CN-122021286-A - Method for predicting residual life of corrosion-prone pipeline based on wall thickness monitoring data

CN122021286ACN 122021286 ACN122021286 ACN 122021286ACN-122021286-A

Abstract

The invention belongs to the technical field of data processing, and particularly relates to a method for predicting the residual life of a corrosion-prone pipeline based on wall thickness monitoring data. The method comprises the steps of firstly collecting ultrasonic wall thickness time sequence monitoring data of a pipeline, judging and checking data through instantaneous gradient and local density in a dual mode, combining anchor point vectors and global corrosion attenuation reference curves to complement blank values, extracting derivative features such as corrosion increment, average corrosion rate, wall thickness fluctuation variance and the like, splicing the derivative features into feature vectors, then inputting an improved bidirectional time sequence attention gating circulating unit to extract fusion time sequence features, finally inputting the fusion features into an improved LSTM network with a time sequence warranty forgetting gate, and outputting a residual life prediction result of the pipeline. The method improves the adaptability of the data quality and the model to the corrosion time sequence rule, improves the prediction accuracy and stability, and provides support for the safe operation of the industrial pipeline.

Inventors

WEI BAO
WEI CI
ZHANG YUGE
Fu Maorui
SONG KE

Assignees

鲁西科安特种设备检测有限公司

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (9)

