Pipeline leakage identification method based on improved convolutional neural network

Abstract

The present application relates to the field of offshore platform pipeline leakage identification, and discloses a pipeline leakage identification method based on an improved convolutional neural network, first, a pipeline leakage model is constructed in the laboratory, leakage signals of various working conditions are collected, real background noise is obtained at the positions of 19m and 29m of the 'Lianwan 3-1' offshore platform, and the real background noise is injected into the laboratory leakage signals to simulate the real working environment, finally, the traditional one-dimensional convolutional neural network is improved, the core of which is to introduce residual connection to increase the stability of training, introduce maximum pooling and average pooling to reduce the feature map size and reduce the calculation amount, use the Dropout regularization technology to prevent model overfitting, and use the adaptive moment estimation optimization algorithm to adjust the network weight, then the convolutional neural network structure is adaptively designed, the method realizes high-accuracy pipeline leakage identification and has good robustness, and solves the problems of serious dependence on manual intervention and complex identification process of the traditional machine learning method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of offshore platform pipeline leakage identification, in particular to a pipeline leakage identification method based on an improved convolutional neural network. BACKGROUND

[0002] Exploration and development of marine oil and gas resources play a vital role in the national energy strategy and are crucial to alleviating energy supply pressure. With the increasing service life of marine oil and gas platform pipeline systems, their safety problems are becoming increasingly serious. In the complex marine environment, both natural gas pipelines and oil processing facilities will vibrate due to the presence of various vibration sources. These long-term cumulative vibration effects can cause fatigue and even damage to the pipeline structure. If not detected and prevented in time, a leakage accident may pose a serious safety threat to the marine oil and gas platform and even cause catastrophic consequences such as explosions, resulting in huge economic losses and social impact. Therefore, the leakage protection of offshore oil and gas pipelines has become an important concern in the industry. However, traditional theoretical analysis methods are difficult to operate in many practical situations and cannot fully ensure the safety and reliability of equipment. For this reason, acoustic emission technology has been gradually introduced and widely applied in real-time monitoring and fault diagnosis. Acoustic emission detection technology is a dynamic, real-time and non-destructive evaluation method that can accurately detect the health status of the overall structure and achieve early warning of potential damage by capturing the acoustic signals spontaneously generated by the internal structure of the material under stress or environmental changes. This technology does not require an external excitation source and has the advantages of efficient defect identification capability, comprehensive evaluation performance and wide application adaptability.

[0003] However, in order to accurately identify pipeline leakage using acoustic emission technology, the leakage signal of the pipeline needs to be collected first. However, in actual operation, it is difficult to obtain high-quality leakage signals. Especially in the complex working environment of offshore platforms, strong noise interference makes signal collection even more difficult. For example, mechanical noise generated by pipeline vibration, sea wave and wind noise, etc. all interfere with signal collection, further increasing the challenge of obtaining high-quality acoustic signals.

[0004] For the collected acoustic emission signals, machine learning methods are one of the commonly used methods, but traditional machine learning methods face some technical challenges. For example, although methods such as empirical mode decomposition (EMD), local mean decomposition (LMD), and variational mode decomposition (VMD) achieve high recognition accuracy, they result in a four-stage process in feature extraction: preprocessing, decomposition, feature extraction, and classification. The processing of this process is complex, greatly increasing the computational burden of the system.

[0005] Another major limitation is the strong reliance on empirical methods in feature engineering, especially in feature selection, which heavily depends on human intervention. This expert-driven process risks information loss and is often insufficient to capture subtle fault features in nonlinear systems. The challenge becomes even more severe when encountering unknown fault types, as predefined feature extraction strategies often lack the flexibility and generality required for robust performance. Moreover, feature selection and dimensionality reduction techniques tend to be inefficient when dealing with high-dimensional datasets, especially those involving large sample sizes. As the dimensionality of signals and the amount of data increase, these issues become more pronounced, negatively impacting the responsiveness and overall performance of systems in real-world applications.

[0006] Therefore, it is necessary to provide a new technical solution to solve the above problems, and the present application proposes a pipeline leakage identification method based on an improved convolutional neural network. SUMMARY

[0007] The present application aims to provide a pipeline leakage identification method based on an improved convolutional neural network to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purposes, the present application is implemented by the following technical solutions: a pipeline leakage identification method based on an improved convolutional neural network, comprising the following steps:

[0009] S1, a pipeline leakage model is built in a laboratory controlled environment to collect pipeline leakage signals under different working conditions. The experimental model mainly consists of a pressure pipeline, a pressure pump, a pressure gauge, and a series of control valves. The pressure pipeline is used to simulate the working pipeline of an offshore platform, the pressure pump is used to provide pressure to make the fluid flow in the pipeline, the pressure gauge is used to reflect the current pressure in the pipeline, and the control valves are used to control the size of the leakage to simulate different leakage conditions.

