A data analysis-based ultrahigh-pressure pipeline residual life prediction system and method

Abstract

The application discloses a kind of based on data analysis's ultrahigh pressure pipeline residual life prediction system and method, it is related to industrial data analysis technical field, through the collaborative data acquisition of image acquisition unit and sensor group, pipeline surface image and sensor multidimensional data are synchronously acquired;Image significant change is determined using similarity threshold, sensor data is compressed through phase space reconstruction and principal component analysis;The image feature vector after normalization is spliced with sensor feature vector, and then a synchronous sample is formed by time-weighted fusion;A joint degradation modeling network based on multi-task learning is used, a two-layer fully connected network is used in the shared feature extraction layer to learn the common degradation law of image features and sensor features, the degradation features of each data type are learned by the independent task output layer respectively, and the high-order interaction fusion of the two types of features is realized by matrix multiplication through the bilinear fusion layer, and finally the pipeline degradation risk value is output.

Description

Technical Field

[0001] This invention relates to the field of industrial data analysis technology, specifically a data analysis-based system and method for predicting the remaining life of ultra-high pressure pipelines. Background Technology

[0002] Ultra-high pressure pipelines, as a core component of key pressure-bearing equipment in EVA (Electrode Vapor Absorption Regulator) devices, play a crucial role in transporting high-temperature, high-pressure, and corrosive media. These pipelines are subjected to complex thermal loads and chemical corrosion over long periods, leading to gradual deterioration of the internal microstructure of their metallic materials, continuous thinning of the pipeline walls, and a sustained degradation of the overall structural integrity of the pipeline.

[0003] Existing technologies for monitoring ultra-high pressure pipelines in EVA devices suffer from the following problems: Within the same monitoring time window, the degradation characteristics of pipeline surface images change extremely slowly; macroscopic cracks often require weeks or even months to accumulate from initiation to significant, identifiable propagation; image frames exhibit extremely high similarity and very low effective information density; while the process data collected by the sensor system during the same period is enormous, with multi-channel data such as temperature, pressure, and ultrasound continuously entering at high sampling frequencies, and the data sequences containing rich process dynamic information. This timescale mismatch between image features and sensor data makes it difficult to establish an effective spatiotemporal correlation between the two using traditional simple time-series alignment methods.

[0004] Therefore, how to effectively solve the time scale mismatch problem of heterogeneous multi-source data and achieve high-precision prediction of the remaining life of ultra-high pressure pipelines has become a key technical problem that needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a data analysis-based system and method for predicting the remaining life of ultra-high pressure pipelines, in order to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the remaining life of ultra-high pressure pipelines based on data analysis;

[0007] The method includes the following specific steps: Step 1: Setting up detection areas and coordinating data acquisition: Detection areas are set up at valve positions, tee positions, and elbow positions along the ultra-high pressure pipeline. Through the coordinated data acquisition of the image acquisition unit and the sensor group, the pipeline surface image and multi-dimensional sensor data are acquired synchronously; Step 2: Image significant change recognition and image observation group construction: Images at the initial moment and images at subsequent moments are acquired. The similarity between the two images is compared. When the similarity is lower than a preset similarity threshold, a significant difference is determined. The images at the initial moment and the images at subsequent moments are combined into an image observation group; Step 3: Obtaining sensor data in the time interval: For the pre-... The sensor data within the time window is compressed, and the massive sensor data is transformed into low-dimensional feature vectors through phase space reconstruction and principal component analysis. Step 4 establishes synchronous samples: the image observation group is associated and fused with the sensor features within the corresponding time window. The fused synchronous sample features are formed through feature extraction, normalization, vector concatenation, and time-weighted fusion. Step 5 spatiotemporal fusion and joint degradation modeling: a joint degradation modeling method based on multi-task learning is adopted. Through a deep neural network architecture with a shared feature extraction layer, an image feature independent task output layer, a sensor feature independent task output layer, and a bilinear fusion layer, the pipeline degradation risk value is output.

[0008] Furthermore, in step 1, in the ultra-high pressure pipeline system of the EVA device, there are several key hydrodynamic characteristic locations distributed along the pipeline. The pipeline walls at these locations are subjected to complex stress distribution and media scouring, and are sensitive areas for performance degradation. Therefore, the locations for setting detection areas include valve locations, tee locations, and elbow locations. Each detection area corresponds to a unique area code and spatial coordinate parameters.

[0009] Furthermore, in step 1, the image acquisition unit and the sensor group work together to acquire data, enabling the synchronous acquisition of pipe surface images and multi-dimensional sensor data. The image acquisition unit is used to capture the microscopic texture and defect features of the pipe surface, while the sensor group is used to capture multi-dimensional physical parameters of the pipe's operating status, including three types: temperature sensors, pressure sensors, and ultrasonic sensors.

[0010] Furthermore, in step 2, images of the pipe surface are periodically acquired at preset time intervals, and valid samples are intelligently filtered based on the degree of change in image content. Image acquisition is performed at fixed time intervals, with the image acquired at time t0 serving as the initial image and the image acquired at time t0+Δt serving as the control image. For each acquired original image frame, correction and enhancement processing are first performed through an image preprocessing module, including distortion correction, contrast enhancement, and noise filtering, to eliminate the impact of changes in acquisition conditions on image quality. Subsequently, an image feature extraction model is used to extract features from the preprocessed images, obtaining a feature image representation for each image. This feature image is a multi-dimensional numerical vector that encodes the texture information and structural features of the pipe surface. After obtaining the feature representations of the initial and control images, the system calculates the similarity between the two feature images. When time Δt has elapsed, i.e., when the image at time t0+Δt is acquired, the similarity with the initial image is below a threshold. Therefore, the two images are grouped into an image observation group, and the acquisition timestamp and detection area code of this observation group are associated and recorded as the basic data unit for subsequent synchronous sample construction.