1. The method for predicting the residual life of the corrosion-prone pipeline based on the wall thickness monitoring data is characterized by comprising the following steps of: S1, collecting wall thickness time series monitoring data of ultrasonic measurement of an easy-to-erode pipeline, performing data verification, and then complementing the deleted abnormal data and the blank data of the abnormal data to obtain data after data verification and complementation; S2, extracting derivative features of the wall thickness time sequence from the data after data verification and complementation, wherein the derivative features comprise corrosion increment of each time point, average corrosion rate in a sliding window and wall thickness fluctuation variance; S3, constructing an improved bidirectional time sequence attention gate control circulation unit, inputting the obtained feature vector into the network, and outputting a fusion time sequence feature; and S4, finally, inputting the output fusion time sequence characteristics into an improved LSTM network for training, and outputting a final pipeline life prediction result.
2. The method for predicting the residual life of a corrosion-prone pipe based on wall thickness monitoring data according to claim 1, wherein the implementation of the data verification in step S1 includes: firstly, calculating the instantaneous gradient and local density of wall thickness time sequence data, constructing a double abnormal judgment basis, and obtaining the wall thickness time sequence data at any time point Instantaneous gradient Characterizing the rate of wall thickness change at adjacent time points: , wherein, For the wall thickness values at adjacent points in time, For monitoring the time interval; Taking out The front and back n data form a local window, and the local density of the data in the window is calculated The calculation mode is as follows: wherein j is an index, For the point in time The corresponding pipe wall thickness monitors the original value, For the point in time The corresponding pipe wall thickness monitors the original value, Respectively the minimum value of the maximum values of all wall thickness values in the local window; If the calculated gradient value is larger than the set gradient abnormality threshold value and the local density of the data in the window is smaller than the set local density threshold value, judging the data as an abnormal value, and if only a single condition is met, marking the data as a suspicious point; Respectively setting for suspicious points And Respectively calculating the instantaneous gradient and the local density of the suspicious point under the two windows, respectively carrying out normalization operation, calculating the cooperative matching degree of the suspicious point for each window, and quantifying the abnormal deviation degree of the suspicious point in the window: , wherein, Respectively normalizing the instantaneous gradient and the local density; Calculating suspicious points at Global window Collaborative matching degree difference of fine window: wherein And respectively determining the matching degree of the global window and the matching degree of the fine window, if the difference value is more than 0.2, reserving data, and if the difference value is less than or equal to 0.2, deleting abnormal data.
3. The method for predicting the residual life of the corrosion-prone pipeline based on wall thickness monitoring data according to claim 2, wherein after data verification, the deleted abnormal data and the empty data of the abnormal data need to be complemented, and is specifically implemented as follows: For missing segments To the point of Constructing anchor point vectors according to the neighborhood effective data to generate initial filling values, and extracting the wall thickness of the first 2 effective points , Corrosion increment Local fluctuation variance Obtaining the wall thickness of the two effective points after the missing , Corrosion increment Local fluctuation variance ; Calculating an initial filling value: , wherein, For the missing time point Is used to determine the initial filling value of (c), As a weighting coefficient, it is realized as: , wherein, The average corrosion increment and fluctuation variance of the whole time sequence are shown; fitting a global corrosion decay reference curve based on the full time sequence data, and correcting the initial filling value by the reference curve: , wherein, For the global corrosion decay reference curve, Local average wall thickness of effective data of the missing segment neighborhood; For a pair of Constructing a sliding window containing k data points before and after each, and weighting each data point in the window and the weight of each data point to the window Is inversely proportional to the time distance of (a), the weight formula is: , wherein, Normalizing the weights for any time point in the window to ensure that the sum of the weights is 1, and normalizing the weights of the data in the window The weighted summation yields the final fill value.
4. The method for predicting the residual life of a corrosion-prone pipeline based on wall thickness monitoring data according to claim 1, wherein the constructing an improved bidirectional time sequence attention gating cycle unit in the step S3 comprises a bidirectional GRU time sequence primitive layer, an in-gate embedded time sequence attention mechanism layer and a bidirectional characteristic mutual feedback coupling layer; The bidirectional GRU time sequence primitive layer is used for completing bidirectional time sequence feature basic extraction of wall thickness feature vectors; The in-gate embedded time sequence attention mechanism layer is used for being embedded into a gating core of the bidirectional GRU time sequence primitive layer to realize synchronous completion of feature screening and time sequence attention distribution; the bidirectional characteristic mutual feedback coupling layer is used for establishing a real-time interaction channel of bidirectional GRU forward and reverse characteristics, and realizing deep complementation and coupling fusion of the characteristics.
5. The method for predicting the residual life of a corrosion-prone pipeline based on wall thickness monitoring data as claimed in claim 4, wherein the specific implementation of the bidirectional GRU time sequence primitive layer for completing bidirectional time sequence feature basic extraction of wall thickness feature vectors is to input the obtained single-time-step feature vectors into a forward GRU branch according to time positive sequence to obtain a forward hiding state at a t-th moment Simultaneously, inputting the single time step feature vector into a backward GRU branch according to the time reverse order to obtain a backward hidden state at the t moment 。
6. The method for predicting the residual life of a corrosion-prone pipeline based on wall thickness monitoring data according to claim 4, wherein the in-gate embedded time sequence attention mechanism layer is used for being embedded and integrated into a gating core of the bidirectional GRU time sequence primitive layer, and the implementation of synchronous completion of feature screening and time sequence attention distribution is as follows: Calculating a time-series attention score at the gating core for the resulting forward and backward hidden states The score characterizes the importance degree of the characteristics at the t moment to the corrosion trend of the pipeline, and the calculation mode is as follows: , wherein, For the feature projection vector of the attention mechanism, In order to project the matrix of the light, As a result of the bias term, Performing feature vector splicing operation; Normalizing the attention scores of all time steps to obtain normalized attention weights Ensuring the sum of weights to be 1; will normalize the attention weights And carrying out feature recalibration on the forward and backward hidden features in the gating unit to obtain the attention-enhanced bidirectional hidden features, wherein the formula is as follows: wherein For the enhanced attention bi-directional concealment feature in the door at time t, Is Hadamard product.
7. The method for predicting the residual life of an easy-to-erode pipeline based on wall thickness monitoring data according to claim 4, wherein the bidirectional characteristic mutual feedback coupling layer is used for establishing a real-time interaction channel of bidirectional GRU forward and backward characteristics, and the specific implementation of realizing depth complementation and coupling fusion of the characteristics is as follows: the forward hidden characteristic is input into the backward branch for characteristic correction, and meanwhile, the backward hidden characteristic is input into the forward branch for characteristic correction, so that bidirectional mutual feedback of the characteristics is realized, and the calculation mode is as follows: , wherein, Respectively a forward characteristic and a backward characteristic after mutual feedback correction, Is a feature coupling matrix; coupling and splicing the features subjected to mutual feedback correction and the attention enhancement features obtained in the previous step to obtain coupling enhancement features: ; And (3) performing time sequence feature aggregation on the coupling enhancement features of all time steps, integrating the discrete single-step features into a complete time sequence feature sequence, and finally outputting a fusion time sequence feature.
8. The method for predicting the residual life of a perishable pipeline based on wall thickness monitoring data according to claim 1, wherein the improved LSTM network is an improved time sequence protection forgetting gate, and is used for carrying out high-weight preservation on historical time sequence characteristics of the network, weakening forgetting duty ratio of invalid redundant information, adapting time sequence accumulation rules of corrosion characteristics and avoiding loss of key time sequence information.
9. The method for predicting the residual life of a corrosion-prone pipeline based on wall thickness monitoring data according to claim 8, wherein the improved time sequence protection forgetting gate is used for carrying out high weight preservation on network history time sequence characteristics, weakening forgetting duty ratio of invalid redundant information, adapting time sequence accumulation rules of corrosion characteristics and avoiding key time sequence information loss, and is specifically realized as follows: Calculating basic forgetting gate weight, and obtaining an original forgetting judgment value: wherein Forgetting the gate weight for the basis of the t-th time step, For the sigmoid activation function, The feature weight matrix is input for the forget gate, For the time-step t to fuse the timing characteristics, Hiding the state weight matrix for the history of the forgetting gate, Hidden state for the network history at time step t-1, Bias for forgetting the door; Then, the weighting correction of the weight-preserving coefficient is carried out to realize the forced constraint of high weight: Wherein, the method comprises the steps of, Forgetting the gate weight for the pre-protection weight of the t time step; Outputting final weight-preserving forgetting weight, and locking a value interval: Wherein, the method comprises the steps of, Forgetting the gate weight for final protection of the t time step; the protection weight acts on the historical cell state to finish the characteristic high-weight retention update: Wherein, the method comprises the steps of, Historical cell status after the protection for the t time step, Network historical cell status for time step t-1; and finally, updating the cell state to completely update the connection formula: Wherein, the method comprises the steps of, The post-selection features for the input gate of time step t, Weights are screened for the input gates of time step t.

