Submarine pipeline migration fault diagnosis method, system and equipment based on bidirectional diffusion bridge

Abstract

The invention discloses a submarine pipeline migration fault diagnosis method, system and equipment based on a bidirectional diffusion bridge mechanism, belongs to the technical field of energy infrastructure fault detection and intelligent diagnosis, and aims to solve the problem of performance degradation caused by field distribution offset when an existing model processes a variable-working-condition submarine oil pipeline monitoring task. And the false alarm and missing alarm risks in the fault detection process are effectively reduced. According to the technical key points, cross entropy loss and boundary expansion loss are utilized to train a fault diagnosis model based on source domain data; a forward diffusion bridge related to a target domain and a reverse diffusion bridge related to a source domain are cascaded to realize effective conversion from the source domain to the target domain; in the reasoning stage, target domain dominant information is reserved in the middle distribution track through a knowledge retention mechanism; and combining the real data set and the conversion data set, and realizing accurate fault diagnosis by using the trained model. As a plug-and-play assembly for executing submarine pipeline cross-domain fault diagnosis, the submarine pipeline cross-domain fault diagnosis method has the following three important characteristics: (1) expandability: the application is wide, and the submarine pipeline cross-domain fault diagnosis method is suitable for any type of data including time sequences or images; (2) flexibility: a channel from a target domain to a source domain is constructed by deconstructing and combining two completely independent diffusion models, which means that an inverse bridge of any specific domain can be reused without retraining; and (3) conciseness: the robustness of the model can be effectively improved by processing diversity of operation conditions introduced by a conversion sample.

Description

Technical Field

[0001] The present invention relates to a submarine pipeline migration fault diagnosis method based on a bidirectional diffusion bridge, relates to submarine oil pipeline fault diagnosis technology, and belongs to the field of energy infrastructure fault detection and intelligent diagnosis technology. Background Art

[0002] As critical infrastructure for the transportation of marine oil and gas resources, submarine pipelines play an irreplaceable and important role in ensuring national energy strategic security and the efficient development of deep-sea resources. However, with the continuous expansion of deepwater oil and gas production and the increase in service life, pipeline systems are facing severe operating fluctuations, such as pressure pulsation, sudden temperature changes, ocean current disturbances, and other non-steady-state operating environments. These variable operating conditions significantly increase the risk of submarine pipeline failures and the complexity of operation and maintenance. Therefore, the research of high-precision intelligent fault diagnosis methods that adapt to variable operating conditions has significant practical value and engineering significance for improving the safety and stability of submarine pipeline operations.

[0003] As a cutting-edge technology in the current field of industrial intelligent diagnosis, deep learning, relying on the automatic feature extraction capabilities of multi-layer neural networks, has made positive progress in many continuous operating condition fault identification scenarios. However, in complex and variable operating conditions, the performance of deep learning models often faces significant challenges. On the one hand, model training relies heavily on datasets that cover all operating conditions. However, in actual operation, abnormal events are scarce and uneven, which limits the model's performance in learning fault samples. On the other hand, the data distribution offset caused by variable operating conditions significantly weakens the generalization ability of the trained model under the target operating conditions, reducing the model's diagnostic accuracy and stability.

[0004] In recent years, research on submarine pipeline fault diagnosis has increased significantly. For example, document CN109632942B provides a pipeline defect size inversion method based on SL. The method includes: performing data interpolation preprocessing on a sample set of triaxial magnetic flux leakage signals of defects with known defect sizes; performing multidimensional feature extraction to extract the signal feature set [F(t), F(δ)] in the time domain; constructing the wavelet energy feature set F(w) in the frequency domain using a wavelet transform-based method; constructing a multidimensional feature set F = [F(t), F(δ), F(w)] for nsample defects; employing an iterative StackingLeaning network to automatically determine the final network structure; and predicting the size of the magnetic flux leakage signal of the defect to be tested to obtain a predicted pipeline defect size. This prior art uses a multidimensional feature extraction method in the time-frequency domain to construct defect features, comprehensively analyze the information contained in the magnetic flux leakage signal, and improve the network's ability to predict the size of complex defects.