[0010] S2, actual background noise signals are collected from 19-meter and 29-meter platforms of "Lianwan 3-1" offshore platform. The reason for choosing the 19-meter platform pipeline is that it has a strong pipeline vibration intensity and typical mechanical vibration noise characteristics, and is prone to fatigue damage. The reason for choosing the 29-meter platform pipeline is that it is in the cooling process after the compressor is pressurized, and the pipeline at this location bears a large pressure, especially the elbow part, which is prone to damage due to stress concentration.

[0011] S3, the actual background noise signals are injected into the pipeline leakage signals in proportion to generate labeled training datasets simulating real working conditions, and the convolutional neural network structure is improved, including:

[0012] a) Introducing residual connection: After the initial convolutional layer, a residual block is connected in series, each residual block contains two convolutional layers, each layer is followed by batch normalization and ReLU activation function, and residual connection is set before the first average pooling layer and after the second average pooling layer;

[0013] b) Combining max pooling and average pooling: Insert an average pooling layer between the residual blocks, and set a max pooling layer after the input layer;

[0014] S4, regularization and optimization algorithm are introduced, dropout layer is added before the fully connected layer, the dropout rate is set, and Adam optimization algorithm is used to dynamically adjust the network weight;

[0015] S5, the convolutional neural network structure is adaptively designed and configured in sequence, which is respectively input layer, initial convolutional layer, max pooling layer, multiple residual blocks, convolutional layer, dropout layer, fully connected layer and output layer, wherein an average pooling layer is connected after each residual block.

[0016] S6, the labeled training data set is divided into training set and test set according to the proportion, the batch size and the training period are set, and the training is carried out, and the confusion matrix and the accuracy-loss curve evaluation pipeline leakage model performance are output.

[0017] Compared with the prior art, the beneficial effects of the present application are:

[0018] 1. The present application solves the problem of difficult acquisition of high-quality pipeline leakage signal to a certain extent. By constructing a pipeline leakage model in a laboratory controlled environment, collecting pipeline leakage signals under various working conditions, collecting actual background noise signals of "Liwang 3-1" offshore platform, and injecting the actual background noise signals into the leakage signals collected in the laboratory to simulate the leakage signals of the pipeline in the real environment, the data set constructed is more close to the real leakage situation, and the problem of difficult acquisition of high-quality leakage signal is solved.

[0019] 2. The present application proposes an improved convolutional neural network structure, which greatly enhances the performance of the model by introducing residual connection, max pooling, average pooling, dropout regularization and Adam optimization algorithm, improves the recognition accuracy and robustness of the model.

[0020] 3. The improved convolutional neural network structure proposed by the present application can directly extract features from raw data without manual feature extraction, solving the problem that traditional machine learning methods rely heavily on manual intervention to select features, resulting in loss of important information, and avoiding the problem that feature selection and dimension reduction technology is low in efficiency when processing high-dimensional data sets. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is the overall flowchart of the present application;

[0022] Figure 2 Sensor installation position map for offshore platform of the present application;

[0023] Figure 3 Spectrum and time-frequency diagram after noise signal injection for the present application

[0024] Figure 4 Improved convolutional neural network structure diagram for the present application;

[0025] Figure 5 Diagnostic flowchart for the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0027] Please refer to the accompanying Figure 1 , Figure 1 It is a whole flowchart of a pipeline leakage identification method based on an improved convolutional neural network according to the present application, which includes the following steps:

[0028] S1, a pipeline leakage model is established in a laboratory controlled environment, the pipeline leakage model includes a pressure pipeline, a pressure pump, a pressure gauge and a control valve, and pipeline leakage signals are collected by an acoustic emission sensor under not less than three leakage aperture conditions;

[0029] S2, a field investigation is carried out on an offshore platform, and actual background noise signals of the pipeline are collected by an acoustic emission sensor at a predetermined height of the offshore platform;

[0030] S3, the actual background noise signals are injected into the pipeline leakage signals in proportion to generate a labeled training data set simulating real working conditions, and the convolutional neural network structure is improved, including:

[0031] a), residual connection is introduced: a residual block is connected in series after the initial convolutional layer, each residual block contains two convolutional layers, each layer is followed by batch normalization and ReLU activation function, and the residual connection is set before the first average pooling layer and after the second average pooling layer;

[0032] b), maximum pooling and average pooling are combined: an average pooling layer is inserted between the residual blocks, and a maximum pooling layer is set after the input layer;

[0033] S4, introduce regularization and optimization algorithm, add Dropout layer before full connection layer, set dropout rate, and use Adam optimization algorithm to dynamically adjust network weight;

[0034] S5, self-adaptive design of convolutional neural network structure and sequential configuration, respectively for input layer, initial convolutional layer, maximum pooling layer, multiple residual blocks, convolutional layer, Dropout layer, full connection layer and output layer, wherein each residual block is connected with an average pooling layer;

[0035] In order to make the improved convolutional neural network can both stable learning basic features and gradually optimize deep parameters in the training process, improve the recognition accuracy of the model to the pipeline leakage signal, the adaptive design in S5 realizes the phased training strategy according to the following details:

