reshape function. The specific process is

[0112] Step 4: Pipeline damage identification based on convolutional neural network

[0113] Identification and location of damage to the flow pipeline of an offshore platform based on a convolutional neural network. After wavelet packet analysis extracts a series of damage features from the real sensor strain modal data of the flow pipeline of an offshore platform, a mapping relationship between damage features and damage positions is established through a convolutional neural network. The present invention proposes a coefficient stacking technique here to stack a two-dimensional matrix composed of a series of damage features into a three-dimensional space, and the height of the stack is the number of channels input to the convolutional neural network. The convolutional neural network is divided into an input layer, a hidden layer, and an output layer. The output layer is a fully connected layer and a LogSoftmax layer, which are used to obtain the probability of damage at each position of the pipeline. The hidden layer structure is three layers, and each layer structure is a convolution layer-BN layer-activation layer-pooling layer-regularization, the convolution kernel size is (5, 5), the pooling layer size is 2, the activation layer function is the ReLU() function, the regularization ratio is 0.1, the optimizer uses the RMSprop function, the damage function uses the cross entropy CrossEntrop function, the initial learning rate is set to 0.05, and the learning rate decreases by 75% every 5 iterations, and the maximum number of iterations is set to 200.

[0114] Step 5: Optimize sensor placement based on the improved whale optimization algorithm

[0115] Based on the previous steps, the damage identification and location of the offshore platform flow pipeline of a certain set of sensor layout schemes can be obtained. Based on this, the present invention improves the whale optimization algorithm to optimize the sensor layout location. The damage identification and location accuracy is used as the objective function, and the number of sensors and the layout spacing are used as optimization variables, as shown below:

[0116] var = [num, inter]

[0117] Among them, var is the optimization variable, num is the number of sensors deployed, and inter is the spacing between sensors. After that, the whale algorithm is used to find the optimization variable that makes the damage identification and positioning accuracy the highest. First, initialize the global variables of the whale algorithm, such as the number of whales, variable interval, maximum number of iterations, parameters b and l. After entering the iteration, the specific evolutionary strategy executed by the current whale individual is determined according to a uniform random number r. When r≤0.5, the impact search strategy is executed, and when r>0.5, the spiral search strategy is executed. In the direct search strategy, the whale individual to be directly searched is determined according to the following formula.

[0118]

[0119]

[0120] in, is a random number ranging from 0 to 1. It is worth noting that and are all vectors, which means that different dimensions of the current whale individual can have different search objects. When , a whale individual is randomly selected from the population for search. The specific formula is as follows:

[0121]

[0122]

[0123]

[0124] in, is a randomly selected whale individual. When , the current global best individual is selected as the object to perform a direct search strategy. The specific formula is as follows:

[0125]

[0126]

[0127]

[0128] in, is the global best individual. The spiral search strategy is as follows:

[0129]

[0130]

[0131] After the current whale individual has executed the evolutionary strategy, it is necessary to check whether its offspring have crossed the boundary and repair the out-of-bounds variables. The present invention proposes a new out-of-bounds repair method to better meet the optimization of sensor deployment in the flow pipeline of the offshore platform. The data format of the number of sensor deployments and the deployment spacing are not the same. The number of sensor deployments is an integer, and the sensor deployment spacing is a floating point number. Therefore, the two need to be separated and targeted out-of-bounds repair methods are used respectively. When the number of sensor deployments exceeds the variable interval, the following method is used to repair the out-of-bounds number of sensor deployments: First, the specific repair strategy executed by the current whale individual is determined according to a random decimal p. When p≤0.5, the number of sensors deployed by the current best individual is selected to replace the current number of out-of-bounds sensors. When p>0.5, the number of sensors deployed by the global best individual is selected to replace the current number of out-of-bounds sensors. After that, the number of sensors deployed by a whale individual in the population is randomly selected. When the number of sensors deployed by the random individual is equal to the number of sensors deployed by the global best individual, the number of sensors deployed by the offspring of the current iteration individual is randomly perturbed. The specific formula is as follows:

[0132]

[0133]

[0134] else:num=num+round(normal(0,1))

