Method for predicting residual life of buried pipeline

Abstract

The invention discloses a buried pipeline residual life prediction method, relates to the technical field of pipeline prediction in computer measurement, can realize multi-step prediction of the residual life of a buried pipeline, and is small in accumulated error and higher in precision. According to the buried pipeline residual life prediction method, a training set is input into a TFT neural network for network training, a test set is input into the trained TFT neural network for testing, TFT model parameters are stored, and training of the neural network is completed; the TFT neural network comprises a gating residual network GRN module used for processing a nonlinear correlation relation between input data and a prediction target, a variable selection network module used for carrying out important feature screening on static and dynamic variables, and a multi-head attention layer module used for learning a long-term dependency relation of a pipeline life sequence. And a long and short term memory network (LSTM) module, and a quantile regression module for outputting the residual life of the pipeline based on regression of different quantiles.

Description

Technical Field

[0001] The invention relates to the technical field of pipeline prediction in computer measurement, and in particular to a method for predicting the remaining life of a buried pipeline. Background Art

[0002] At present, buried pipelines, as important means of transporting fluids and gases, play a very important role in fields related to national economy and people's livelihood, such as petroleum and petrochemicals, natural gas, water supply and drainage.

[0003] However, due to the complex soil environment outside the buried pipeline and the large span of the laying area, it is very easy for the pipeline to fail, and even cause major safety accidents. Therefore, it is particularly important to accurately predict the remaining life of the buried pipeline, and its health status is of great significance in ensuring the safe operation of the pipeline.

[0004] In the prior art, the research methods for the remaining life of pipelines at home and abroad are mainly divided into two categories: model-driven and data-driven. Model-driven prediction of the remaining life of pipelines is mainly carried out by constructing empirical or semi-empirical mechanism models of corrosion rate. However, due to the complexity and diversity of corrosion factors affecting pipeline life, and the impact of pipeline operation status on its own degradation is not considered, it has certain limitations. The data-driven method predicts its expected life by mining the behavioral characteristics implicit in the historical operation data of the pipeline and combining the current pipeline status. It mainly includes two categories: probabilistic statistical methods and intelligent algorithms. Among them, the probabilistic statistical method mainly uses Monte Carlo, Markov, Bayesian and other methods to estimate the remaining life by calculating the probability distribution of pipeline data. The prediction of pipeline life based on deep learning algorithms is still in the exploratory stage, and the prediction method is basically a single-step output, which will inevitably produce cumulative errors during recursive use. Summary of the invention

[0005] The object of the present invention is to provide a method for predicting the remaining life of a buried pipeline, which can realize multi-step prediction of the remaining life of the buried pipeline with smaller cumulative error and higher accuracy.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme: A method for predicting the remaining life of a buried pipeline comprises the following steps: Data acquisition: Collect the data accumulated from on-site pipeline inspections in recent years and the soil environmental data along the pipeline as the original data set for model training and testing; Variable division: The factors that affect the remaining life of the pipeline are divided into static variables and time-varying variables according to their relationship with time; static variables are variables that do not change with time, including pipeline material, location, medium flow rate, medium temperature, and medium sulfur content; time-varying variables are variables that change with time, including medium particle content, soil moisture content, soil salt content, and soil pH value; Training the Temporal Fusion Transformer (TFT) neural network: The original data set is divided into a training set of 80% and a test set of 20% in proportion, the training set is input into the TFT neural network for network training, the test set is input into the trained TFT neural network for testing, and the TFT model parameters are saved to complete the training of the neural network; wherein, the TFT neural network includes: a gated residual network GRN module, a variable selection network module, a multi-head attention layer module, a long short-term memory network LSTM module, and a quantile regression module; the gated residual network GRN module uses the nonlinear correlation between the input data and the prediction target for processing; the variable selection network module is used to screen important features of static and dynamic variables; the multi-head attention layer module is used to learn the long-term dependency of the pipeline life sequence; the quantile regression module outputs the remaining life of the pipeline based on the regression of different quantiles; Pipeline life prediction: The trained TFT neural network is used to operate on the preprocessed pipeline data and output the predicted pipeline life value. Specifically, by setting the model history time step number m and the prediction time step number n, the pipeline life information of the past m years is input into the TFT model to predict the remaining life of the pipeline n years in the future. At the same time, the three quantiles of each prediction point, 10%, 50% and 90%, are regressed to the mean as the output of the point.