[0036] First of all, the overall framework of phased training is clear, according to the network depth and the number of residual blocks, the initial, intermediate and final three training stages are divided, and the weight parameters of different residual blocks are locked or unlocked in each stage to guide the model to learn features from shallow to deep, avoid training fluctuations caused by synchronous parameter update;

[0037] The initial training stage takes the residual block as the frozen unit, keeps the weight parameters of the first two residual blocks locked, only allows the subsequent convolutional layer and full connection layer to participate in parameter update, and the locking proportion is set according to the overall network depth. The first two residual blocks are responsible for extracting the basic features of the signal, such as the low frequency vibration of the leakage signal and the basic frequency spectrum mode of the noise, locking its weight can avoid the random fluctuation of the initial training to destroy the basic feature learning, the locking proportion is set according to the overall network depth, for example, when the network contains 5 residual blocks, the first 2 layers are locked, the locking proportion is 40%, to ensure the stability of the basic feature extraction module, at this time, the subsequent convolutional layer (used for refining features) and full connection layer (used for classification mapping) can dynamically adjust the parameters according to the input data, so that the basic features and high-level features are preliminarily matched, laying a foundation for the intermediate training;

[0038] The middle training stage adopts a dynamic unfreezing mechanism. After completing a preset number of training cycles, the weight locking state of a residual block is released. The unfreezing order is adjusted according to the network depth. The preset training cycle is set according to the data complexity, for example, one residual block is unfreezed every 10 cycles. First, the third layer residual block (responsible for extracting medium complexity features, such as the mixed mode of leakage signals and noise) is unfreezed, so that its weights can be updated with training and optimized with the parameters of the subsequent layers that have been trained. After completing the next 10 cycles, the fourth layer residual block is unfreezed, which is responsible for extracting high complexity features, such as the difference between signals of different leakage apertures. In this way, the logic of dynamic unfreezing is that the basic features of shallow residual blocks have been preliminarily stabilized, and gradually unlocking deep residual blocks allows the model to adjust the extraction method of medium-level features when learning high-level features, achieving precise matching between feature levels and avoiding training chaos caused by unlocking all parameters at once.

[0039] In the final training stage, all network parameters are opened for joint adjustment. The learning rate is adjusted according to the initial value, and the adjustment round is set according to the total training period. When all residual blocks are unlocked, the model enters the global optimization stage. At this time, the learning rate needs to be reduced, such as reducing the initial learning rate to 1 / 10, so that the parameter fine-tuning is more precise, and the large step update does not destroy the formed feature mapping relationship. The adjustment round is set to 20% of the total training period, for example, if the total training period is 100, the learning rate is adjusted once at the 80th, 90th, and 100th rounds, and gradually converges to the optimal solution. The purpose of joint adjustment is to form a whole synergy of parameters from basic feature extraction to classification output, ensuring that the model's recognition ability for complex leakage signals, such as small aperture leakage in strong noise, reaches the optimal.

[0040] Through this phased training strategy, the model first stabilizes the learning of basic features, then gradually optimizes deep parameters, and finally globally fine-tunes, ensuring the stability of the training process and improving the model's adaptability to complex working conditions, so that the improved convolutional neural network can more accurately identify pipeline leakage features from noisy signals.

[0041] In order to make the residual block more flexible to capture the multi-scale features of the pipeline leakage signal, while considering the computational efficiency and the pertinence of feature extraction, the inner convolutional layer of the residual block adopts a mixed kernel size configuration, which specifically includes:

[0042] First, design an adaptive convolution kernel size for the first convolutional layer. The convolution kernel size of the first convolutional layer is determined by the formula , where represents the length dimension of the current residual block input feature vector, The core logic is to adapt the size of the convolution kernel to the length of the input feature, so as to more accurately cover the basic mode of the feature, such as the periodic fluctuation of the leakage signal and the continuous frequency band of the noise. For example, when the input feature vector length is longer, the convolution kernel size is increased accordingly to ensure that the long-period characteristics of the signal can be captured. When the input feature vector is shorter, the convolution kernel size is reduced to focus on local detailed features. Through this adaptive design, the first convolution layer can preliminarily filter out the key basic features related to the leakage signal.

[0043] The second convolution layer performs a depth separable convolution operation with a fixed kernel size, and introduces a fixed dilated convolution factor. The depth separable convolution separates the standard convolution into a depth convolution and a point-by-point convolution. The input channel is first applied to the convolution kernel (fixed size), and then the channel information is integrated through 1x1 convolution. This greatly reduces the computational complexity while ensuring the feature extraction capability, and is suitable for processing high-dimensional data mixed with pipeline leakage signals and noise. The introduction of the dilated convolution mechanism is to expand the receptive field. The fixed dilated convolution factor enables the convolution kernel to cover a larger range of feature regions without increasing the number of parameters, thereby capturing long-distance associated features in the signal, such as the propagation and attenuation pattern of the leakage signal in the pipeline, and improving the recognition ability of complex signal patterns.