[0011] Furthermore, in step 3, the sensor data within the time window [t0, t0+Δt] is compressed. For the original time series data X of the j-th sensor channel... j = (x 1j x 2j , ..., x nj (where n is the total number of data points). First, the optimal delay time τj and the optimal embedding dimension mj need to be determined. The optimal delay time τj is determined using the autocorrelation function method, which calculates the autocorrelation function of the time series. Where τ is the delay, R(τ) decreases as τ increases, and the value of τ corresponding to the first decrease of R(τ) to R(0) / e is determined as the optimal delay time τj. The optimal embedding dimension mj is determined by the pseudo-nearest neighbor method. This method determines the optimal dimension by calculating the pseudo-nearest neighbor ratio of adjacent points in the embedding space. The minimum dimension corresponding to the pseudo-nearest neighbor ratio being lower than a preset threshold is the optimal embedding dimension mj. In a preferred embodiment of the present invention, the threshold of the autocorrelation function method is set to 1 / e, and the threshold of the pseudo-nearest neighbor method is set to 0.1.

[0012] Based on the optimal delay time τj and the optimal embedding dimension mj, construct the delay coordinate vector Y. j (i) The formula for constructing the delayed coordinate vector is: ;

[0013] Where i is the index of the data point in the phase space, ranging from 1 to N, and N = n - (mj - 1) × mj, which is the number of valid data points after phase space reconstruction. Delay coordinate vector Y j(i) has a dimension of mj. Phase space reconstruction maps a one-dimensional time series to a set of data point trajectories in an mj-dimensional phase space, and establishes a phase space with a dimension of N×mj.

[0014] Furthermore, a covariance matrix C is constructed on the reconstructed phase space, and the elements C of the covariance matrix C of the j-th sensor channel are... pq The covariance between the p-th and q-th dimensions is expressed by the following formula: , This represents the mean of the N phase points in the p-th dimension. Let Y(i) represent the mean of the N phase points in the q-th dimension. p Let Y(i) represent the p-th dimension coordinate value of the i-th phase point. q Let represent the q-th dimension coordinate value of the i-th phase point. Perform eigenvalue decomposition on the covariance matrix C to obtain mj eigenvalues ​​λ1, λ2, λ3, ..., λmj. Arrange the eigenvalues ​​in descending order as λ1≥λ2≥λ3≥……≥λmj.

[0015] The variance contribution rate of each principal component is calculated, and the first pc principal components are extracted so that the cumulative contribution rate meets the preset condition. pc is the final number of principal components extracted. The extracted pc-dimensional principal components are used as the compressed feature vector of the sensor channel. The dimension of the feature vector is reduced from the original mj dimensions to pc dimensions, thus achieving effective data compression.

[0016] Furthermore, in step 4, the establishment of synchronous samples involves associating and fusing the image observation group with the sensor features within the corresponding time window to form complete sample data that can be used for subsequent joint degradation modeling.

[0017] Let the image at time t0 be the initial image and the image at time t0+Δt be the control image. For the initial image frame, the extracted feature vector is denoted as FIX, and for the control image, the extracted feature vector is denoted as FIy. The difference between FIX and FIy is calculated to obtain the image change feature vector ΔFI = FIy - FIX.

[0018] The processing of sensor features requires integrating the compressed features of the three sensor channels. Let the compressed features of the temperature sensor channel be Stemp, the compressed features of the pressure sensor channel be Spress, and the compressed features of the ultrasonic sensor channel be Sultra. The three feature vectors are concatenated to obtain the fused sensor feature vector S.

[0019] Furthermore, in step 4, since the image feature vector ΔFI and the fused sensor feature vector S come from different data domains and have different numerical distribution ranges and physical dimensions, normalization processing is required to eliminate the influence of dimensional differences.

[0020] The formula for normalizing image feature vectors is: Inorm = (ΔFI - Imin) / (Imax - Imin);

[0021] Where Imin represents the minimum value of the image feature vector, Imax represents the maximum value of the image feature vector, and Inorm represents the normalized image feature vector. The minimum and maximum values ​​are calculated based on historically accumulated sample statistics and are updated periodically to adapt to changes in data distribution. The formula for normalizing the sensor feature vector is: Snorm = (S - Smin) / (Smax - Smin);

[0022] Where Smin represents the minimum value of the sensor feature vector, Smax represents the maximum value of the sensor feature vector, and Snorm represents the normalized sensor feature vector.

[0023] After normalization, the normalized image feature vector Inorm is concatenated with the sensor feature vector Snorm to obtain the joint feature vector Fconcat, whose dimension is the sum of the dimensions of Inorm and Snorm. The formula for vector concatenation is: Fconcat = [Inorm, Snorm]. The joint feature vector Fconcat integrates image appearance change information and sensor internal state information, and serves as the input feature for subsequent spatiotemporal fusion modeling.

[0024] Furthermore, in step 4, adaptive weighted fusion based on temporal distance is used to handle the fusion problem of multiple sets of synchronous samples. Since multiple sets of synchronous samples accumulate during long-term pipeline monitoring, and the collection times of different samples vary from the current time, their ability to represent the current degradation state also differs. Samples closer to the current time contain richer degradation information, while samples farther away have relatively lower reference value. Therefore, a weighted fusion method based on temporal distance is adopted, dynamically allocating fusion weights according to the temporal position of the synchronous samples.

[0025] Let the reference time be tnow, and the acquisition time of the u-th synchronous sample be tu. Define the time difference Δtu = tnow - tu, where Δtu represents the time interval between the u-th synchronous sample and the present. The formula for calculating the fusion weight is: w u =exp(-λ×Δtu;

[0026] Where λ is a preset decay coefficient used to control the rate at which the weight decays over time, and a typical value for λ is 0.1 / unit of time. u is the time weight of the u-th synchronized sample, and its value ranges from 0 to 1. The closer the sample is to the current time, the greater the weight, and the farther the sample is from the current time, the smaller the weight.

[0027] Adaptive weighted fusion based on temporal distance is used to handle the fusion problem of multiple sets of synchronous samples. During long-term pipeline monitoring, the system accumulates multiple sets of synchronous samples. The different samples are collected at varying distances from the current moment, and their ability to represent the current degradation state also differs. Samples closer to the current moment contain richer degradation information and more accurately reflect the current state; samples farther from the current moment have relatively lower reference value, and their information utility gradually decays over time. Therefore, the system adopts a weighted fusion method based on temporal distance, dynamically assigning fusion weights according to the temporal location of the synchronous samples.