[0135] in, is the best individual in the world, is the current best individual, round() is the rounding function, and normal(0,1) is the standard normal distribution. When the sensor spacing exceeds the variable interval, it must be considered that the sensor spacing is related to the number of sensors. It cannot be simply replaced with the sensor spacing of the global or local best individual or the sensor spacing of the global or local best individual is disturbed. Therefore, repairing the sensor spacing that exceeds the interval is divided into two parts: one is the sensor spacing replacement object, and the other is the sensor spacing replacement method. The determination of the sensor spacing replacement object is divided into three steps. First, according to the fitness value from small to large, find individuals with the same number of sensors in the historical global best individuals. If there are the same individuals, then the historical global best individual is the replacement object of the sensor spacing; if not, then in the current population according to the fitness value from small to large, find individuals with the same number of sensors. If there are, then the individual is determined as the replacement object of the sensor spacing. It should be noted that it cannot be the same as the current iteration individual; if not, there is no replacement object. The sensor spacing replacement method changes according to whether there is a replacement object. When the sensor spacing replacement object is the best individual in history, the sensor spacing to be repaired is replaced by the sensor spacing of the current repair object; when the sensor spacing replacement object is the current group individual, the sensor spacing to be repaired is replaced by the sensor spacing of the replacement object under random disturbance, that is, adding a normally distributed random number; when there is no sensor spacing replacement object, a random number within a variable interval is used for repair.

[0136] After each whale individual executes the evolutionary strategy, it needs to execute the global elitist strategy. Unlike the previous elitist strategy, in the global elitist strategy, the offspring of the current iteration individual are not only compared with their parents, but also need to be compared with the best individual in history and the best individual in the current group, so as to keep the best individual used in each whale iteration real-time. Through this method, when the number of groups becomes larger, the group convergence speed can be effectively accelerated.

[0137] When the iteration is finished, the optimized variables of the best individual in history are used to deploy sensors.

[0138] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Any features such as structure / module name, control mode, algorithm, process or composition ratio not clearly described in this technical solution shall be regarded as common technical features disclosed in the prior art.

[0139] Example 1

[0140] The present invention provides a specific implementation case of a method for locating damage in a flow transmission pipeline of an offshore platform and optimizing sensor layout.

[0141] Step 1: Simulate the flow pipeline of the offshore platform according to the actual engineering environment. The pipeline simulation model is a steel pipe with a length of 1000mm, an outer diameter of 60mm, a wall thickness of 3mm, and a material of structural steel. The starting position of the damage simulation is 30mm, and the ending position is 970mm. The damage length is 3mm, the width is 1mm, and the depth is 2mm. Starting from 30mm of the pipeline, the damage simulation is set at intervals of 3.5mm. A total of 278 groups of valid data were collected. The damage starting position is used as the label, and the strain modal data is used as the training set.

[0142] Step 2:

[0143] (1) According to formula (1), noise simulation is performed on the strain modal data obtained in step 1 to obtain noise strain modal data.

[0144] (2) The splrep function of the interpolate class in the Python.Scipy library is used to perform cubic spline interpolation on the noise strain modal data to obtain the interpolated strain modal data.

[0145] (3) According to formula (2), the interpolated strain modal data is normalized to obtain normalized strain modal data.

[0146] Step 3: For the normalized strain modal data obtained in step 2, use the Pywt.WaveletPacket function in Python to implement wavelet packet analysis, select the db1 series function as the wavelet packet basis function, decompose the number of layers into 9 layers, and obtain 512 frequency coefficients. The specific command is pywt.WaveletPacket (mode = 'symmetric', wavelet = 'db1', maxlevel = 9). The result of the wavelet packet analysis is a two-dimensional matrix of (512, 128), that is, the coefficient characteristic length of each node is 128. Use the wavelet packet stacking technique to convert the two-dimensional matrix into a three-dimensional matrix of (4, 128, 128).