[0044] Finally, a feature map channel number matching device is set in the feedforward connection. When the input and output channel numbers do not match, a convolution kernel is used for channel alignment. The activation function uses a parameterized linear rectifier unit. Since the number of channels in the convolution layer in the residual block may change with the feature extraction requirement, such as increasing from 16 channels to 32 channels, the input and output channel numbers may differ. In this case, a 1x1 convolution kernel is used to adjust the channels of the input feature map, such as mapping 16 channels to 32 channels, to ensure that the residual connection is consistent in dimension when the input feature and the output feature are added together. The parameterized linear rectifier unit is used as the activation function, which retains the non-linear characteristics of the traditional linear rectifier unit while allowing negative features to be preserved through learnable parameters. This avoids the loss of potential leakage signal details due to feature truncation, such as weak leakage with low amplitude. The output features of the residual block are more comprehensive. Through this hybrid kernel scale configuration, the first convolution layer adaptively captures the basic features, and the second convolution layer efficiently extracts complex associated features. Combined with the channel matching mechanism and the optimized activation function, the residual block can flexibly adapt to different input features and accurately extract multi-scale features of the pipeline leakage signal, providing more reliable feature input for the subsequent classification layer.

[0045] S6, proportionally divide the labeled training data set into a training set and a test set, set the batch size and training period for training, and output the confusion matrix and accuracy-loss curve to evaluate the performance of the pipeline leakage model.

[0046] To further improve the adaptability of the pipeline leakage identification model to complex working conditions and reduce misjudgment caused by differences between training data and real scenes, the following post-processing steps are included after the training of S6 is completed:

[0047] S6.1, based on the confusion matrix and the accuracy-loss curve, extract the uncertain samples in the model prediction, the uncertain samples are the leakage signal classification results with a confidence lower than a preset threshold, including the number of correctly classified and misclassified samples; the accuracy-loss curve reflects the convergence state of the model in the training process, if the curve fluctuates greatly or the accuracy stagnates at a low level, it means that there are samples that are difficult to accurately identify. The uncertain sample is defined as the leakage signal classification result with a confidence lower than a preset threshold, and the preset threshold is set according to the accuracy requirement of the model application scene, for example, set to 0.7, that is, the sample with a confidence lower than 70% in the prediction result is marked as uncertain, for example, a sample with a 3mm aperture leakage is actually determined as a 5mm aperture leakage with a confidence of 0.6, this sample is extracted as an uncertain sample, and these samples are the key breakthrough for improving the performance of the model;

[0048] Then perform S6.2, time-frequency graph analysis on the uncertain samples, specifically including:

[0049] According to the spectrum graph and time-frequency graph processing method, a time-frequency graph is generated by applying short-time Fourier transform, the energy distribution characteristics in the time-frequency graph are extracted, and compared with normal samples to identify abnormal patterns, the time-frequency graph can simultaneously show the time and frequency characteristics of the signal, and is suitable for analyzing the spectral distribution difference of pipeline leakage signals and noises at different times, in specific operation, the uncertain samples are processed by short-time Fourier transform, the one-dimensional acoustic emission signal is converted into a two-dimensional time-frequency graph, the horizontal axis of the graph is time, the vertical axis is frequency, and the color depth represents the signal energy. Extract the energy distribution characteristics in the time-frequency graph, such as the energy peak value of the leakage signal in a specific frequency band, the frequency range and duration of noise interference, and compare them with normal samples, such as the typical time-frequency graph of 1mm, 3mm and 5mm aperture leakage, to identify abnormal patterns, for example, in the time-frequency graph of the uncertain sample, the energy distribution of a certain frequency band deviates greatly from the normal sample, or the proportion of noise energy in the high frequency band abnormally increases, these abnormal patterns are often the core reason for the model misjudgment;

[0050] S6.3, feedback the abnormal patterns to the S3 step, dynamically adjust the noise injection ratio, wherein the noise ratio adjustment formula is:

[0051]

[0052] wherein, is the original noise ratio, For the new noise ratio, derive from the confusion matrix, for the preset value, retrain the pipeline leakage model using the adjusted labeled training dataset, iteratively optimize the confusion matrix, and the abnormal pattern reflects the deviation of the noise simulation in the current training data from the real scene. For example, if the abnormal pattern shows that the high-frequency noise interference is insufficient, the noise injection ratio needs to be increased; if the noise type does not match the actual one, the selection weight of the noise sample needs to be adjusted. When adjusting, according to the misclassification rate derived from the confusion matrix, that is, the proportion of uncertain samples in the total samples and the preset target accuracy, the new noise injection ratio is determined, for example, the original noise injection ratio is 30%, the misclassification rate is 10%, and the target accuracy is 95%. Then, increase the noise injection ratio in proportion, so that the training data is closer to the real noise environment. Re-input the adjusted labeled training dataset into the improved convolutional neural network, repeat the training process, and gradually reduce the misclassification samples in the confusion matrix through multiple iterations, until the recognition accuracy of the model for uncertain samples reaches the expected target.