[0005] The prior art with the document number CN114964476B discloses a fault diagnosis method, device and equipment of an oil and gas pipeline system moving equipment, which comprises the following steps: collecting vibration signals of an oil and gas pipeline system moving equipment to be detected; performing continuous wavelet transform on the vibration signals to obtain a two-dimensional time-frequency image; inputting the two-dimensional time-frequency image into a pre-constructed convolutional neural network fault diagnosis model, using the convolutional neural network fault diagnosis model to perform fault diagnosis on the oil and gas pipeline system moving equipment to be detected, and obtaining a fault diagnosis result; and using a LIME algorithm to analyze the fault diagnosis result to obtain fault analysis information, wherein the fault analysis information is to mark different energy intensities on the two-dimensional time-frequency image, and different energy intensities represent different contribution degrees of the time-frequency regions corresponding to the energy intensities to the fault diagnosis result. The method can avoid a tedious feature extraction process, break the uncontrollability of the diagnosis result, and improve the fault diagnosis accuracy.

[0006] However, the prior art does not propose to use a technical means based on a bidirectional diffusion bridge to change a traditional transfer learning method based on statistical alignment or feature reconstruction, and does not mention how to alleviate the recognition degradation problem of a deep model in a distribution offset scene in pipeline fault diagnosis. SUMMARY

[0007] The technical problem to be solved by the present application is:

[0008] In view of the above problems, the present application provides a submarine pipeline transfer fault diagnosis method based on a bidirectional diffusion bridge, which effectively converts target domain data under variable working conditions into a source domain style relied on for training by introducing a diffusion generation mechanism, thereby alleviating the recognition degradation problem of a deep model in a distribution offset scene.

[0009] The technical solution adopted by the present application to solve the above technical problems is:

[0010] A submarine pipeline transfer fault diagnosis method based on a bidirectional diffusion bridge, the method comprising the following steps:

[0011] Step 1: Collecting pressure parameters using an indoor pipeline laboratory to construct a source domain data set

[0012] Using an indoor pipeline laboratory platform, by arranging multiple-point pressure sensors at key positions, collecting pressure data during pipeline operation, obtaining original pressure data covering typical working conditions such as normal operation, leakage and blockage, and constructing a source domain training data set Wherein represents the i-th pressure sequence composed of N sampling points,

[0013] is the corresponding working condition label, C is the total number of fault types, n s is the total number of source domain samples.

[0014] Step 2: Train the source domain pipeline fault diagnosis model using the source domain dataset

[0015] Step 2.1: Use long short-term memory network to build feature extraction and classification module in fault diagnosis model

[0016] The long short-term memory network is selected as the feature extraction and classification module, and its core structure includes input gate, forget gate and output gate, which can effectively model the long-term dependence features in time series and is suitable for fault pattern recognition tasks in pipeline pressure data.

[0017] Step 2.2: Train the source domain fault diagnosis model using cross-entropy loss and boundary expansion loss

[0018] The model training uses cross-entropy loss function and boundary expansion loss function, which are defined as follows:

[0019]

[0020] Where, represents the i-th sample of the source domain belonging to the c-th class, and z represents the signal feature extracted by the diagnosis model; e m and e n represent the classification error of the m-th sample and the n-th sample of the source domain, respectively; represents the indicator function, which is 1 if e m >e n , otherwise 0; Δ m, n is the confidence difference, which is calculated as follows:

[0021]

[0022] Where, P i (c) and P j (c) are the probabilities of the i-th source domain sample and the j-th source domain sample belonging to the c-th class given by the diagnosis model.

[0023] Step 2.3: Use Adam optimizer for parameter update

[0024] The training process uses Adam optimizer for parameter update, and its update rule is:

[0025]

[0026] Where, θ t is the model parameter of the t-th iteration; is the gradient of the t-th iteration; α = 0.011 is the learning rate; β1 = 0.9, β2 = 0.999 are the exponential decay rates of momentum terms; ∈ = 10 -8 is a small constant to prevent division by zero; m tand v t are the first and second order moment estimates; and are the bias-corrected estimates.

[0027] The source domain fault diagnosis model parameters are iteratively updated through the optimization strategy, so that they have the ability to accurately classify multiple working conditions on the source domain data.

[0028] Step 4: Collect real pressure data of submarine pipelines and build target domain data set

[0029] Using the submarine pipeline online monitoring system, by laying multiple point pressure sensors in key pipe sections, real-time acquisition of pressure signals in the pipeline operation process, obtaining original pressure data covering typical fault conditions such as normal operation, leakage, blockage, corrosion, etc., to build a target domain training data set wherein represents the pressure sequence composed of N sampling points, is the corresponding working condition label, K is the total number of fault types, n t is the total number of target domain samples.