[0147] Step 4: Process the 278 sets of data obtained in step 1 through steps 2 and 3. Input the processed data into the convolutional neural network. The structural parameters of the convolutional neural network in the specific implementation case are as follows:

[0148] Table 1 Convolutional neural network structure parameters

[0149]

[0150] The training set accounts for 90% of the total number of samples, the test set accounts for 10% of the total number of samples, the number of iterations is 1000, the loss function is the cross entropy function, the convolutional neural network is built based on Pytorch, the operating environment is Python3.8, Numpy1.23, scipy1.9.3, Pytorch1.12, cuda11.3, the hardware environment is: Intel(R)Xeon(R)Silver4210R CPU@2.40GHz, Nvidia 3090, 128GB DDR4 2400, batch_size is 16, and the learning rate is 0.05. The recognition results are shown in the following table:

[0151] Table 2 Damage location identification results

[0152]

[0153] As can be seen from the table, for the 20 test samples, the convolutional neural network has an accuracy rate of 85% in identifying the damage location. For the sample conditions with incorrect identification, the highest error is 5.9%, indicating that the identification result for a single damage location is relatively accurate.

[0154] Step 5: Set the global parameters of the improved whale optimization algorithm: the number of whales is set to 50, the number of sensors is in the range of [50, 200], the sensor layout interval is in the range of [5, 20], b is set to 1, l is set to 1, the maximum number of iterations is set to 100, and the fitness is the damage location recognition accuracy. After the improved whale optimization algorithm is iterated, the best individual in history is used as the sensor layout setting.

[0155] After the improved whale optimization algorithm is iterated, the best individual in history is used as the sensor deployment parameter. The optimization results are shown in Table 3:

[0156] Table 3 Damage location optimization identification results

[0157]

[0158] Table 4 Sensor layout

[0159]

[0160] After the optimization of the sensor layout, for 20 test samples, the convolutional neural network has an accuracy rate of 95% in identifying the damage location. For the sample conditions with incorrect identification, the maximum error in identification is 0.2%, indicating that the improved whale optimization algorithm has a very obvious effect on the optimization of sensor layout.

[0161] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for optimizing the placement of sensors in a flow pipeline on an offshore platform, characterized in that: The following steps are involved: S1: Obtain parameter information of the flow pipeline of the offshore platform, build a finite element simulation model, build pipeline damage information in the finite element simulation model, set the data acquisition starting point, obtain strain modal data on the axial path of the pipeline, obtain strain modal data at each sensor position based on the original sensor layout plan, and use the pipeline damage position as a label for the data to construct a training data set; S2: performing noise simulation, data interpolation, and linear normalization on the training data set to obtain a preprocessed continuous sequence data set; S3: performing time-frequency analysis on the continuous sequence data set using wavelet packet analysis to obtain a damage feature information set of the strain modal data of the sensor corresponding to the flow pipeline of the offshore platform; S4: Based on the damage feature information set, establishing a damage feature-damage location mapping relationship set through a convolutional neural network; S5: Based on the damage feature-damage location mapping relationship set and using the improved whale optimization algorithm, the sensor deployment position is optimized and adjusted. The process of optimizing the sensor placement includes: The damage identification and positioning accuracy is used as the objective function, and the number and spacing of sensors are used as optimization variables, as shown below: in, To optimize the variables, is the number of sensors deployed, Set up spacing for sensors; Afterwards, the whale algorithm is used to find the optimization variables that maximize the accuracy of damage identification and location, including the following steps: Initialize the global variables of the whale algorithm, and after entering the iteration, according to a uniform random number Determines the specific evolutionary strategy that the current whale individual executes. When , the impact search strategy is executed. When , the spiral search strategy is executed; After the current whale individual has executed the evolutionary strategy, it checks whether its offspring have crossed the boundary and repairs the variables that have crossed the boundary; When the number of sensors deployed exceeds the variable range, the following process is used to repair the number of sensors that have exceeded the range: First, according to a random decimal Determines the specific repair strategy implemented by the current whale individual. When , the number of sensors deployed by the current best individual is selected to replace the number of sensors deployed that are currently out of bounds. , select the number of sensors deployed by the global best individual to replace the current number of out-of-bounds sensors; After that, the number of sensors deployed by a whale individual in the population is randomly selected. When the number of sensors deployed by the random individual is equal to the number of sensors deployed by the global best individual, the number of sensors deployed by the offspring of the current iteration individual is randomly perturbed. The specific formula is as follows: in, is the best individual in the world, is the best individual at present, is the rounding function, is a standard normal distribution; Finally, after each whale individual executes the evolutionary strategy, it executes the global elitism strategy, comparing the offspring of the current iteration individual with their parents, and with the best individual in history and the best individual in the current group, so as to keep the best individual used in each whale iteration instant, thereby accelerating the convergence speed of the group; When the iteration is finished, the optimization variables of the best individual in history are used as the optimal sensor deployment parameters.