[0053] The overall idea of the method is to first obtain leakage signals under various leakage conditions in a laboratory controlled environment, then collect the actual background noise signals of the pipeline, inject the actual background noise signals into the leakage signals, and construct a dataset simulating the working conditions in a real environment. The traditional one-dimensional convolutional neural network is improved by introducing residual connection, max pooling and average pooling, Dropout regularization technology and Adam optimization algorithm to enhance the performance of the convolutional neural network. Then the improved convolutional neural network is used to train and learn the dataset, and the accuracy-loss rate curve of the test set and the training set is used to reflect the recognition accuracy of the model, so that the model can continuously adapt to the complex noise environment in the real scene, gradually improve the recognition accuracy of different leakage conditions, and finally form a pipeline leakage recognition model with stronger robustness.

[0054] The pressure pipeline meets the high-pressure fluid conveying standard, uses seamless steel pipe material, the acoustic emission sensor is installed at the specified position of the experimental section, is fixed using coupling medium, and the pipeline leakage model in S1 is divided into two main areas, namely the experimental section with a preset length and the buffer section. The experimental section is uniformly distributed with high-pressure welded needle valves for simulating and controlling leakage conditions, and the buffer section is connected with a pressurizing device through a flexible rubber pipeline. The pressurizing device selects a Z4DSY type electric pressure self-control pressure test pump, which is as follows:

[0055] To simulate the leakage of the pipeline, a pipeline leakage model was built under laboratory controlled conditions. The model mainly consists of a pressure pipeline, a pressure pump, a pressure gauge and a series of control valves. The selected pipeline model strictly follows the GB / T 8163-2018 standard, uses seamless steel pipe material, and is suitable for high-pressure fluid transportation. The pipeline diameter is 105 mm, the wall thickness is 5 mm, and the total length is 6.2 m. The system is divided into two key areas. First, the experimental section: 5 meters long, evenly distributed with multiple high-pressure welded needle valves for simulating and controlling leakage conditions; then the buffer section: 1.2 meters long, connected to the pressure device through a flexible rubber pipeline. The pressure device uses a Z4DSY type electric pressure self-control test pressure pump with excellent performance. Two R15 sensors are installed at both ends of the pipeline 10 cm away from the boundary of the experimental section, and industrial-grade white vaseline is used as the coupling medium between the pipeline surface and the sensor, then fixed with tape.

[0056] The data acquisition device uses a sound emission detection device produced by the United States Physical Acoustics Company (PAC) with a model number of PCI-2. The device has a wide and precise sampling frequency range (1 kHz to 3 MHz) and built-in efficient data storage function, which can store acoustic emission waveform information on the hard disk at a speed of 10 million sampling points per second. The AEwin software matched with the PCI-2 device is used for operation and control in the data acquisition stage. The sensor used in the experiment is R15 type, with a center frequency set to 150 kHz. In order to further optimize the signal quality and enhance the contrast between signal and noise, this experiment configures 2 / 4 / 6 channel preamplifiers, whose gains can be adjusted to 20 / 40 / 60 dB respectively. The acquisition parameters are set and optimized in advance on the AEwin software platform. Then the lead fracture experiment was carried out to verify the coupling state of the sensor and the pipeline and the stability of signal transmission, and to confirm whether the data acquisition device storage function is normal. The final adjustable gain is set to 40db, and the sampling frequency is set to 1MHz. After debugging, multiple groups of pipeline leakage signals were collected.

[0057] Please refer to the attached Figure 2 , Figure 2 is the position map of the measuring point of the offshore platform in the present application, which shows the pipeline position for collecting the actual background noise.

[0058] The collection position of the actual background noise signal in S2 is located at the 19m and 29m platforms of the offshore platform, the installed acoustic emission sensor type and parameter setting are consistent with the equipment in the laboratory, and are fixed through the coupling medium. The sampling mode of the acoustic emission sensor adopts a long-term timing sampling mode, which is as follows:

[0059] In order to collect the actual background noise signal of the pipeline, a detailed field survey was conducted on the offshore platform, and finally the pipelines of the 19-meter platform and the 29-meter platform were selected as the collection objects. The 19-meter platform has typical mechanical vibration noise characteristics due to the strong pipeline vibration intensity, and is prone to fatigue damage. The 29-meter platform is in the cooling treatment link after the compressor boosting, and the pipeline at this position bears a large pressure, especially the elbow part is prone to damage due to stress concentration. Therefore, these two positions with typical noise characteristics and potential risks are selected as the collection points of the field noise signal.

[0060] Please refer to the attached Figure 3 , Figure 3 In the present application, the frequency spectrum and the time-frequency diagram of the actual background noise injected into the leakage signal are obtained.