[0007] In practical applications, after data acquisition, data expansion is also included: analyzing the distribution characteristics of the data set, and using Euclidean distance and angle as similarity criteria to expand the number of data sets to improve the generalization ability of the model and the accuracy of prediction.

[0008] Among them, the gated residual network GRN module includes: input layer, fully connected layer and output layer; the input layer linearly transforms the initial input vector and context by multiplying the ELU function, and activates it through the Sigmoid function; the fully connected layer performs global average pooling on the output vector of the input layer, randomly inactivates some neurons, and outputs a 1x1xC long vector; the output layer normalizes and sums the output of the fully connected layer.

[0009] Specifically, the variable selection network module includes: a GRN function and a Softmax function; the GRN function performs nonlinear transformation on the input vector; the Softmax function normalizes and maps the output vector of the nonlinear transformation of the GRN function into a sparse weight; the output vector of the GRN function is multiplied by the sparse weight to generate a set of transformed features, and these features are added to obtain the output vector.

[0010] Furthermore, a multi-head attention layer module is used to capture the long-term dependencies between pipeline life and variables, including: an initialization function and a forward transfer function; the initialization function accepts a dropout parameter and a scale parameter, the dropout parameter specifies how much dropout to apply before softmax, and the scale parameter specifies whether the dot product result should be scaled; the forward transfer function accepts a q parameter, a k parameter, a v parameter, and a mask parameter, the q parameter, the k parameter, and the v parameter are input vectors of queries, keys, and values, and the mask parameter is an optional masking vector and is used to mask certain positions in the input.

[0011] Optionally, the forward transfer function is calculated as follows: Calculate the dot product of the q parameter and the k parameter to get the attention matrix attention. If the module is initialized to scaling mode, the result will be divided by the square root of the length of the k parameter vector when calculating the dot product to reduce the value range of the dot product result to avoid gradient vanishing or gradient exploding. Determine whether the mask parameter vector is an empty vector. If the mask parameter is not an empty vector, fill -1e in the attention. 9 , so as to produce a probability of 0 in the softmax operation; The attention matrix attention is normalized into a probability distribution through the softmax function, and the dropout operation is applied to prevent overfitting; The output vector output is obtained by dot product calculation; At output, the forward function returns the output vector output and the attention matrix attention.

[0012] Furthermore, the LSTM module is used to capture the short-term dependency between pipeline life and variables, including input gate, forget gate and output gate logic, and contains two states: cell state and hidden state; the cell state is a relatively stable memory of the entire time series, and the hidden state focuses on the output of the current time step and the previous time step; the input gate is used to control how much of the current input data of the network is saved; the forget gate is used to control which information to keep and which information to forget, in order to avoid gradient vanishing and gradient exploding phenomena; the output gate is used to calculate the final output information; Preferably, the forget gate receives h(t-1) and the current input data x(t) as input parameters, performs dot product calculation with the sigmoid function, and obtains the forget gate output; the input gate receives the output of the forget gate and the current input data x(t), performs dot product with the sigmoid function, and then multiplies with the tanh function to obtain updated state data; the output gate performs dot product with the sigmoid function, and then normalizes to [-1, 1] through the tanh function to obtain a new cell state.

[0013] Furthermore, the quantile regression module is used to calculate the difference between the predicted sequence and the target sequence, and set the 10%, 50% and 90% quantiles, and calculate the weighted error based on the quantiles, and then return a tensor representing the quantized loss.