[0030] Step 5: Build diffusion model

[0031] The purpose of the forward process of the diffusion model is to convert the data into a Gaussian distribution after T time steps. First, define the stochastic differential equation of the forward diffusion process, which is in the form of:

[0032]

[0033] where x t represents the state of the diffusion process at time t; f(t) and g(t) are time-dependent functions; ∈ θ is a parameterized noise estimation function; σ t is a positive scalar function at time t; ∈ θ is a noise prediction model.

[0034] The forward stochastic differential equation is equivalent to a deterministic probability flow deterministic ordinary differential equation, which is convenient for solving and numerical calculation. The form of this deterministic ordinary differential equation is:

[0035]

[0036] where, is the distribution gradient of the diffusion process.

[0037] Further, using the DPM-Solver numerical method, formula (8) is solved to obtain the mapping from the initial state x0to the target state x T . The parameterized velocity field where λ represents the noise schedule), represents the velocity function of the model at parameters θ. The mapping of the state is achieved by integrating the expression:

[0038]

[0039] The integral starts from the initial value of the noise schedule λ0to the final value λ T , which completes the reverse sampling from the noise space to the target data distribution.

[0040] Step 6: Construct the source domain diffusion model and the target domain diffusion model based on step 5

[0041] According to the diffusion forward process described in step 5, the forward process and the reverse process of the source domain diffusion model can be obtained, which are as follows:

[0042]

[0043] where x (i) represents the latent space sample, and the latent space follows a Gaussian distribution; represents the source domain diffusion model; represents the converted source domain sample.

[0044] The forward process and the reverse process of the target domain diffusion model are as follows:

[0045]

[0046] where x (i) represents the latent space sample, and the latent space follows a Gaussian distribution; represents the target domain diffusion model; represents the converted target domain sample.

[0047] Step 7: Train the source domain and target domain diffusion models respectively

[0048] Based on formula (8), the purpose of the diffusion model is to learn a parameterized distribution to estimate p (s) (x0). The diffusion probability model is a latent variable model:

[0049]

[0050]

[0051] The training of the diffusion model needs to optimize the variational lower bound of the data likelihood, so as to improve the ability of the model to accurately reconstruct the data according to the prior knowledge. In practice, we use a simplified loss to replace the variational lower bound, based on the source domain data in step 1, the target domain data in step 4, formula (14), formula (15), to train the source domain and target domain diffusion models. The specific loss function form is as follows:

[0052]

[0053] Where, alpha t is a positive scalar value.

[0054] Step 8: Construct the diffusion bridge from the target domain to the source domain to complete the inference process

[0055] Step 8.1: Complete the forward inference from the target domain to the latent space

[0056] According to the trained target domain diffusion model obtained in step 7, the forward diffusion bridge from the target domain to the latent space can be obtained, and the latent space sample can be obtained in the form as follows:

[0057]

[0058] Step 8.2: Complete the reverse inference from the latent space to the source domain

[0059] According to the trained source domain diffusion model obtained in step 7 and the latent space sample obtained in formula (20), the reverse diffusion bridge from the latent space to the source domain can be obtained, and the converted source domain sample can be obtained in the form as follows:

[0060]

[0061] Step 9: Input the converted source domain sample obtained in step 8 into the source domain fault diagnosis model obtained in step 2 to obtain the fault diagnosis result.

[0062] The present application has the following beneficial technical effects:

[0063] The submarine pipeline migration fault diagnosis method based on the bidirectional diffusion bridge provided by the present application has stronger generation ability and stable step-by-step inversion characteristics compared with the traditional migration learning method based on statistical alignment or feature reconstruction, can guide the target domain sample to converge to the source domain layer by layer, and can enhance the explainability and compatibility of the data in the source model space while maintaining the key structural features. Compared with the prior art, the present application has the following advantages and effects:

[0064] (1) Different from general intelligent fault diagnosis methods based on deep learning technology, the present application mainly focuses on the problem of fault diagnosis under small sample and variable working conditions. In practical application of submarine pipelines, fault samples are scarce and working conditions change dramatically, making it difficult for traditional deep learning methods to handle the complex scenario of small samples and variable working conditions. The present application solves the problem of small sample and variable working conditions by introducing the transfer learning technology of bidirectional diffusion bridge, which can significantly improve the precision and stability of the fault diagnosis model under variable working conditions by adaptively adjusting the data distribution of the target domain in the case of data scarcity, thereby improving the practical application value of intelligent fault diagnosis methods.