[0061] S3 divides the pipeline leakage signal according to the length of a preset number of sampling points to obtain leakage signal samples, synchronously divides the collected actual background noise signal according to the length of the preset number of sampling points to construct noise signal samples, and finally forms a labeled training data set simulating the actual working condition by fusing the leakage signal samples and the noise signal samples, as follows:

[0062] Since the laboratory noise level is relatively low and cannot reflect the actual noise situation of the offshore platform, the measured field noise is fused into the leakage signal collected in the laboratory to verify the acoustic emission pipeline leakage detection method. Since both are continuous signals, the signal power is used to determine the noise percentage, that is:

[0063] Among them is the noise power, is the signal power, , are the root mean square values of the noise and the leakage signal respectively, . Therefore, the leakage signal fused with noise can be obtained by the following formula:

[0064] Among them is the collected leakage signal; is the measured noise signal.

[0065] In order to make samples of leakage signals containing noise, the pipeline leakage signal collected in the laboratory is first uniformly divided according to the length of 1024 sampling points, so as to obtain 900 leakage signal samples, and the actual noise signal collected on the offshore platform is also divided into paragraphs with a length of 1024 sampling points, thereby constructing 900 noise signal samples. Finally, by fusing the leakage signal samples and the noise signal samples, an original data set containing 1800 groups of samples is formed.

[0066] Please refer to the attached Figure 4 , Figure 4 is the improved convolutional neural network structure diagram proposed in the present application.

[0067] a) The residual connection, max-pooling and average-pooling are introduced. The specific reference positions of the residual connection are before the first average-pooling layer, after the second average-pooling layer and before the second average-pooling layer. Each residual block is composed of two convolutional layers, followed by batch normalization and ReLU activation function. This architecture adds the input to the output connection without changing the forward propagation, and adjusts the gradient backpropagation. The introduction of max-pooling and average-pooling is used to reduce the size of the feature map and reduce the amount of calculation. The specific implementation is as follows:

[0068] In the traditional one-dimensional convolutional neural network, the residual connection is introduced to increase the stability of training. The residual connection can be used to solve the degradation problem in the training process of deep neural networks. Mathematically, the output of the residual block can be represented as:

[0069] wherein, is the input of the residual block, represents a nonlinear transformation in the residual block with weights , is the output of the residual block. This structure allows the input to be directly added to the output of the block through a short circuit, thereby helping the gradient to be directly backpropagated to the deeper network without affecting the forward propagation, effectively alleviating the gradient vanishing problem. In this embodiment, by adding residual connections between convolutional layers, the model can maintain the stability and efficiency of training while increasing the depth. This design enables the network to learn more complex and abstract feature representations.

[0070] Max-pooling and average-pooling are introduced to reduce the size of the feature map and reduce the amount of calculation. Max-pooling is a commonly used pooling technique in convolutional neural networks, which is used to reduce the size of the feature map, reduce the amount of calculation, and provide translational invariance to some extent. Max-pooling achieves this by extracting the maximum value of the covered region, which helps to retain the most prominent parts of texture and internal features. Given a feature map region, the operation of max-pooling can be represented as:

[0071] wherein, is the set of feature map regions covered by the pooling window, is the maximum value extracted from .

[0072] Average pooling reduces the size of the feature map by calculating the average value of each small region on the input feature map. Assuming within a region, the operation of average pooling can be represented as:

[0073] wherein, is the output after average pooling, is the value of the position on the input feature map, is the size of the pooling window. Average pooling can make the feature representation more robust to small translations. Even if there is a slight change in the position of the object in the image, the features extracted by average pooling can remain consistent.

[0074] By introducing maximum pooling and average pooling, the dimension and parameter number of the feature map can be reduced while extracting key features, thereby reducing the amount of calculation.

[0075] The Dropout layer in S4 adopts the Dropout regularization technique and sets the dropout rate, which is used to randomly discard neurons in the training process. The remaining neurons continue to participate in the calculation of the convolutional neural network to prevent overfitting of the pipeline leakage model. Finally, the Adam optimization algorithm is used to calculate the adaptive learning rate of each parameter to adjust the weights of the convolutional neural network, as follows:

[0076] The Dropout regularization technique is introduced to prevent network overfitting. In the training process, the Dropout technique randomly discards a portion of the neurons in the network (i.e., sets their output to zero), which can be represented as:

[0077] wherein, is the input vector of a neuron, is a randomly generated 0-1 vector of the same dimension, represents element multiplication. In this way, Dropout forces the network to learn more robust features because it cannot rely on the activation output of any one neuron, thereby improving the generalization ability of the model. In this embodiment, the dropout rate is set to 0.4, i.e., there is a 40% probability that a neuron will be discarded and a 60% probability that it will be retained for network calculation.

[0078] The Adam optimization algorithm is introduced to adjust the network weights. The Adam optimization algorithm is an adaptive learning rate optimization algorithm that calculates the adaptive learning rate of each parameter, which can be represented as:

[0079] wherein, is the parameter, is the learning rate, and respectively, are bias corrections for the first and second moment estimations, is a very small number to prevent division by zero. In this embodiment, the improved one-dimensional convolutional neural network performs better in the offshore platform pipeline leakage identification task through the joint action of the above improvements.