[0014] Compared with the prior art, the method for predicting the remaining life of buried pipelines described in the present invention has the following advantages: In the method for predicting the remaining life of a buried pipeline provided by the present invention, a prediction model for the remaining life of a buried pipeline is built based on a time domain multivariate fusion Transformer network (Temporal Fusion Transformer, TFT), and factors affecting the pipeline life, such as medium flow velocity, medium temperature, medium sulfur content, medium particulate matter content, soil moisture content, soil salt content, soil pH value, pipeline material, pipeline pressure, pipeline position, etc., are divided into static variables, dynamic time-varying variables, and dynamic time-invariant variables. Then, the data sequence features are screened based on a gated residual network (GRN), and short-term and long-term dependencies are captured based on a long short-term memory (LSTM) network and a multi-head attention mechanism, respectively. Finally, the remaining life of the pipeline is predicted through quantile regression. From this analysis, it can be seen that the remaining life prediction method of buried pipelines provided by the present invention constructs a remaining life prediction model for buried pipelines based on the TFT network, realizes automatic feature screening by directly inputting historical data, and quickly obtains prediction results; captures short- and long-term dependencies through the LSTM network and the multi-head attention mechanism, and when performing multi-step prediction of the remaining life of the pipeline, the cumulative error is smaller and the accuracy is higher; at the same time, in order to avoid the problem of model overfitting caused by too few pipeline life data samples, the data expansion method is used to improve the accuracy of the model; in addition, in actual application, the time step can be adjusted according to different types of pipelines, and the network weight parameters can be trained separately, so that the model can be applied to the prediction of the remaining life of other types of pipelines. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a process flow of a method for predicting the remaining life of a buried pipeline provided by an embodiment of the present invention; Figure 2A schematic diagram of the structure of a TFT neural network in a method for predicting the remaining life of a buried pipeline provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a gated residual network GRN module of a TFT neural network in a method for predicting the remaining life of a buried pipeline provided in an embodiment of the present invention; Figure 4 A schematic diagram of the structure of a variable selection network module of a TFT neural network in a method for predicting the remaining life of a buried pipeline provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a multi-head attention layer module of a TFT neural network in a method for predicting the remaining life of a buried pipeline provided in an embodiment of the present invention; Figure 6 A schematic diagram of the structure of a long short-term memory network (LSTM) module of a TFT neural network in a method for predicting the remaining life of a buried pipeline provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] For ease of understanding, the method for predicting the remaining life of a buried pipeline provided by an embodiment of the present invention is described in detail below in conjunction with the accompanying drawings.

[0017] The present invention provides a method for predicting the remaining life of a buried pipeline. Figure 1-Figure 6 As shown, the following steps are included: Step S100: data acquisition: Collect the data accumulated from on-site pipeline inspections in recent years, as well as the soil environment data along the pipeline as the original data set for model training and testing (for example, collect the pipeline wall thickness monitoring data of a certain buried pipeline for nearly 20 years, and retrieve the soil environment data along the pipeline to construct 1,200 sets of samples as the original data set for model training and testing); Step S200: Data expansion: Analyze the distribution characteristics of the data set, and use Euclidean distance and angle as similarity criteria to expand the number of data sets to improve the generalization ability of the model and the accuracy of prediction; Step S300: variable division: The factors that affect the remaining life of the pipeline are divided into static variables and time-varying variables according to their relationship with time; static variables are variables that do not change with time, including pipeline material, location, medium flow rate, medium temperature, and medium sulfur content; time-varying variables are variables that change with time, including medium particle content, soil moisture content, soil salt content, and soil pH value; Step S400: training the Temporal Fusion Transformer (TFT) neural network: The original data set is divided into a training set of 80% and a test set of 20% in proportion, the training set is input into the TFT neural network for network training, the test set is input into the trained TFT neural network for testing, and the TFT model parameters are saved to complete the training of the neural network; wherein, the TFT neural network includes: a gated residual network GRN module, a variable selection network module, a multi-head attention layer module, a long short-term memory network LSTM module, and a quantile regression module; the gated residual network GRN module uses the nonlinear correlation between the input data and the prediction target for processing; the variable selection network module is used to screen important features of static and dynamic variables; the multi-head attention layer module is used to learn the long-term dependency of the pipeline life sequence; the quantile regression module outputs the remaining life of the pipeline based on the regression of different quantiles; Step S500: Pipeline life prediction: The trained TFT neural network is used to operate on the preprocessed pipeline data and output the predicted pipeline life value; specifically, by setting the model historical time step number m and the prediction time step number n, the pipeline life information of the past m years is input into the TFT model to predict the remaining pipeline life n years in the future, and the three quantiles 10%, 50% and 90% regression mean of each prediction point are used as the output of the point (for example, the pipeline wall thickness data and surrounding soil environment data collected for the first 20 years of a new pipeline are input into the trained TFT neural network, and the prediction step is set to 20, that is, the pipeline life information of the past 20 years is used to predict the remaining pipeline life of the next 20 years; the TFT neural network operates on the new pipeline data, and the TFT neural network outputs the three quantiles 10%, 50% and 90% regression values ​​at each prediction step according to the operation results, and the mean is used as the output of the prediction step).