[0028] The weighted fusion formula for multiple sets of synchronized samples is: ;

[0029] Where M is the total number of synchronous samples, F con,u Let F be the joint feature vector of the u-th synchronized sample. fu The features of the fused synchronous samples serve as input for subsequent joint degradation modeling.

[0030] Furthermore, step 5 employs a joint degradation modeling method based on multi-task learning, which achieves deep fusion of image features and sensor features through a deep neural network architecture, and predicts the degradation risk value of the pipeline.

[0031] The joint degradation model's overall architecture comprises four core layers: a shared feature extraction layer, an image feature independent task output layer, a sensor feature independent task output layer, and a bilinear fusion layer. The shared feature extraction layer employs a multilayer perceptron structure to learn the common degradation patterns of image and sensor features, capturing the underlying degradation modes shared between the two types of heterogeneous data. The independent task output layer sets independent output branches for image and sensor features respectively, learning the private degradation features of each data type. The bilinear fusion layer deeply interacts and fuses the image and sensor outputs through bilinear operations, generating more expressive fused features. The final output layer predicts pipeline degradation risk values ​​based on the fused features.

[0032] The specific structure of the shared feature extraction layer is a two-layer fully connected network.

[0033] The first fully connected layer has a weight matrix of W1, a bias vector of b1, and uses the ReLU activation function. The output of the first fully connected layer is calculated as H1 = ReLU(W1 × F). fu +b1).

[0034] The weight matrix of the second fully connected network is W2, the bias vector is b2, and the final output of the shared feature extraction layer, i.e. the shared feature vector, is calculated as: Fsh=ReLU(W2×H1+b2).

[0035] The shared feature vector Fsh encodes the common degradation information between image features and sensor features, and is a key input for subsequent bilinear fusion.

[0036] The image feature-independent task output layer is used to predict image degradation metrics based on shared features.

[0037] Let Wimage be the weight matrix of the independent output layer and bimage be the bias vector. The calculation formula for the output of the image feature independent task is: Oimage = Wimage × Fsh + bimage; where Oimage is used to characterize the degree of degradation of the pipe surface reflected by the image observation.

[0038] An independent task output layer based on sensor features is used to predict sensor degradation metrics based on shared features.

[0039] Let Wsensor be the weight matrix of the independent output layer, and Fsh be the bias vector. The formula for calculating the independent task output of the sensor features of bsensor is: Osensor = Wsensor × Fsh + bsensor; where Osensor is used to characterize the degree of degradation of the internal state of the pipe reflected by the sensor data.

[0040] The core operation of the bilinear fusion layer achieves deep interactive fusion of image output and sensor output through matrix multiplication.

[0041] Set the fusion scale K, preferably 16, 32, 64 or other integer powers of 2;

[0042] For the k-th layer, the bilinear fusion vector fk is calculated, where k is an integer from 1 to K. Wb k Let blin be the bilinear weight matrix of the k-th layer. k The fusion bias for the k-th layer;

[0043] The bilinearly fused feature vector Fbin is obtained by concatenating the bilinearly fused vectors of the K-layer rows.

[0044] Bilinear fusion operation establishes a correspondence between the image output space and the sensor output space through the Wb matrix, which can capture high-order interaction information between the two types of features.

[0045] The final pipeline degradation risk prediction output is calculated through the output layer. Let the output layer weight matrix be Wout, the bias vector be bout, the activation function be the sigmoid function, and the output calculation formula be: R = σ(Wout × Fbin + bout);

[0046] Where σ represents the Sigmoid activation function, and R is the pipeline degradation risk value, ranging from 0 to 1. The closer the R value is to 1, the higher the pipeline degradation risk; the closer the R value is to 0, the better the pipeline condition.

[0047] A data analysis-based system for predicting the remaining life of ultra-high pressure pipelines includes: a collaborative acquisition module, an image comparison module, a data compression module, a feature fusion module, and a joint modeling module.

[0048] The collaborative acquisition module is used to deploy acquisition units in the preset detection area of ​​the EVA device to collaboratively acquire image data of the pipe surface and multi-dimensional monitoring data of the sensor group, so as to realize the spatiotemporal synchronous acquisition of the two types of data.

[0049] The image comparison module is used to acquire images of the pipe surface at the initial time and subsequent time, calculate the similarity between two frames of images, and determine that there is an appearance difference when the similarity is lower than a preset threshold. The initial time image and the subsequent time image are paired to construct an image observation group.

[0050] The data compression module is used to extract the raw sensor data within a preset time window, and to perform dimensionality reduction and compression on the data through phase space reconstruction and principal component analysis to generate feature vectors that characterize the temporal characteristics of the sensor.

[0051] The feature fusion module is used to associate and fuse the image observation group with the sensor feature vectors within the corresponding time window. Through feature extraction, normalization, vector concatenation and time-weighted fusion, synchronous sample features are generated.

[0052] The joint modeling module is used for spatiotemporal feature fusion and joint degradation modeling based on synchronous sample features. It constructs a network architecture that includes a shared feature extraction layer, an image feature independent task output layer, a sensor feature independent task output layer, and a bilinear fusion layer, and outputs pipeline degradation risk values.

[0053] Furthermore, the image comparison module includes: a time-series image acquisition unit, a similarity determination unit, and an observation group construction unit; the time-series image acquisition unit is used to acquire pipe surface images at the initial time and subsequent time according to a preset sampling frequency, and establish the correspondence between images and timestamps; the similarity determination unit is used to calculate the similarity between two frames of images acquired in the same detection area, compare the similarity with a preset similarity threshold, and determine that there is an apparent difference when the similarity is lower than the preset threshold.

[0054] The observation group construction unit is used to pair the initial time-series images that are determined to have apparent differences with the subsequent time-series images to construct an image observation group that includes region coding, timestamps, and image data.