[0065] (2) Different from traditional domain adaptation methods based on metric indicators or adversarial learning, the present application constructs a data adaptation method based on bidirectional diffusion bridge. Traditional domain adaptation methods usually focus on reducing the distribution difference between source domain and target domain, while the present application uses adaptive technology to map target domain data to source domain style, and dynamically adjusts the distribution of data according to the characteristics of variable working conditions to adapt to different working conditions. This data adaptation method does not rely on static distribution assumption, but makes real-time adaptive adjustment according to the actual changes of the target domain, greatly improving the accuracy and reliability of pipeline fault diagnosis, while avoiding the overfitting risk in traditional methods.

[0066] (3) Considering the demand of intelligent fault diagnosis technology in practical application, this method has high practical application value. For the complex variable working condition environment of submarine pipelines, the present application has verified its effectiveness in multiple experiments and application scenarios, especially in handling small sample and high variable working condition pipeline fault diagnosis, and has achieved remarkable results. The successful application of this technology provides a reliable solution for the intelligent monitoring and fault warning of submarine pipelines, and has strong engineering popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 is a flow chart of submarine pipeline migration fault diagnosis based on bidirectional diffusion bridge according to an embodiment of the present application;

[0068] Figure 2 is the performance quantization evaluation result of fault diagnosis of the present application embodiment;

[0069] Figure 3 is the performance quantization evaluation of the disclosed data set of the present application embodiment;

[0070] Figure 4 is the data distribution change graph of the present application embodiment (left for source domain, right for target domain);

[0071] Figure 5 is the parameter sensitivity analysis graph of the present application embodiment. DETAILED DESCRIPTION

[0072] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate the present application and, together with the description given below, serve to Figures 1-5 The present application is further described:

[0073] In order for those skilled in the art to better understand the technical solutions in the present application, 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.

[0074] Figure 1 is the flow chart of the submarine pipeline migration fault diagnosis based on the bidirectional diffusion bridge in the technical solutions of the present application, and the overall process is realized by programming in Python language. Referring to Figure 1 , a submarine pipeline migration fault diagnosis method based on a bidirectional diffusion bridge, comprising the following steps:

[0075] Step 1: Collecting pressure parameters in a laboratory pipeline laboratory to construct a source domain data set

[0076] Using the laboratory pipeline experiment platform, by arranging multiple point pressure sensors at key positions, collecting pressure data during pipeline operation, obtaining original pressure data covering typical working conditions such as normal operation, leakage and blockage, and constructing a source domain training data set Among them represents the i-th pressure sequence composed of N sampling points, is the corresponding working condition label, C is the total number of fault types, n s is the total number of source domain samples.

[0077] Step 2: Using the source domain data set to train the source domain pipeline fault diagnosis model

[0078] Step 2.1: Using long short-term memory network to construct feature extraction and classification module in fault diagnosis model

[0079] The model structure selects long short-term memory network as the feature extraction and classification module, and its core structure includes input gate, forget gate and output gate, which can effectively model the long-term dependence features in time series, and is suitable for fault pattern recognition tasks in pipeline pressure data.

[0080] Step 2.2: Training the source domain fault diagnosis model using cross-entropy loss and boundary expansion loss

[0081] The model training adopts cross-entropy loss function and boundary expansion loss function, and the specific definitions are as follows:

[0082]

[0083] Among them, represents the i-th sample belonging to the c-th class in the source domain, and z represents the signal features extracted by the diagnosis model; e m and e nclassification error of the mth sample and the nth sample in the source domain, respectively; is an indicator function, which is 1 if e m n is 0 otherwise; Δ m,n is the confidence difference, which is calculated as follows:

[0084]

[0085] where P i (c) and P j (c) are the probabilities of the ith sample and the jth sample in the source domain belonging to class c given by the diagnostic model, respectively.