[0018] Compared with the prior art, the method for predicting the remaining life of a buried pipeline according to the embodiment of the present invention has the following advantages: In the method for predicting the remaining life of a buried pipeline provided by an embodiment of the present invention, a prediction model for the remaining life of a buried pipeline is built based on a temporal multivariate fusion Transformer network (Temporal Fusion Transformer, TFT), and factors affecting the pipeline life, such as medium flow velocity, medium temperature, medium sulfur content, medium particulate matter content, soil moisture content, soil salt content, soil pH value, pipeline material, pipeline pressure, and pipeline position, are divided into static variables, dynamic time-varying variables, and dynamic time-invariant variables. Then, the data sequence features are screened based on a gated residual network (GRN), and short-term and long-term dependencies are captured based on a long short-term memory (LSTM) network and a multi-head attention mechanism, respectively. Finally, the remaining life of the pipeline is predicted through quantile regression. From this analysis, it can be seen that the remaining life prediction method for buried pipelines provided by the embodiment of the present invention constructs a remaining life prediction model for buried pipelines based on the TFT network, realizes automatic feature screening by directly inputting historical data, and quickly obtains prediction results; captures short-term and long-term dependencies through the LSTM network and the multi-head attention mechanism, and when performing multi-step prediction of the remaining life of the pipeline, the cumulative error is smaller and the accuracy is higher; at the same time, in order to avoid the problem of model overfitting caused by too few pipeline life data samples, the data expansion method is used to improve the accuracy of the model; in addition, in actual application, the time step can be adjusted according to different types of pipelines, and the network weight parameters can be trained separately, so that the model can be applied to the prediction of the remaining life of other types of pipelines.

[0019] Among them, Figure 3 As shown in FIG. 1 , the gated residual network GRN module may include: an input layer, a fully connected layer, and an output layer; the input layer linearly transforms the initial input vector and the context (if any) by multiplying the ELU function, and activates it through the Sigmoid function; the fully connected layer performs global average pooling on the output vector of the input layer, and randomly inactivates some neurons, and outputs a 1x1xC long vector; the output layer normalizes and sums the output of the fully connected layer.

[0020] Specifically, Figure 4 As shown, the variable selection network module may include: a GRN function and a Softmax function; the GRN function performs nonlinear transformation on the input vector; the Softmax function normalizes and maps the output vector of the nonlinear transformation of the GRN function into a sparse weight; the output vector of the GRN function is multiplied by the sparse weight to generate a set of transformed features, and these features are added to obtain the output vector.