[0055] Furthermore, the data compression module includes a phase space reconstruction unit and a principal component analysis unit. The phase space reconstruction unit is used to reconstruct the phase space of the sensor raw data arranged in chronological order. It uses the autocorrelation function method to solve for the optimal delay time. The delay corresponding to the first decrease of the autocorrelation function value to its maximum value at a preset attenuation ratio is determined as the optimal delay time. Based on the optimal delay time and the optimal embedding dimension, a delay coordinate vector is constructed to map the one-dimensional time series into the phase point trajectory in the multi-dimensional phase space, and a phase space matrix with the dimension being the product of the number of phase points and the embedding dimension is constructed.

[0056] The principal component analysis unit is used to construct a covariance matrix based on the phase space matrix, perform eigenvalue decomposition on the covariance matrix, obtain each principal component and its corresponding eigenvalue, extract the top several principal components whose cumulative contribution rate reaches the preset contribution rate threshold, and use them as compressed feature vectors.

[0057] The image feature extraction unit is used to extract features from two frames of images in the image observation group, calculate the difference between the two feature vectors, and obtain the image change feature vector.

[0058] Furthermore, the feature fusion module includes an image feature extraction unit, a sensor feature stitching unit, and a fusion processing unit; the sensor feature stitching unit is used to stitch together the compressed features of each channel of the sensor to obtain a fused sensor feature vector.

[0059] The fusion processing unit is used to normalize the image change feature vector and the fused sensor feature vector respectively, and combine them with a time weighting strategy to complete feature fusion and form the final synchronous sample features.

[0060] Furthermore, the joint modeling module includes a shared feature extraction unit, an independent task output unit, a bilinear fusion unit, and a risk output unit. The shared feature extraction unit processes the features of synchronous samples using a two-layer fully connected network structure. The output of the first fully connected network is processed by an activation function and then input into the second fully connected network to output a shared feature vector. The independent task output unit includes an image feature independent task output subunit and a sensor feature independent task output subunit, which output the image feature independent task results and the sensor feature independent task results respectively based on the shared feature vector through matrix operations. The bilinear fusion unit performs matrix multiplication on the transpose of the image feature independent task output results and the sensor feature independent task output results using a bilinear weight matrix, and then adds a fusion bias term to obtain the fused features.

[0061] The risk output unit is used to calculate and output the pipeline degradation risk value based on the fusion features through an activation function.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] 1. By organically combining image significant change recognition and sensor data compression technology, based on the judgment of whether there are significant changes in the image, only the image observation group containing effective information about changes is retained. At the same time, phase space reconstruction and principal component analysis compression are performed on the sensor data to transform massive sensor data into low-dimensional feature vectors, achieving effective alignment and correlation of the two types of data on the time scale.

[0064] 2. The adaptive weighted fusion strategy based on time distance uses an exponential decay function to calculate the time weight of each sample, making full use of the pipeline degradation patterns contained in historical samples, and dynamically allocating fusion weights according to time position to improve the accuracy of joint degradation modeling.

[0065] 3. By capturing the underlying degradation patterns shared between the two types of heterogeneous data, a correspondence between the image output space and the sensor output space is established, enabling high-level interactive fusion of the two types of data and fully integrating image appearance change information and sensor internal state information. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating a data analysis-based method for predicting the remaining life of ultra-high pressure pipelines according to the present invention.

[0067] Figure 2 This is a schematic diagram of the structure of an ultra-high pressure pipeline remaining life prediction system based on data analysis according to the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example: Figures 1-2 As shown, the present invention provides a technical solution: a method for predicting the remaining life of ultra-high pressure pipelines based on data analysis.

[0070] The method includes five core processing steps: Step 1: setting the detection area and coordinating data acquisition; Step 2: identifying significant changes in the image and constructing the image observation group; Step 3: acquiring sensor data in the time interval; Step 4: establishing synchronous samples; and Step 5: spatiotemporal fusion and joint degradation modeling.

[0071] In step 1, during the setting of the detection area and coordinated data acquisition, the EVA device's ultra-high pressure pipeline system has several key hydrodynamic features along its length. The pipe walls at these locations experience complex stress distributions and media erosion, making them sensitive areas for performance degradation. Specifically, the detection areas include three typical types: valve locations, tee locations, and elbow locations. Each detection area corresponds to a unique area code and spatial coordinate parameters. The area code uses a structured coding scheme, such as a "location-location type-serial number" format. The spatial coordinate parameters are represented in a three-dimensional Cartesian coordinate system, including the coordinates of the detection area's center point and the boundary parameters of the coverage area, providing a reference for subsequent image acquisition and positioning calibration.

[0072] In step 2, the image significant change recognition and image observation group construction, the system periodically acquires digital images of the pipe surface at preset time intervals and intelligently filters valid samples based on the degree of change in image content. The specific operation process is as follows: The system performs image acquisition tasks at fixed time intervals Δt, acquiring the image at time t0 as the initial image and the image at time t0+Δt as the control image. For each acquired original image frame, it first undergoes correction and enhancement processing through an image preprocessing module, including distortion correction, contrast enhancement, and noise filtering, to eliminate the impact of changes in acquisition conditions on image quality. Subsequently, a pre-trained deep convolutional neural network model is used to extract features from the preprocessed image, obtaining a feature image representation for each image. This feature image is a multi-dimensional numerical vector that encodes the texture information and structural features of the pipe surface. After obtaining the feature representations of the initial and control images, the similarity between the two feature images is calculated using the cosine similarity metric.

[0073] In step 3, which involves acquiring sensor data within the time interval, the multi-channel time series data from the sensor within the [t0, t0+Δt] time window is compressed to transform the original data into low-dimensional feature vectors, enabling effective correlation and fusion with the image observation group. This step comprises two core processing stages: the first stage is time series feature extraction based on phase space reconstruction, and the second stage is data dimensionality reduction and compression based on principal component analysis.

[0074] In the first stage of phase space reconstruction processing, for the raw time series data X of the j-th sensor channel in a certain detection region... j = (x 1j x 2j , ..., x nj (where n is the total number of data points). First, the optimal delay time τj and the optimal embedding dimension mj need to be determined. The optimal delay time τj is determined using the autocorrelation function method, which calculates the autocorrelation function of the time series. Where τ is the delay, R(τ) decreases as τ increases, and the value of τ corresponding to the first decrease of R(τ) to R(0) / e is determined as the optimal delay time τj. The optimal embedding dimension mj is determined by the pseudo-nearest neighbor method. This method determines the optimal dimension by calculating the pseudo-nearest neighbor ratio of adjacent points in the embedding space. The minimum dimension corresponding to the pseudo-nearest neighbor ratio being lower than a preset threshold is the optimal embedding dimension mj.