[0086] Step 2.3: Parameter update using Adam optimizer

[0087] The training process uses the Adam optimizer for parameter update, and its update rule is as follows:

[0088]

[0089] where θ t is the model parameter at the tth iteration; is the gradient at the tth iteration; α = 0.011 is the learning rate; β1 = 0.9, β2 = 0.999 are the exponential decay rates of the momentum term; ∈ = 10 -8 is a small constant to prevent division by zero; m t and v t are the first and second moment estimates, respectively; and are the bias-corrected estimates.

[0090] Through this optimization strategy, the source domain fault diagnosis model parameters are iteratively updated, so that it has the ability to accurately classify multiple working conditions on the source domain data.

[0091] Step 4: Collect real pressure data of submarine pipelines and build target domain dataset

[0092] Using the submarine pipeline online monitoring system, by laying multiple point pressure sensors in key pipe sections, real-time acquisition of pressure signals during pipeline operation is realized, and original pressure data covering typical fault conditions such as normal operation, leakage, blockage, and corrosion are obtained to build the target domain training dataset where represents the ith pressure sequence composed of N sampling points, is the corresponding working condition label, K is the total number of fault types, and n t is the total number of target domain samples.

[0093] Step 5: Build diffusion model ​

[0094] The purpose of the forward process of the diffusion model is to convert the data into a Gaussian distribution after T time steps. First, define the stochastic differential equation of the forward diffusion process, which is in the form of:

[0095]

[0096] where x t represents the state of the diffusion process at time t; F(t) and g(t) are time-dependent functions; ∈ θ is a parameterized noise estimation function; σ t is a positive scalar function at time t; ∈ θ is a noise prediction model.

[0097] The forward stochastic differential equation is equivalent to a deterministic probability flow deterministic ordinary differential equation, which is convenient for solving and numerical calculation. The form of the deterministic ordinary differential equation is:

[0098]

[0099] where, is the distribution gradient of the diffusion process.

[0100] Further, using the DPM-Solver numerical method, formula (8) is solved to obtain the mapping from the initial state x0to the target state x T . Introduce the parameterized velocity field (where λ represents the noise time schedule), which represents the velocity function of the model under the parameter θ. The mapping of the state is realized by the integral expression:

[0101]

[0102] The integral starts from the noise scheduling starting value λ0and ends at the termination value λ T , which completes the reverse sampling from the noise space to the target data distribution.

[0103] Step 6: Construct the source domain diffusion model and the target domain diffusion model based on step 5

[0104] According to the diffusion forward process described in step 5, the forward process and the inverse process of the source domain diffusion model can be obtained, which is in the form of:

[0105]

[0106] where x (i) represents the latent space sample, and the latent space follows a Gaussian distribution; represents the source domain diffusion model; represents the converted source domain sample.

[0107] The forward process and the inverse process of the target domain diffusion model are as follows:

[0108]

[0109] Wherein, x (i) represents a latent space sample, and the latent space is subject to a Gaussian distribution; represents the target domain diffusion model; represents the converted target domain sample.

[0110] Step 7: Training the source domain and target domain diffusion models respectively

[0111] Based on formula (8), the purpose of the diffusion model is to learn a parameterized distribution to estimate p (s) (x0). The diffusion probability model is a latent variable model:

[0112]

[0113] The training of the diffusion model needs to optimize the variational lower bound of the data likelihood, so as to improve the ability of the model to accurately reconstruct the data according to the prior knowledge. In practice, we use a simplified loss to replace the variational lower bound, based on the source domain data of step 1, the target domain data of step 4, formula (14), formula (15), to train the source domain and target domain diffusion models. The specific loss function form is as follows:

[0114]

[0115] Wherein, α t is a positive scalar value.

[0116] Step 8: Constructing the diffusion bridge from the target domain to the source domain to complete the inference process

[0117] Step 8.1: Completing the forward inference from the target domain to the latent space

[0118] According to the trained target domain diffusion model obtained in step 7, the forward diffusion bridge from the target domain to the latent space can be obtained, and the latent space sample can be obtained in the form as follows:

[0119]

[0120] Step 8.2: Completing the reverse inference from the latent space to the source domain

[0121] According to the trained source domain diffusion model obtained in step 7 and the latent space sample obtained in formula (20), the reverse diffusion bridge from the latent space to the source domain can be obtained, and the converted source domain sample can be obtained in the form as follows:

[0122]

[0123] Step 9: input the converted source domain sample obtained in step 8 into the source domain fault diagnosis model obtained in step 2 to obtain a fault diagnosis result.