[0021] Furthermore, if Figure 5As shown, the multi-head attention layer module can be used to capture the long-term dependencies between pipeline life and variables, and may include: an initialization function and a forward transfer function; the initialization function accepts a dropout parameter and a scale parameter, the dropout parameter specifies how much dropout to apply before softmax, and the scale parameter specifies whether the dot product result should be scaled; the forward transfer function accepts a q parameter, a k parameter, a v parameter, and a mask parameter, the q parameter, the k parameter, and the v parameter are input vectors of queries, keys, and values, and the mask parameter is an optional masking vector and is used to mask certain positions in the input.

[0022] Optionally, the calculation steps of the forward transfer function can be as follows: Step S1, calculate the dot product of the q parameter and the k parameter to obtain the attention matrix attention. If the module is initialized to the scaling mode, the result will be divided by the square root of the length of the k parameter vector when calculating the dot product to reduce the value range of the dot product result to avoid gradient vanishing or gradient exploding phenomena; Step S2: Determine whether the mask parameter vector is an empty vector. If the mask parameter is not an empty vector, fill -1e in the attention. 9 , so as to produce a probability of 0 in the softmax operation; Step S3, normalize the attention matrix attention to a probability distribution through the softmax function, and apply the dropout operation to prevent overfitting; Step S4, obtaining an output vector output by dot product calculation; Step S5: At output, the forward function returns the output vector output and the attention matrix attention.

[0023] Furthermore, if Figure 6 As shown in the figure, the LSTM module can be used to capture the short-term dependency between pipeline life and variables, and can include: input gate, forget gate and output gate logic, and can contain two states: cell state and hidden state; the cell state is a relatively stable memory of the entire time series, and the hidden state focuses on the output of the current time step and the previous time step; the input gate is used to control how much current input data the network saves; the forget gate is used to control which information to keep and which information to forget, so as to avoid gradient disappearance and gradient explosion; the output gate is used to calculate the final output information; Preferably, the forget gate receives h(t-1) and the current input data x(t) as input parameters, performs dot product calculation with the sigmoid function, and obtains the forget gate output; the input gate receives the output of the forget gate and the current input data x(t), performs dot product with the sigmoid function, and then multiplies with the tanh function to obtain updated state data; the output gate performs dot product with the sigmoid function, and then normalizes to [-1, 1] through the tanh function to obtain a new cell state.

[0024] Furthermore, the quantile regression module can be used to calculate the difference between the predicted sequence and the target sequence, set the 10%, 50% and 90% quantiles, calculate the weighted error based on the quantiles, and then return a tensor representing the quantized loss.

[0025] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting the remaining life of a buried pipeline, characterized in that: The following steps are involved: Data acquisition: Collect the data accumulated from on-site pipeline inspections in recent years and the soil environmental data along the pipeline as the original data set for model training and testing; Variable division: The factors that affect the remaining life of the pipeline are divided into static variables and time-varying variables according to their relationship with time; static variables are variables that do not change with time, including pipeline material, location, medium flow rate, medium temperature, and medium sulfur content; time-varying variables are variables that change with time, including medium particle content, soil moisture content, soil salt content, and soil pH value; Training the Temporal Fusion Transformer neural network: The original data set is divided into a training set of 80% and a test set of 20% in proportion, the training set is input into the TFT neural network for network training, the test set is input into the trained TFT neural network for testing, and the TFT model parameters are saved to complete the training of the neural network; wherein, the TFT neural network includes: a gated residual network GRN module, a variable selection network module, a multi-head attention layer module, a long short-term memory network LSTM module, and a quantile regression module; the gated residual network GRN module uses the nonlinear correlation between the input data and the prediction target for processing; the variable selection network module is used to screen important features of static and dynamic variables; the multi-head attention layer module is used to learn the long-term dependency of the pipeline life sequence; the quantile regression module outputs the remaining life of the pipeline based on the regression of different quantiles; Pipeline life prediction: The trained TFT neural network is used to operate on the preprocessed pipeline data and output the predicted pipeline life value. Specifically, by setting the model history time step number m and the prediction time step number n, the pipeline life information of the past m years is input into the TFT model to predict the remaining life of the pipeline n years in the future. At the same time, the three quantiles of each prediction point, 10%, 50% and 90%, are regressed to the mean as the output of the point.