[0075] In a preferred embodiment of the present invention, the criterion for determining the optimal delay time using the autocorrelation function method is as follows: when R(τ) first drops to 1 / e times R(0), the value of τ is determined as the optimal delay time τj, where 1 / e is approximately equal to 0.3679, that is, when the value of the autocorrelation function decays to 36.79% of its maximum value, the corresponding delay amount is the optimal delay time parameter.

[0076] The basic principle of the pseudo-nearest neighbor method is to examine the changes in the relationship between adjacent data points in the phase space under different embedding dimensions. For each candidate embedding dimension m, the distance between all phase points and their nearest neighbors is calculated, and each nearest neighbor is determined to be a pseudo-nearest neighbor. The determination criterion is as follows: For the i-th phase point in the phase space and its nearest neighbor i', the distance dm(i,i') between the two phase points in the m-dimensional space is calculated, as well as the distance dm+1(i,i') after extending the phase point to the m+1 dimension. If dm+1(i,i') / dm(i,i') is greater than a set threshold, with a reference value of 2, then the nearest neighbor i' is determined to be a pseudo-nearest neighbor. The proportion of pseudo-nearest neighbors among all phase points is counted as the pseudo-nearest neighbor rate for that embedding dimension m. Within a given embedding dimension search range, such as 2 to 20, the embedding dimension is gradually increased, and the corresponding pseudo-nearest neighbor rate is calculated. When the pseudo-nearest neighbor rate first drops below the preset threshold, the embedding dimension is determined as the optimal embedding dimension mj. In a preferred embodiment of the present invention, the threshold of the pseudo-nearest neighbor method is set to 0.1, that is, the minimum embedding dimension corresponding to the pseudo-nearest neighbor rate being less than 10% is the optimal embedding dimension.

[0077] Based on the optimal delay time τj and the optimal embedding dimension mj, construct the delay coordinate vector Y. j (i) The formula for constructing the delayed coordinate vector is: ;

[0078] Where i is the index of the data point in the phase space, ranging from 1 to N, and N = n - (mj - 1) × mj, which is the number of valid data points after phase space reconstruction. Delay coordinate vector Y j (i) has a dimension of mj. Phase space reconstruction maps a one-dimensional time series to a set of data point trajectories in an mj-dimensional phase space, and establishes a phase space with a dimension of N×mj.

[0079] In the second stage of data dimensionality reduction and compression processing, a covariance matrix C is constructed on the reconstructed phase space, and the elements C of the covariance matrix C of the j-th sensor channel are... pq The covariance between the p-th and q-th dimensions is expressed by the following formula: , This represents the mean of the N phase points in the p-th dimension. Let Y(i) represent the mean of the N phase points in the q-th dimension. p Let Y(i) represent the p-th dimension coordinate value of the i-th phase point. q Let represent the q-th dimension coordinate value of the i-th phase point. Perform eigenvalue decomposition on the covariance matrix C to obtain mj eigenvalues ​​λ1, λ2, λ3, ..., λmj. Arrange the eigenvalues ​​in descending order as λ1≥λ2≥λ3≥……≥λmj.

[0080] The variance contribution rate of each principal component is calculated, and the first pc principal components are extracted to ensure that the cumulative contribution rate meets a preset condition, typically set to a cumulative contribution rate of 85% or higher. pc represents the final number of extracted principal components. The extracted pc-dimensional principal components are used as the compressed feature vector for that sensor channel. The dimension of the feature vector is reduced from the original mj dimensions to pc dimensions, achieving effective data compression. For a sensor group containing temperature, pressure, and ultrasound sensors, the phase space reconstruction and principal component analysis are performed independently for each channel, ultimately yielding the compressed feature vectors for each of the three sensor channels.

[0081] In step 4, the system establishes synchronous samples, which associates and fuses the image observation group with the sensor features within the corresponding time window. Through a series of operations such as feature extraction, normalization, vector concatenation, and time-weighted fusion, the system forms fused synchronous sample features, which serve as input data for subsequent joint degradation modeling.

[0082] The process of extracting the image change feature vector is as follows: Let the image at time t0 be the initial image, and the image at time t0+Δt be the control image. For the initial image frame, the same deep convolutional neural network model as in step 2 is used to extract the feature vector, denoted as FIX; for the control image frame, the same feature extraction network is used to extract the feature vector, denoted as FIy. Both FIX and FIy are high-dimensional feature vectors, and their dimensions depend on the network structure design, typically 512 or 1024 dimensions. The difference between FIX and FIy is calculated to obtain the image change feature vector ΔFI = FIy - FIX. The difference operation is performed element-wise, and the resulting difference vector encodes the change information between the two images. Positive values ​​indicate the enhancement direction of the control image relative to the initial image, and negative values ​​indicate the weakening direction of the control image relative to the initial image.

[0083] The sensor feature integration process is as follows: Let the compressed feature of the temperature sensor channel be Stemp, the compressed feature of the pressure sensor channel be Spress, and the compressed feature of the ultrasonic sensor channel be Sultra. The feature vectors of the three sensor channels may have different dimensions, depending on the degree of variation of the original data of each channel and the number of principal components extracted by principal component analysis. The three feature vectors are concatenated to obtain the fused sensor feature vector S. The vector concatenation adopts the method of connecting the first and last features, arranged in the order of temperature feature, pressure feature, and ultrasonic feature. The dimension of the fused sensor feature vector S is the sum of the dimensions of the three channel features.

[0084] Since the image feature vector ΔFI and the fused sensor feature vector S originate from different data domains and have different numerical distribution ranges and physical dimensions, direct subsequent fusion will affect the model's learning performance. Therefore, normalization processing is required to eliminate the influence of dimensional differences. The normalization processing adopts the min-max normalization method, which linearly maps the original values ​​to a preset numerical range.