Description

Method for predicting residual life of corrosion-prone pipeline based on wall thickness monitoring data Technical Field The invention belongs to the technical field of data processing, and particularly relates to a method for predicting the residual life of a corrosion-prone pipeline based on wall thickness monitoring data. Background The easily corroded pipeline is widely applied to the fields of petrochemical industry, municipal fuel gas, industrial fluid transportation and the like, and when the pipeline is contacted with acid-base medium, humid air or sulfur-containing fluid for a long time, the pipeline wall is damaged by electrochemical corrosion, stress corrosion and the like. The continuous attenuation of the pipe wall thickness can cause safety accidents such as leakage, pipe explosion and the like, and cause economic loss and environmental hazard, so that the residual life of the easily-corroded pipeline is accurately predicted, and the method is a key link for guaranteeing the safety of industrial production and urban operation. The existing method for predicting the residual life of the pipeline has obvious short plates, namely, the wall thickness monitoring data processing mode is rough, the abnormal value judgment is mostly achieved by adopting a single gradient threshold value method, the local data distribution characteristics are not combined, effective data are easy to judge by mistake, the vacancy value filling depends on linear interpolation, and the time sequence change rule of the corrosion rate is ignored, so that the data is distorted. Secondly, the adaptability of the main flow time sequence prediction model is insufficient, the traditional gating circulation unit is difficult to capture nonlinear mutation characteristics of the corrosion process, key historical time sequence information is easy to weaken by a forgetting gate of a long-period memory network, and a corrosion accumulation rule cannot be fitted accurately. These defects result in larger deviation of the predicted result, and it is difficult to meet the requirements of engineering practice for high-precision life prediction. Disclosure of Invention Aiming at the technical problems in the background technology, the invention provides a method for predicting the residual life of a corrosion-prone pipeline based on wall thickness monitoring data. In order to achieve the above purpose, the technical scheme adopted by the invention comprises the following steps: S1, collecting wall thickness time series monitoring data of ultrasonic measurement of an easy-to-erode pipeline, performing data verification, and then complementing the deleted abnormal data and the blank data of the abnormal data to obtain data after data verification and complementation; S2, extracting derivative features of the wall thickness time sequence from the data after data verification and complementation, wherein the derivative features comprise corrosion increment of each time point, average corrosion rate in a sliding window and wall thickness fluctuation variance; S3, constructing an improved bidirectional time sequence attention gate control circulation unit, inputting the obtained feature vector into the network, and outputting a fusion time sequence feature; and S4, finally, inputting the output fusion time sequence characteristics into an improved LSTM network for training, and outputting a final pipeline life prediction result. Preferably, the implementation of the data verification in step S1 includes: firstly, calculating the instantaneous gradient and local density of wall thickness time sequence data, constructing a double abnormal judgment basis, and obtaining the wall thickness time sequence data at any time point Instantaneous gradientCharacterizing the rate of wall thickness change at adjacent time points: , wherein, For the wall thickness values at adjacent points in time,For monitoring the time interval; Taking out The front and back n data form a local window, and the local density of the data in the window is calculatedThe calculation mode is as follows: wherein j is an index, For the point in timeThe corresponding pipe wall thickness monitors the original value,For the point in timeThe corresponding pipe wall thickness monitors the original value,Respectively the minimum value of the maximum values of all wall thickness values in the local window; If the calculated gradient value is larger than the set gradient abnormality threshold value and the local density of the data in the window is smaller than the set local density threshold value, judging the data as an abnormal value, and if only a single condition is met, marking the data as a suspicious point; Respectively setting for suspicious points AndRespectively calculating the instantaneous gradient and the local density of the suspicious point under the two windows, respectively carrying out normalization operation, calculating the cooperative matching degree of the suspicious po