[0124] Embodiment

[0125] The effectiveness of the method is verified by taking fault diagnosis of an oil and gas pipeline network as an example.

[0126] 1. Experimental setup

[0127] The source domain fault diagnosis model in step 2 is composed of three layers of long short-term memory networks. During training, the total number of iterations is 5000, the batch size is 128, the optimizer is Adam, and the learning rate is 1e-4. The algorithms mentioned in the present application are built by the PyTorch framework and trained on an NVIDIA GEFORCE RTX 3090 GPU.

[0128] The source domain diffusion model and the target domain diffusion model in step 7 are composed of 12 layers of Transformers, each using 384 fully connected dimensions and 6 heads of multi-head attention mechanism, the optimizer is Adam, the learning rate is 1e-4, the total number of iterations is 1000, and the batch size is 64.

[0129] 2. Comparative algorithms

[0130] In order to comprehensively evaluate the performance of the proposed fault diagnosis model, the present application selects five kinds of transfer learning models for comparison.

[0131] 3. Quantitative evaluation of metric indicators

[0132] Compared with other methods, the proposed technology (code DBDA) performs particularly outstanding in handling the source-free transfer learning task of pipeline fault diagnosis, with the highest average accuracy of 72.84%, as shown in Figure 2 It is worth noting that as a representative of explicitly modeling domain-invariant features and domain-specific features, OUDA still cannot match DBDA, which indicates that the feature alignment method is not flexible enough in dealing with data privacy scenarios. In addition, DDA, as the most advanced algorithm in computer vision, has failed in many cases of time series classification. On the one hand, the linear low-pass filter of DDA is only suitable for capturing high-level features of images. On the other hand, DDA is based on pair-wise training and cannot adapt to more extensive transmission scenarios. Therefore, traditional diffusion models can handle feature conversion and data adaptation, while the proposed DBDA can effectively handle transmission tasks in the real world without considering the availability of source data. In addition, the performance of the proposed technology on public datasets still surpasses the comparative algorithms, as shown in Figure 3 .

[0133] 4. Visualization evaluation

[0134] To verify the data transformation capability of the DBDA model, a representation analysis was conducted on the concentric circles (CR) extracted from the complex two-dimensional distribution. As shown in the left column, in the source domain, each data point is assigned a unique color to uniquely identify its instance. These points are mapped to the target domain through the latent space, and the color label remains consistent, thereby realizing the visual tracking of the transformation trajectory thereof. As shown in the right column, the spatial proximity and continuity of the same color points across domains indicate that the proposed model effectively captures the structural consistency and maintains the transformation correspondence between the source distribution and the target distribution. Figure 4 Figure 4 5. Performance analysis test

[0135] We evaluated the hyperparameter sensitivity of DBDA to the following factors: learning rate, Transformer, and loss balancing parameter. The results are shown in

[0136] We found that the learning rate is the most common influencing factor and must be carefully selected. In addition, increasing the number of blocks in DBDA does not necessarily improve performance, indicating that larger configurations are not always beneficial. In addition, the balancing parameter has little effect on the performance of DBDA. Figure 5 In summary, the method proposed by the present application has better quantitative evaluation indicators and performance tests than existing comparative methods, effectively reducing the risk of false positives and false negatives, and improving the accuracy of submarine pipeline fault diagnosis. The method described in the present application has been verified through simulation experiments and practical applications, and the claimed technical effects and practicality of the present application have been verified.

[0137] The algorithm (method) proposed by the present application is the underlying technical kernel of the present application, and various products can be derived based on the algorithm.

[0138] Based on the algorithm (method) proposed by the present application, a submarine pipeline migration fault diagnosis system based on a bidirectional diffusion bridge is developed using a program language. The system has program modules corresponding to the steps of the above technical solution, and when running, it executes the steps in the above-mentioned submarine pipeline migration fault diagnosis method based on a bidirectional diffusion bridge.

[0139] The computer program of the developed system (software) is stored on a computer readable storage medium, and the computer program is configured to realize the steps of the above-mentioned submarine pipeline migration fault diagnosis method based on a bidirectional diffusion bridge when called by a processor. That is, the present application is materialized on a carrier to become a computer program product.