[0085] The formula for normalizing image feature vectors is Inorm = (ΔFI - Imin) / (Imax - Imin), where Imin represents the minimum component of the image feature vector, Imax represents the maximum component, and Inorm represents the normalized image feature vector. The minimum and maximum values ​​are calculated based on historically accumulated sample statistics and are updated periodically to adapt to changes in data distribution. This normalization method maps each component of the image feature vector to a numerical range between zero and one.

[0086] The formula for normalizing the sensor feature vector is Snorm = (S - Smin) / (Smax - Smin), where Smin represents the minimum component of the sensor feature vector, Smax represents the maximum component, and Snorm represents the normalized sensor feature vector. After normalization, the normalized image feature vector Inorm and the sensor feature vector Snorm are concatenated. The concatenation formula is Fconcat = Inorm, Snorm, and the dimension of the joint feature vector Fconcat is the sum of the dimensions of Inorm and Snorm. The joint feature vector Fconcat integrates image appearance change information and sensor internal state information, serving as the input feature for subsequent spatiotemporal fusion modeling.

[0087] Let the reference time be tnow, and the acquisition time of the u-th synchronous sample be tu. Define the time difference Δtu = tnow - tu, where Δtu represents the time interval between the u-th synchronous sample and the present. The formula for calculating the fusion weight is: w u=exp(-λ×Δtu;

[0088] Where λ is a preset decay coefficient used to control the rate at which the weight decays over time, and a typical value for λ is 0.1 / unit of time. u is the time weight of the u-th synchronized sample, and its value ranges from 0 to 1. The closer the sample is to the current time, the greater the weight, and the farther the sample is from the current time, the smaller the weight.

[0089] The weighted fusion formula for multiple sets of synchronized samples is: ;

[0090] Where M is the total number of synchronous samples, F con,u Let F be the joint feature vector of the u-th synchronized sample. fu The features of the fused synchronous samples serve as input for subsequent joint degradation modeling.

[0091] In step 5, the spatiotemporal fusion and joint degradation modeling, the system adopts a joint degradation modeling method based on multi-task learning. It achieves deep fusion of image features and sensor features through a deep neural network architecture and predicts the degradation risk value of the pipeline.

[0092] The joint degradation model's overall architecture comprises four core layers: a shared feature extraction layer, an image feature independent task output layer, a sensor feature independent task output layer, and a bilinear fusion layer. The shared feature extraction layer employs a multilayer perceptron structure to learn the common degradation patterns of image and sensor features, capturing the underlying degradation modes shared between the two types of heterogeneous data. The independent task output layer sets independent output branches for image and sensor features respectively, learning the private degradation features of each data type. The bilinear fusion layer deeply interacts and fuses the image and sensor outputs through bilinear operations, generating more expressive fused features. The final output layer predicts pipeline degradation risk values ​​based on the fused features.

[0093] The specific structure of the shared feature extraction layer is a two-layer fully connected network.

[0094] Let F be the feature of the fused synchronous samples. fu The dimension is din, the weight matrix of the first fully connected network is W1, the bias vector is b1, and the activation function is ReLU. The output of the first fully connected network is calculated as H1 = ReLU(W1 × F). fu +b1).

[0095] The weight matrix of the second fully connected network is W2, the bias vector is b2, and the final output of the shared feature extraction layer, i.e. the shared feature vector, is calculated as: Fsh=ReLU(W2×H1+b2).

[0096] The shared feature vector Fsh encodes the common degradation information between image features and sensor features, and is a key input for subsequent bilinear fusion.

[0097] The image feature-independent task output layer is used to predict image degradation metrics based on shared features.

[0098] Let Wimage be the weight matrix of the independent output layer and bimage be the bias vector. The calculation formula for the output of the image feature independent task is: Oimage = Wimage × Fsh + bimage; where Oimage is used to characterize the degree of degradation of the pipe surface reflected by the image observation.

[0099] An independent task output layer based on sensor features is used to predict sensor degradation metrics based on shared features.

[0100] Let Wsensor be the weight matrix of the independent output layer, and Fsh be the bias vector. The formula for calculating the independent task output of the sensor features of bsensor is: Osensor = Wsensor × Fsh + bsensor; where Osensor is used to characterize the degree of degradation of the internal state of the pipe reflected by the sensor data.

[0101] The core operation of the bilinear fusion layer achieves deep interactive fusion of image output and sensor output through matrix multiplication.

[0102] Set the fusion scale K, preferably 16, 32, 64 or other integer powers of 2;

[0103] For the k-th layer, the bilinear fusion vector fk is calculated, where k is an integer from 1 to K. Wb k Let blin be the bilinear weight matrix of the k-th layer. k The fusion bias for the k-th layer;

[0104] The bilinearly fused feature vector Fbin is obtained by concatenating the bilinearly fused vectors of the K-layer rows.

[0105] Bilinear fusion operation establishes a correspondence between the image output space and the sensor output space through the Wb matrix, which can capture high-order interaction information between the two types of features.

[0106] The final pipeline degradation risk prediction output is calculated through the output layer. Let the output layer weight matrix be Wout, the bias vector be bout, the activation function be the sigmoid function, and the output calculation formula be: R = σ(Wout × Fbin + bout);

[0107] Where σ represents the Sigmoid activation function, and R is the pipeline degradation risk value, ranging from 0 to 1. The closer the R value is to 1, the higher the pipeline degradation risk; the closer the R value is to 0, the better the pipeline condition.

[0108] In a certain detection area, a temperature sensor collected 12,000 data points (n=12,000) at a sampling frequency of 100 Hz within a time window [t0, t0+Δt]. The autocorrelation function of this time series data was calculated as follows: when the delay τ=10, the autocorrelation function value R(10) equals 0.68R(0); when the delay τ=15, the autocorrelation function value R(15) equals 0.38R(0), which is lower than the threshold of 0.3679R(0) for the first time. Therefore, the optimal delay time τj is determined to be 15 sampling intervals, i.e., 0.15 seconds.