2. The method for optimizing the placement of sensors in a flow pipeline of an offshore platform according to claim 1, characterized in that: In S1, the parameter information of the offshore platform fluid delivery pipeline includes length, outer diameter, inner diameter, fluid, material, and surrounding working environment parameters.

3. The method for optimizing the placement of sensors in a flow pipeline of an offshore platform according to claim 2, characterized in that: In S1, the process of constructing pipeline damage information in the finite element simulation model includes: Pipeline damage construction: Pipeline damage with a length of L mm, a width of W mm, and a depth of H mm is obtained by rotary cutting on the inner wall of the pipeline simulation model; Obtaining the position mark of the damage: Obtain the distance X mm between one end point of the damage and the end of the pipe corresponding to the end point, and use the X mm as the position mark of the damage.

4. The method for optimizing the placement of sensors in a flow pipeline of an offshore platform according to claim 3, characterized in that: In S1, the process of building a training dataset includes: The Y mm at one end of the offshore platform flow pipeline simulation model is set as the data collection initial point, and the L-30 mm at one end of the offshore platform flow pipeline simulation model is set as the data collection end point. The strain modal data on the pipeline axial path is obtained by finite element calculation. Then, according to the sensor layout plan, the strain modal data at each sensor layout position is taken out, which is a single training data, and the pipeline damage position Xmm is used as the label of the data.

5. The method for optimizing the placement of sensors in a flow pipeline of an offshore platform according to claim 1, characterized in that: In S2, during the noise simulation, white noise is added to the simulated strain modal data to simulate the generation of noise in actual engineering; Among them, random numbers based on standard normal distribution are used to generate white noise: in, is the standard normal distribution, For input data, is the data after adding noise; In the data interpolation process, three data interpolations are performed based on the original data features; In the linear normalization process, the range of each order strain mode is mapped to [0, 1], and the dimension of the real sensor strain modal data of any order is unified. The normalization formula is as follows: in, is the minimum function, is the maximum value function.

6. The method for optimizing the placement of sensors in a flow pipeline of an offshore platform according to claim 1, characterized in that: In S3, the wavelet packet analysis process includes: The number of iterations for each data in the continuous sequence data set is The binary decomposition of When the decomposition is performed, the frequency domain is divided into nodes, each node corresponds to a frequency band, and each frequency band contains the frequency domain feature information of the original signal within this range. The wavelet packet analysis process is shown in the following formula: In the formula, For the The iteration The frequency domain coefficients of nodes, and is a set of wavelet basis functions; The result of the wavelet packet analysis is the coefficient vector of all nodes in the last iteration , defined as follows: in, For the last iteration, For the nodes, is the frequency characteristic, is the number of frequency features; Coefficient vector based on all nodes A damage feature information set of the strain modal data of sensors corresponding to the flow pipeline of the offshore platform is constructed.

7. The method for optimizing the placement of sensors in a flow pipeline of an offshore platform according to claim 6, characterized in that: In S4, in the process of establishing the damage feature-damage location mapping relationship set through the convolutional neural network: The coefficient stacking method is used to stack a series of two-dimensional matrices composed of damage features into a three-dimensional space. The height of the stacking is the number of channels input into the convolutional neural network. During the stacking process, set the number of stacking layers to , the specific process is as follows: in, represents the frequency characteristics of all nodes after the last iteration in wavelet packet analysis, Represented as a length , width is A two-dimensional matrix, represents the folded frequency feature matrix, Indicates a high , the length is the quotient of numres divided by fold, and the width is The three-dimensional matrix, in It is represented as a folding process, which is implemented by calling the open source library numpy.reshape function. The specific process is numpy.reshape(( )); The convolutional neural network is divided into an input layer, a hidden layer and an output layer; The output layer is a fully connected layer and a LogSoftmax layer, which are used to obtain the probability of damage at each position of the pipeline. The hidden layer structure is three layers, and each layer is a convolution layer-BN layer-activation layer-pooling layer-regularization layer. The convolution kernel size is (5,5), the pooling layer size is 2, the activation layer function is the ReLU() function, the regularization ratio is 0.1, the optimizer uses the RMSprop function, and the damage function uses the cross entropy CrossEntrop function. The initial learning rate is set to 0.05, and the learning rate decreases by 75% every 5 iterations. The maximum number of iterations is set to 200. The damage feature-damage location mapping relationship set is established through the convolutional neural network.