[0080] After that, the convolutional neural network structure is adaptively designed, and a network structure capable of effectively processing one-dimensional marine platform pipeline acoustic emission data is constructed. The main part of the network is composed of multiple core components. First, there is an initial convolutional layer, followed by a series of residual blocks. Each residual block is followed by an average pooling layer. Finally, another convolutional layer is used for feature extraction. The last part of the network is a fully connected layer sequence, which is used to map deep features to output categories.

[0081] Specifically, the initial layer is composed of a one-dimensional convolution (Conv1d) layer, which receives a single-channel input and expands to 16 feature channels through a convolution kernel with a kernel size of 5 and a step size of 1. This layer also includes a batch normalization layer and a ReLU activation function, as well as a max pooling layer with a step size of 4, which is used to reduce the feature dimension and improve computational efficiency.

[0082] The core component is the residual block, which includes two convolutional layers inside each residual block, each followed by a batch normalization and ReLU activation function. These residual blocks are specifically designed to handle cases where the number of input and output channels is different or the step size is not 1. By introducing a conditional downsampling path, the input of the residual block can be added to the output after two convolution operations. This design allows information to flow more efficiently in the network while alleviating the problem of gradient vanishing.

[0083] Between consecutive residual blocks, the dimension of the features is further reduced through an average pooling layer, which helps to reduce computational load and increase the abstraction ability of the network. After passing through three residual blocks and the corresponding pooling layers, the data is further extracted through another convolutional layer, and then a Dropout layer is used to reduce overfitting. At this stage, the output of the previous layer is first flattened to meet the input format requirements of the fully connected layer. The fully connected layer sequence first reduces the 4096-dimensional feature vector to 1024 dimensions, and then maps it to the final output through an output layer, thus completing the entire classification and prediction process.

[0084] Please refer to the accompanying Figure 5 , Figure 5 is the overall flowchart of the method proposed by the present application.

[0085] The improved one-dimensional convolutional neural network is used to identify the established data set, including the following steps:

[0086] (1) Data collection: Collecting the acoustic emission signal data generated during the operation of the pipeline to form a data set as the basis for model training and testing.

[0087] (2) Model design: Designing an improved convolutional neural network structure integrating residual connection, max pooling, average pooling, and introducing Dropout technology and Adam optimization algorithm to enhance the model's feature learning ability, reduce overfitting, and improve the model's generalization performance.

[0088] (3) Model training: Training the CNN model using the labeled data set, optimizing network parameters through forward propagation and back propagation algorithms to minimize errors.

[0089] (4) Model evaluation: Evaluating the model performance using the test set to ensure that the model has good generalization ability and can accurately diagnose unseen data.

[0090] During the model training process, the batch size is set to 128 and a total of 400 cycles are trained. Finally, the confusion matrix, accuracy-loss curve of the training set, and accuracy-loss curve of the test set are obtained as the final output. Analysis shows that the model proposed by the present application has achieved stable convergence on both the training set and the test set. At the end of training, the classification accuracy of the training set and the validation set is more than 99%, while the loss value continues to decrease and stabilizes on the convergence platform. These results highlight the strong convergence behavior and effective learning ability of the model. By analyzing the confusion matrix, it can be concluded that the model correctly classifies all acoustic signals related to normal operating conditions, proving the high reliability of the model in complex acoustic conditions.

[0091] Through the method described in this embodiment, a data set simulating real offshore platform pipeline leakage signals is successfully constructed, and the problems of complex process and serious dependence on manual intervention for feature extraction in traditional machine learning methods, which leads to loss of key information, are solved, providing an important basis for the structural analysis and optimization of offshore platform pipelines.