[0140]

[0141] ​​The submarine pipeline migration fault diagnosis device comprises at least one processor and a memory connected with the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned submarine pipeline migration fault diagnosis method based on the bidirectional diffusion bridge, so as to realize fault diagnosis and early warning of abnormal conditions of the oil pipeline.

[0142] Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0143] Computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and / or object-oriented programming language, and / or in assembly / machine language. As used in this application, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0144] It should be understood that the steps shown in the various forms of flow above can be reordered, added to, or deleted from. For example, the steps recited in the present application can be performed in parallel, in series, or in a different order, as long as the desired results of the technology disclosed in the present application are achieved, and are within the scope of protection of the present application.

Claims

1. A submarine pipeline migration fault diagnosis method based on a bidirectional diffusion bridge, the method is used for submarine pipeline fault diagnosis, characterized in that: The implementation process of the method is: Step 1: Use an indoor pipeline laboratory to collect pressure parameters and build a source domain dataset Using the indoor pipeline experimental platform, by deploying multiple pressure sensors at key locations, we collect pressure data during pipeline operation, obtain raw pressure data covering typical operating conditions such as normal operation, leakage, and blockage, and construct a source domain training dataset. in represents the i-th pressure sequence consisting of N sampling points, is the corresponding working condition label, C is the total number of fault types, n s is the total number of source domain samples; Step 2: Use the source domain dataset to train the source domain pipeline fault diagnosis model Step 2.1: Use LSTM to build the feature extraction and classification module in the fault diagnosis model The model structure uses a long short-term memory network as the feature extraction and classification module. Its core structure includes input gates, forget gates, and output gates. It can effectively model long-term dependency features in time series and is suitable for fault pattern recognition tasks in pipeline pressure data. Step 2.2: Train the source domain fault diagnosis model using cross entropy loss and margin extension loss The model training uses the cross entropy loss function and the boundary extension loss function, which are defined as follows: in, represents the i-th sample in the source domain belonging to the c-th class, z represents the source domain signal feature extracted by the diagnostic model; e m and e n Represent the classification errors of the mth sample and the nth sample in the source domain respectively; represents the indicator function, if e m >e n 1, otherwise 0; Δ m,n is the confidence difference, calculated as follows: Among them, P i (c) and P j (c) are the probabilities that the i-th source domain sample and the j-th source domain sample belong to class c given by the diagnostic model; Step 2.3: Update parameters using Adam optimizer The training process uses the Adam optimizer to update parameters, and its update rule is: Among them, θ t is the model parameter of the tth iteration; is the t-th iteration gradient; α = 0.011 is the learning rate; β1 = 0.9, β2 = 0.999 is the exponential decay rate of the momentum term; ∈ = 10 -8 A minimum constant to prevent division by zero; m t and v t are the first-order and second-order moment estimates; and is the bias-corrected estimate; The optimization strategy is used to iteratively update the source domain fault diagnosis model parameters, enabling it to accurately classify various working conditions based on source domain data. Step 4: Collect real submarine pipeline pressure data and build the target domain dataset Using the submarine pipeline online monitoring system, by deploying multiple pressure sensors in key pipeline sections, real-time pressure signals are collected during pipeline operation. Raw pressure data covering typical fault conditions such as normal operation, leakage, blockage, corrosion, etc. is obtained to construct a target domain training dataset. in represents the i-th pressure sequence consisting of N sampling points, is the corresponding working condition label, K is the total number of fault types, n t is the total number of samples in the target domain; Step 5: Build a diffusion model The purpose of the forward process of the diffusion model is to convert the data into a Gaussian distribution after T time steps. First, the stochastic differential equation of the forward diffusion process is defined, which is in the form of: Among them, x t represents the state of the diffusion process at time t; f(t) and g(t) are time-dependent functions; ∈ θ is the parameterized noise estimation function; σ t is a positive scalar function at time t; ∈ θ is the noise prediction model; The forward stochastic differential equation is equivalently converted into a deterministic probability flow deterministic ordinary differential equation, which is convenient for solution and numerical calculation; the form of the deterministic ordinary differential equation is: in, is the distribution gradient of the diffusion process; Furthermore, the DPM-Solver numerical method is used to solve formula (8) and obtain the equation from the initial state x0 to the target state x T Mapping; Introducing parameterized velocity field (where λ represents the noise schedule), which represents the speed function of the model under the parameter θ; the state mapping is achieved through the integral expression: The integral is integrated from the noise scheduling starting value λ0 to the ending value λ T , complete the reverse sampling from the noise space to the target data distribution; Step 6: Build the source domain diffusion model and the target domain diffusion model based on step 5 According to the forward diffusion process described in step 5, the forward process and reverse process of the source domain diffusion model can be obtained, which are as follows: Among them, x (i) Represents a sample of the latent space, where the latent space follows a Gaussian distribution; represents the source domain diffusion model; Represents the converted source domain samples; The forward and reverse processes of the target domain diffusion model are as follows: Among them, x (i) Represents a sample of the latent space, where the latent space follows a Gaussian distribution; represents the target domain diffusion model; Represents the target domain sample after conversion; Step 7: Train the source domain and target domain diffusion models separately Based on formula (8), the purpose of the diffusion model is to learn a parameterized distribution To estimate p (s) (x0); the diffusion probability model is a latent variable model: The training of the diffusion model requires optimizing the variational lower bound of the data likelihood, thereby improving the model's ability to accurately reconstruct data based on prior knowledge. In practice, a simplified loss is used instead of the variational lower bound. Based on the source domain data in step 1, the target domain data in step 4, formulas (14) and (15), the source and target domain diffusion models are trained. The specific loss function is as follows: Among them, α t is a positive scalar value; Step 8: Build a diffusion bridge from the target domain to the source domain to complete the reasoning process Step 8.1: Complete forward inference from the target domain to the latent space According to the trained target domain diffusion model obtained in step 7, the forward diffusion bridge from the target domain to the latent space can be obtained, and the latent space sample form is obtained as follows: Step 8.2: Complete reverse inference from latent space to source domain According to the trained source domain diffusion model obtained in step 7 and the latent space samples obtained by formula (20), the reverse diffusion bridge from the latent space to the source domain can be obtained, and the converted source domain samples are obtained as follows: Step 9: Input the converted source domain samples obtained in step 8 into the source domain fault diagnosis model obtained in step 2 to obtain the fault diagnosis results.