[0109] The pseudo-nearest neighbor method was applied to determine the optimal embedding dimension for this time series. When the embedding dimension m=5, the pseudo-nearest neighbor rate was 0.15; when the embedding dimension m=6, the pseudo-nearest neighbor rate was 0.08, which was lower than the threshold of 0.1 for the first time. Therefore, the optimal embedding dimension mj was determined to be 6. The number of effective data points N was calculated as N=n-(mj-1)×τj=12000-(6-1)×15=11925, and the matrix dimension after phase space reconstruction was 11925×6.

[0110] A 6×6 covariance matrix is ​​constructed on the reconstructed phase space. After eigenvalue decomposition, six eigenvalues ​​are obtained: λ1=45.3, λ2=28.7, λ3=15.2, λ4=8.6, λ5=3.4, and λ6=1.8, with a sum of 103.0. The variance contribution rates of each principal component are calculated to be 43.98%, 27.86%, 14.76%, 8.35%, 3.30%, and 1.75%, respectively. The cumulative contribution rate of the first four principal components reaches 95.95%, exceeding the 85% threshold. Therefore, the first four principal components are extracted as compressed feature vectors, reducing the feature dimension from 6 to 4.

[0111] In the bilinear fusion layer, the output Oimage of the image feature-independent task is designed to be 32-dimensional, the output Osensor of the sensor feature-independent task is designed to be 32-dimensional, and the fusion scale K is set to 16. Then, each bilinear weight matrix Wb... k The dimension is 32×32, and the bilinear fusion operation is performed. The calculation result is a 32-dimensional vector. The 16 layers of bilinear fusion vectors are concatenated to obtain a bilinear fusion feature vector Fbin with dimensions of 16 × 32 = 512. Finally, the pipeline degradation risk value R calculated by the output layer is mapped to the zero-to-one range using the Sigmoid function. When the R value exceeds a preset risk threshold, such as 0.7, a warning is output, indicating that the pipeline's current degradation risk is at a high level, requiring further detailed inspection or maintenance.

[0112] The model training employs a supervised learning method, utilizing a historically accumulated pipeline monitoring dataset. The dataset's sample structure remains consistent with the inference phase; each sample contains a set of image observations, sensor data for the corresponding time window, and manually labeled degradation risk level tags. Tags are obtained based on historical detection results and maintenance records. For example, when a pipeline subsequently fails or is repaired / replaced, the tags for the corresponding historical samples are assigned based on the failure time and severity.

[0113] The goal of model training is to minimize the error between the predicted output and the true label, using mean squared error loss as the loss function. In each training epoch, the training data is first divided into several batches, and the fused batches are synchronized with the sample features F. fu The shared feature extraction layer is used to calculate the shared feature vector Fsh. Then, Oimage and Osensor are calculated through the image feature independent task output layer and the sensor feature independent task output layer, respectively. Next, the bilinear fusion layer is used to calculate the fused feature vector Fbin. Finally, the predicted degradation risk value Rpred is calculated through the output layer. Based on the loss function value between the predicted value and the true label, the gradient of the loss function with respect to the parameters of each layer is calculated. The gradient is then propagated from the output layer to the bilinear fusion layer, the two independent task output layers, and the shared feature extraction layer via backpropagation. The Adam optimizer updates the weight matrix and bias vector of each layer based on the calculated gradients. During training, the model performance is periodically evaluated on the validation dataset. If the validation loss does not decrease for 20 consecutive training epochs, early stopping is used to terminate training, and the model parameters with the lowest validation loss are saved as the final model.

[0114] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for predicting the remaining life of ultra-high pressure pipelines based on data analysis, characterized in that: A detection area is set up in the EVA device to collaboratively collect image data and sensor data. Through the collaborative data acquisition of the image acquisition unit and the sensor group, the synchronous acquisition of pipe surface images and sensor data is achieved. Images from the initial moment and images from subsequent moments are acquired. The similarity between the two images is compared. If the similarity is lower than a preset similarity threshold, it is determined that there is a difference. The images from the initial moment and the images from subsequent moments are then combined into an image observation group. The sensor data within the preset time window is compressed, and the sensor data is transformed into feature vectors through phase space reconstruction and principal component analysis. The image observation group is associated and fused with the sensor features within the corresponding time window. The fused synchronous sample features are formed through feature extraction, normalization, vector concatenation and time-weighted fusion. Spatiotemporal fusion and joint degradation modeling are performed based on synchronous sample features. The joint degradation modeling includes a network architecture consisting of a shared feature extraction layer, an image feature independent task output layer, a sensor feature independent task output layer, and a bilinear fusion layer, which outputs pipeline degradation risk values.

2. The method for predicting the remaining life of ultra-high pressure pipelines based on data analysis according to claim 1, characterized in that: Methods for collaboratively acquiring image data and sensor data include: The detection areas are set to include valve locations, tee locations, and elbow locations. Each detection area corresponds to a unique area code and coordinate parameters. Through collaborative data acquisition by an image acquisition device and a sensor group, the surface image of the pipeline and sensor data are acquired synchronously. The sensor group includes a temperature sensor, a pressure sensor, and an ultrasonic sensor.

3. The method for predicting the remaining life of ultra-high pressure pipelines based on data analysis according to claim 1, characterized in that: Methods for compressing sensor data include: The original sensor data arranged in chronological order is reconstructed, and the optimal delay time in the phase space is obtained by using the autocorrelation function method. The delay corresponding to the first drop of the autocorrelation function value to a preset attenuation ratio of its maximum value is the optimal delay time. The minimum dimension corresponding to the pseudo-nearest neighbor ratio being lower than a preset threshold is the optimal embedding dimension. A covariance matrix is ​​constructed on the reconstructed phase space. The elements of the covariance matrix represent the covariance relationship between different dimensions in the phase space. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain the principal components and their corresponding eigenvalues ​​for each dimension. The eigenvalues ​​are arranged in descending order, and the variance contribution rate of each principal component is calculated. The top few principal components that make the cumulative contribution rate reach or exceed the preset contribution rate threshold are extracted as compressed feature vectors.

4. The method for predicting the remaining life of ultra-high pressure pipelines based on data analysis according to claim 3, characterized in that: Based on the optimal delay time and optimal embedding dimension, a delay coordinate vector is constructed to map a one-dimensional time series into a set of data point trajectories in the phase space. The dimension of the delay coordinate vector is the embedding dimension, and after the phase space is reconstructed, a phase space with a dimension equal to the number of phase points multiplied by the embedding dimension is established.