2. The method for predicting the remaining life of a buried pipeline according to claim 1, characterized in that: After data acquisition, data expansion is also included: analyzing the distribution characteristics of the data set, and expanding the number of data sets using Euclidean distance and angle as similarity criteria.

3. The method for predicting the remaining life of a buried pipeline according to claim 1 or 2, characterized in that: The gated residual network GRN module includes: input layer, fully connected layer and output layer; the input layer linearly transforms the initial input vector and context by multiplying the ELU function, and activates it through the Sigmoid function; the fully connected layer performs global average pooling on the output vector of the input layer, randomly inactivates some neurons, and outputs a 1x1xC long vector; the output layer normalizes and sums the output of the fully connected layer.

4. The method for predicting the remaining life of a buried pipeline according to claim 3, characterized in that: The variable selection network module includes: GRN function and Softmax function; GRN function performs nonlinear transformation on input vector; Softmax function normalizes and maps the output vector of nonlinear transformation of GRN function into sparse weight; multiplies the output vector of GRN function with sparse weight to generate a set of transformed features, and adds these features to obtain output vector.

5. The method for predicting the remaining life of a buried pipeline according to claim 4, characterized in that: The multi-head attention layer module is used to capture the long-term dependencies between pipeline life and variables, including: initialization function and forward transfer function; the initialization function accepts dropout parameter and scale parameter, the dropout parameter specifies how much dropout to apply before softmax, and the scale parameter specifies whether the dot product result should be scaled; the forward transfer function accepts q parameter, k parameter, v parameter and mask parameter, q parameter, k parameter and v parameter are input vectors of query, key and value, and mask parameter is an optional mask vector and is used to mask certain positions in the input.

6. The method for predicting the remaining life of a buried pipeline according to claim 5, characterized in that: The calculation steps of the forward transfer function are as follows: Calculate the dot product of the q parameter and the k parameter to get the attention matrix attention. If the module is initialized to scaling mode, the result will be divided by the square root of the length of the k parameter vector when calculating the dot product to reduce the value range of the dot product result. Determine whether the mask parameter vector is an empty vector. If the mask parameter is not an empty vector, fill -1e in the attention. 9 , so as to produce a probability of 0 in the softmax operation; The attention matrix attention is normalized into a probability distribution through the softmax function, and the dropout operation is applied to prevent overfitting; The output vector output is obtained by dot product calculation; At output, the forward function returns the output vector output and the attention matrix attention.

7. The method for predicting the remaining life of a buried pipeline according to claim 5, characterized in that: The LSTM module of the long short-term memory network is used to capture the short-term dependency between the pipeline life and the variables, including: input gate, forget gate and output gate logic, and contains two states: cell state and hidden state; the cell state is a relatively stable memory of the entire time series, and the hidden state focuses on the output of the current time step and the previous time step; the input gate is used to control how much current input data the network saves; the forget gate is used to control which information to retain and which information to forget; the output gate is used to calculate the final output information.

8. The method for predicting the remaining life of a buried pipeline according to claim 7, characterized in that: The forget gate receives h(t-1) and the current input data x(t) as input parameters, performs dot product calculation with the sigmoid function, and obtains the forget gate output; the input gate receives the output of the forget gate and the current input data x(t), performs dot product with the sigmoid function, and then multiplies it with the tanh function to obtain the updated state data; the output gate performs dot product with the sigmoid function, and then normalizes it to [-1, 1] through the tanh function to obtain the new cell state.

9. The method for predicting the remaining life of a buried pipeline according to claim 7, characterized in that: The quantile regression module is used to calculate the difference between the predicted sequence and the target sequence, set the 10%, 50%, and 90% quantiles, calculate the weighted error based on the quantiles, and then return a tensor representing the quantized loss.

Images