8. The method for optimizing the placement of sensors in a flow pipeline of an offshore platform according to claim 1, characterized in that: In S5, in the direct search strategy, the whale individuals to be directly searched are determined according to the following formula; in, is a random number ranging from 0 to 1; when When , a whale individual is randomly selected from the population for search. The specific formula is as follows: in, For randomly selected whale individuals; when When , the current global best individual is selected as the object to execute the spiral search strategy. The specific formula is as follows: in, is the best individual in the world; The spiral search strategy is as follows: 。 9. A sensor layout optimization system for a flow pipeline on an offshore platform, characterized in that: include: The simulation and data construction acquisition module obtains parameter information of the offshore platform flow pipeline, constructs a finite element simulation model, constructs pipeline damage information in the finite element simulation model, sets the data acquisition starting point, obtains strain modal data on the pipeline axial path, obtains strain modal data at each sensor position based on the original sensor layout plan, and uses the pipeline damage position as a label for the data to construct a training data set; A data post-processing module performs noise simulation, data interpolation, and linear normalization on the training data set to obtain a pre-processed continuous sequence data set; A feature analysis module performs time-frequency analysis on the continuous sequence data set using wavelet packet analysis to obtain a damage feature information set of the strain modal data of the sensor corresponding to the flow pipeline of the offshore platform; A convolution calculation module, based on the damage feature information set, establishes a damage feature-damage location mapping relationship set through a convolutional neural network; An optimization calculation module optimizes and adjusts the sensor layout position based on the damage feature-damage position mapping relationship set and adopts an improved whale optimization algorithm; The process of optimizing the sensor placement includes: The damage identification and positioning accuracy is used as the objective function, and the number and spacing of sensors are used as optimization variables, as shown below: in, To optimize the variables, is the number of sensors deployed, Set spacing for sensors; Afterwards, the whale algorithm is used to find the optimization variables that maximize the accuracy of damage identification and location, including the following steps: Initialize the global variables of the whale algorithm, and after entering the iteration, according to a uniform random number Determines the specific evolutionary strategy that the current whale individual executes. When , the impact search strategy is executed. When , the spiral search strategy is executed; After the current whale individual has executed the evolutionary strategy, it checks whether its offspring have crossed the boundary and repairs the variables that have crossed the boundary; When the number of sensors deployed exceeds the variable range, the following process is used to repair the number of sensors that have exceeded the range: First, according to a random decimal Determines the specific repair strategy implemented by the current whale individual. When , the number of sensors deployed by the current best individual is selected to replace the number of sensors deployed that are currently out of bounds. , select the number of sensors deployed by the global best individual to replace the current number of out-of-bounds sensors; After that, the number of sensors deployed by a whale individual in the population is randomly selected. When the number of sensors deployed by the random individual is equal to the number of sensors deployed by the global best individual, the number of sensors deployed by the offspring of the current iteration individual is randomly perturbed. The specific formula is as follows: in, is the best individual in the world, is the best individual at present, is the rounding function, is a standard normal distribution; Finally, after each whale individual executes the evolutionary strategy, it executes the global elitism strategy, comparing the offspring of the current iteration individual with their parents, and with the best individual in history and the best individual in the current group, so as to keep the best individual used in each whale iteration instant, thereby accelerating the convergence speed of the group; When the iteration is finished, the optimization variables of the best individual in history are used as the optimal sensor deployment parameters.

10. A storage medium containing computer executable instructions, characterized in that: The computer executable instructions, when executed by a computer processor, are used to execute the method for optimizing the deployment of sensors in a flow pipeline of an offshore platform as claimed in any one of claims 1 to 8.

Images