2. The method for diagnosing submarine pipeline migration faults based on a bidirectional diffusion bridge according to claim 1, characterized in that: In step 1, an indoor pipeline experimental platform is used to deploy multi-point pressure sensors at key locations to collect pressure data during pipeline operation, and obtain original pressure data covering typical operating conditions such as normal operation, leakage, and blockage.

3. The method for diagnosing submarine pipeline migration faults based on a bidirectional diffusion bridge according to claim 1, characterized in that: In step 2, the source domain fault diagnosis model is trained using the cross entropy loss function and the boundary extension loss function.

4. The method for diagnosing submarine pipeline migration faults based on a bidirectional diffusion bridge according to claim 1, characterized in that: In step 5, the diffusion model is established and solved using the DPM-Solver numerical method.

5. The method for diagnosing submarine pipeline migration faults based on a bidirectional diffusion bridge according to claim 1, characterized in that: In step 6, a bidirectional diffusion bridge is constructed for data transformation using the target domain forward bridge and the source domain reverse bridge.

6. The method for diagnosing submarine pipeline migration faults based on a bidirectional diffusion bridge according to claim 1, characterized in that: In step 8, the target domain data is converted to the source domain data using a bidirectional diffusion bridge.

7. The method for diagnosing submarine pipeline migration faults based on a bidirectional diffusion bridge according to claim 1, characterized in that: In step 9, the converted source domain samples are input into the source domain fault diagnosis model to obtain the fault diagnosis results.

8. A submarine pipeline migration fault diagnosis method system based on a bidirectional diffusion bridge, characterized by: The system has a program module corresponding to the steps of any one of claims 1 to 7, and executes the steps of the submarine pipeline migration fault diagnosis method based on a bidirectional diffusion bridge when running.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of a submarine pipeline migration fault diagnosis method based on a bidirectional diffusion bridge according to any one of claims 1 to 7 when called by a processor.

10. A submarine pipeline migration fault diagnosis device, characterized by: The submarine pipeline relocation fault diagnosis device includes at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to execute the submarine pipeline relocation fault diagnosis method based on a bidirectional diffusion bridge according to any one of claims 1 to 7, thereby achieving fault diagnosis and abnormality warning for the submarine pipeline.

Images