5. The method for predicting the remaining life of ultra-high pressure pipelines based on data analysis according to claim 1, characterized in that: Methods for correlating and fusing image observation groups with sensor features within corresponding time windows include: Feature extraction is performed on each image frame in the image observation group, and the difference between the two feature vectors is calculated to obtain the image change feature vector. The compressed features of the sensor channels are concatenated to obtain the fused sensor feature vector. Then, the fused sensor feature vector and the image change feature vector are normalized and concatenated.

6. The method for predicting the remaining life of ultra-high pressure pipelines based on data analysis according to claim 1, characterized in that: Joint degradation modeling methods include: The shared feature extraction layer adopts a two-layer fully connected network structure. The output of the first fully connected network is processed by an activation function and then input into the second fully connected network to obtain the shared feature vector. The image feature independent task output layer obtains the image feature independent task output based on the shared feature vector through matrix operations. The sensor feature independent task output layer obtains the sensor feature independent task output based on the shared feature vector through matrix operations. The bilinear fusion layer performs matrix operations on the transposes of the image feature independent task output and the sensor feature independent task output through a bilinear weight matrix and adds a fusion bias to obtain the fused feature. Finally, the output layer calculates the pipeline degradation risk value based on the fused feature through an activation function.

7. A data analysis-based system for predicting the remaining life of ultra-high pressure pipelines, used to execute the data analysis-based method for predicting the remaining life of ultra-high pressure pipelines as described in any one of claims 1-6, characterized in that: The system includes: The system includes a collaborative acquisition module, an image comparison module, a data compression module, a feature fusion module, and a joint modeling module. The collaborative acquisition module is used to deploy acquisition units in the preset detection area of ​​the EVA device to collaboratively acquire image data of the pipe surface and monitoring data of the sensor group, so as to realize the spatiotemporal synchronous acquisition of the two types of data. The image comparison module is used to acquire images of the pipe surface at the initial time and subsequent time, calculate the similarity between the two frames of images, and determine that there is an apparent difference when the similarity is lower than a preset threshold. The initial time image and the subsequent time image are paired to construct an image observation group. The data compression module is used to extract the original sensor data within a preset time window, and to perform dimensionality reduction and compression on the data through phase space reconstruction and principal component analysis to generate feature vectors characterizing the temporal characteristics of the sensor. The feature fusion module is used to associate and fuse the image observation group with the sensor feature vectors within the corresponding time window. After feature extraction, normalization, vector concatenation and time-weighted fusion, synchronous sample features are generated. The joint modeling module is used to perform spatiotemporal feature fusion and joint degradation modeling based on the synchronous sample features, and to construct a network architecture that includes a shared feature extraction layer, an image feature independent task output layer, a sensor feature independent task output layer and a bilinear fusion layer, and outputs pipeline degradation risk values.

8. The data analysis-based system for predicting the remaining life of ultra-high pressure pipelines according to claim 7, characterized in that: The image comparison module includes: a time-series image acquisition unit, a similarity determination unit, and an observation group construction unit; The time-series image acquisition unit is used to acquire pipe surface images at the initial and subsequent times according to a preset sampling frequency, and to establish the correspondence between images and timestamps; The similarity determination unit is used to calculate the similarity between two frames of images collected in the same detection area, compare the similarity with a preset similarity threshold, and determine that there is an apparent difference when the similarity is lower than the preset threshold. The observation group construction unit is used to pair the initial time-series images that are determined to have apparent differences with the subsequent time-series images to construct an image observation group that includes region coding, timestamps, and image data.

9. A data analysis-based system for predicting the remaining life of ultra-high pressure pipelines according to claim 7, characterized in that: The data compression module includes a phase space reconstruction unit and a principal component analysis unit; the feature fusion module includes an image feature extraction unit, a sensor feature stitching unit, and a fusion processing unit. The phase space reconstruction unit is used to reconstruct the phase space of the sensor raw data arranged in time sequence. The optimal delay time is determined by the autocorrelation function method. The delay amount corresponding to the first drop of the autocorrelation function value to its maximum value at a preset attenuation ratio is determined as the optimal delay time. Based on the optimal delay time and the optimal embedding dimension, a delay coordinate vector is constructed to map the one-dimensional time series into the phase point trajectory in the phase space, and a phase space matrix with the dimension of the product of the number of phase points and the embedding dimension is constructed. The principal component analysis unit is used to construct a covariance matrix based on the phase space matrix, perform eigenvalue decomposition on the covariance matrix, obtain each principal component and its corresponding eigenvalue, extract the top several principal components whose cumulative contribution rate reaches the preset contribution rate threshold, and use them as compressed feature vectors. The image feature extraction unit is used to extract features from two frames of images in the image observation group, calculate the difference between the two feature vectors, and obtain the image change feature vector. The sensor feature stitching unit is used to stitch together the compressed features of each channel of the sensor to obtain a fused sensor feature vector; The fusion processing unit is used to normalize the image change feature vector and the fusion sensor feature vector respectively, and combine them with a time weighting strategy to complete feature fusion and form the final synchronous sample features.

10. A data analysis-based system for predicting the remaining life of ultra-high pressure pipelines according to claim 7, characterized in that: The joint modeling module includes a shared feature extraction unit, an independent task output unit, a bilinear fusion unit, and a risk output unit; The shared feature extraction unit is used to process the features of synchronous samples using a two-layer fully connected network structure. The output of the first fully connected network is processed by an activation function and then input into the second fully connected network to output a shared feature vector. The independent task output unit includes an image feature independent task output subunit and a sensor feature independent task output subunit, which respectively output the image feature independent task result and the sensor feature independent task result based on the shared feature vector through matrix operations. The bilinear fusion unit is used to perform matrix multiplication on the transpose of the output results of the image feature independent task and the sensor feature independent task through the bilinear weight matrix, and then superimpose the fusion bias term to obtain the fused features. The risk output unit is used to calculate and output the pipeline degradation risk value based on the fusion features through an activation function.

Images