[0092] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A pipeline leakage recognition method based on an improved convolutional neural network, characterized in that, The method comprises the following steps: S1, establishing a pipeline leakage model in a laboratory controlled environment, the pipeline leakage model comprising a pressure pipeline, a pressure pump, a pressure gauge and a control valve, collecting pipeline leakage signals through an acoustic emission sensor under at least three leakage aperture conditions; S2, conducting an on-site investigation on an offshore platform, collecting pipeline actual background noise signals through an acoustic emission sensor at a preset height of the offshore platform; S3, injecting the actual background noise signals into the pipeline leakage signals in proportion to generate a labeled training data set simulating real working conditions, and improving the structure of a convolutional neural network, comprising: a) introducing a residual connection: connecting residual blocks in series after the initial convolutional layer, each residual block containing two convolutional layers, each layer followed by batch normalization and a ReLU activation function, and the residual connection being arranged before the first average pooling layer and after the second average pooling layer; b) combining maximum pooling and average pooling: inserting an average pooling layer between the residual blocks, and arranging a maximum pooling layer after the input layer; S4, introducing regularization and optimization algorithms, adding a Dropout layer before the fully connected layer, setting the dropout rate, and using the Adam optimization algorithm to dynamically adjust the network weights; S5, adaptively designing and sequentially configuring the convolutional neural network structure, respectively for the input layer, the initial convolutional layer, the maximum pooling layer, the plurality of residual blocks, the convolutional layer, the Dropout layer, the fully connected layer and the output layer, wherein an average pooling layer is connected after each residual block; The adaptive design in S5 realizes a phased training strategy according to the following details: In the initial training stage, the residual block is taken as the frozen unit, the weight parameters of the first two residual blocks are locked, only the subsequent convolutional layer and fully connected layer are allowed to participate in parameter updating, and the locking proportion is set according to the overall network depth; In the middle training stage, a dynamic unfreezing mechanism is adopted, and after completing a preset number of training cycles, the weight locking state of one residual block is released, and the unfreezing order is adjusted according to the network depth; In the final training stage, all network parameters are jointly adjusted, the learning rate is adjusted according to the initial value, and the adjustment round is set according to the total training period; S6, dividing the labeled training data set into a training set and a test set in proportion, setting the batch size and the training period for training, and outputting the confusion matrix and the accuracy-loss curve to evaluate the performance of the pipeline leakage model; After the training of S6, the following post-processing steps are further included: S6.1, based on the confusion matrix and the accuracy-loss curve, extracting uncertain samples in the model prediction, the uncertain samples being the leakage signal classification results with a confidence lower than a preset threshold; S6.2, performing time-frequency graph analysis on the uncertain samples, specifically including: According to the spectrum graph and time-frequency graph processing method, a time-frequency graph is generated by applying a short-time Fourier transform, energy distribution features in the time-frequency graph are extracted, and abnormal patterns are identified by comparing with normal samples; S6.3, feeding the abnormal patterns back to the S3 step to dynamically adjust the noise injection ratio, wherein the noise ratio adjustment formula is: Wherein, β old is the original noise ratio, β new is the new noise ratio, the misclassification rate is derived from the confusion matrix, the objective accuracy is a preset value, the pipeline leakage model is retrained using the adjusted labeled training data set, and the confusion matrix is iteratively optimized. 2.The pipeline leakage identification method based on the improved convolutional neural network according to claim 1, characterized in that, The pressure pipeline meets the high-pressure fluid conveying standard, uses seamless steel pipe material, and the acoustic emission sensor is installed at the specified position of the experimental section using a coupling medium. 3.The pipeline leakage identification method based on improved convolutional neural network according to claim 1, characterized in that, The pipeline leakage model in the S1 is divided into two main areas, namely an experimental section with a preset length and a buffer section, wherein the experimental section is uniformly distributed with high-pressure welded needle valves for simulating and regulating leakage conditions, and the buffer section is connected with a pressing device through a flexible rubber pipeline, and the pressing device is selected as a Z4DSY type electric pressure self-control pressure test pump. 4.The pipeline leakage identification method based on the improved convolutional neural network according to claim 2, characterized in that, The collection position of the actual background noise signal in the S2 is located at the 19m and 29m platforms of the offshore platform, the installed acoustic emission sensor model and parameter settings are consistent with the equipment in the laboratory, and are fixed through the coupling medium, and the sampling mode of the acoustic emission sensor adopts a long-term timing sampling mode. 5.The pipeline leakage identification method based on improved convolutional neural network according to claim 1, characterized in that, The pipeline leakage signal in the S3 is uniformly divided according to the length of a preset number of sampling points to obtain a leakage signal sample, and the collected actual background noise signal is divided according to the length of the preset number of sampling points to construct a noise signal sample, and finally, the leakage signal sample and the noise signal sample are fused to form a labeled training data set simulating real working conditions. 6.The pipeline leakage identification method based on improved convolutional neural network according to claim 1, characterized in that, In the a), residual connection, max pooling and average pooling are introduced, and the specific introduction positions of the residual connection are before the first average pooling layer, after the second average pooling layer and before the second average pooling layer, each residual block is composed of two convolution layers, followed by batch normalization and ReLU activation function, the input is added to the output connection by adopting this architecture, the gradient back propagation is adjusted without changing the forward propagation, and the introduction of max pooling and average pooling is used to reduce the size of the feature map and reduce the calculation amount. 7.The improved convolutional neural network based pipeline leak identification method of claim 1, wherein, In the S4, the Dropout layer adopts Dropout regularization technology and sets a dropout rate, the dropout rate is used to randomly discard neurons in the training process, the remaining neurons continue to participate in the calculation of the convolutional neural network, prevent overfitting of the pipeline leakage model, finally, the Adam optimization algorithm is used to calculate the adaptive learning rate of each parameter, and the convolutional neural network weight is adjusted. 8.The improved convolutional neural network based pipeline leak identification method of claim 1, wherein, The inner convolution layer of the residual block adopts a hybrid kernel scale configuration, specifically including: The kernel size of the first convolution layer is determined according to the formula where H input represents the length dimension of the current residual block input feature vector, and k1 is the kernel size of the first convolution layer. The second convolution layer performs a depth separable convolution operation, and the kernel size is a fixed value, while introducing a hollow convolution mechanism containing a fixed expansion convolution factor; The feature map channel number matching device is set in the feedforward connection, when the input and output channel numbers do not match, the channel alignment is performed by using a convolution kernel, and the activation function adopts a parameterized linear rectifier unit.

Images