A method for predicting the life of thermal pipelines

Abstract

The present invention proposes a method for predicting the life of a thermal pipeline, which relates to the field of machine learning. The thermal pipeline life prediction process proposed in the present invention includes constructing a data set related to thermal pipelines, capturing the temporal dynamic characteristics of the data through a gated recurrent unit GRU, then capturing data features at different angles through an adaptive feature selection mechanism, and capturing the global characteristics of the data through a hybrid computing module, so as to make full use of key information related to the life of the thermal pipeline in the data, and after all features are weighted and fused to obtain a comprehensive information feature representation, a fully connected neural network is used to calculate the prediction result of the thermal pipeline life, which effectively enhances the accuracy of the model prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting the life of a thermal pipeline. Background Art

[0002] The life of a thermal pipeline refers to the length of time that the thermal pipeline can operate normally without major failures under its designed conditions of use. With the long-term use of thermal pipelines, the pipeline materials will gradually age due to factors such as high temperature, pressure, corrosion, and fatigue, resulting in a decrease in their carrying capacity, which may eventually cause accidents such as pipeline leakage or rupture, bringing serious safety hazards and economic losses. Accurate prediction of the life of thermal pipelines is of great significance, which can help discover potential problems in advance, perform maintenance or replacement in a timely manner, avoid major accidents, and ensure the stable operation of the heating system.

[0003] Current methods in the field of time series prediction mainly include traditional statistical models, machine learning models, and deep learning models. Although these methods have certain advantages in time dependence and trend modeling, these methods usually only process data from a single perspective and find it difficult to comprehensively capture the deep impact of data on pipeline life from multiple perspectives, resulting in a decrease in prediction accuracy and insufficient adaptability to complex situations.

[0004] Thermal pipeline related data contains many key variables that have a significant impact on the life of the pipeline. By fully analyzing these data and capturing the multi-level information implicit in the data from multiple angles, and dynamically adjusting their importance according to the different degrees of influence of each angle on the target variable, we can effectively capture the impact of data on the life of thermal pipelines and enhance the prediction accuracy and robustness of the model when dealing with complex situations. Summary of the invention

[0005] The present invention provides a method for predicting the life of thermal pipelines, which aims to use the gated recurrent unit (GRU) to model the time dependency of data so as to better capture the temporal dynamic characteristics in the thermal pipeline data. Secondly, the multi-level information implicit in the data is captured from multiple angles, and the importance of the information from each angle is dynamically adjusted in combination with an adaptive weight selection mechanism, so as to extract more representative and effective data features. On this basis, the shared information of the data is obtained through a hybrid computing strategy to combine global and local features to further improve the prediction ability of the model. Finally, all data features are weighted and fused to achieve accurate prediction of the life of the thermal pipeline.

[0006] The technical method adopted by the present invention to achieve the above-mentioned purpose specifically comprises the following steps:

[0007] S1. Collect relevant data of thermal pipelines, including temperature, humidity, pressure, corrosion degree of inner and outer surfaces of pipelines, pipeline material, operation time, cycle load, soil pH, and pre-process the collected data to provide a reliable data basis for subsequent thermal pipeline life prediction;

[0008] S2, use feedforward neural network to embed features of preprocessed data and map the input to high-dimensional feature space;

[0009] S3. The temporal dependency of data is further captured through the gated recurrent unit (GRU). GRU uses the reset gate and update gate mechanisms to selectively retain or forget information from long time series, ensuring that the model can process long time series data while retaining important historical information related to thermal pipeline life prediction to accurately model temporal dynamic characteristics.

[0010] S4. Build an adaptive feature selection mechanism, first use multiple feedforward neural networks to capture data features from different angles, and then use adaptive weights to dynamically fuse feature data from different angles to ensure that the impact of key variables on pipeline life is fully modeled;

[0011] S5. Construct a hybrid computing module to capture contextual information and process the commonalities between features at different angles to obtain global features between features at different angles, so that the model can consider global and local features.

[0012] S6. Fuse multiple data information, comprehensively consider the importance of each data feature, and generate a final unique feature representation for predicting the life of thermal pipelines;

[0013] S7. The weighted fused feature representation is further calculated through a fully connected neural network to generate a prediction result for the thermal pipeline life.

[0014] Preferably, in step S1, data related to thermal pipelines are collected, including temperature, humidity, pressure, degree of corrosion of the inner surface of the pipeline, degree of corrosion of the outer surface of the pipeline, pipeline material, operating time, cycle load and soil pH data, and the mean filling method is used to fill in the missing values ​​in the data set.

[0015] Preferably, in step S2, given Historical observation sequence of data ,in is the observation time step, For the Type of data The observation values ​​at each moment are converted into high-dimensional vectors that can be processed by the model through the feedforward neural network using a point-by-point marking method. The specific calculation process is as follows:

[0016] ;

[0017] In the formula for The high-dimensional vector representation obtained by mapping is is the learnable weight parameter matrix, is the ReLU activation function, is the dimension of the data.

[0018] Preferably, compared with the traditional low-dimensional representation, the high-dimensional representation can capture more implicit patterns and structural information. Each observation point in the time series is mapped to a high-dimensional vector representation through a feedforward neural network, which effectively enhances the model's feature extraction capability for sequence data. The use of the ReLU activation function can also avoid the gradient vanishing problem, ensuring that the model maintains a high training efficiency during the deep learning process.

[0019] Preferably, in step S3, a gated recurrent unit GRU is used to capture the dependency of high-dimensional vectors in the time dimension and model the temporal dynamic characteristics. The specific calculation process is as follows:

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] In the formula For the The reset gate for the time step, For the The update gate for the time step, For the Candidate hidden states for time steps, For the The hidden state at the time step, , , , , , is the weight parameter, , , is the bias parameter, is the sigmoid function, It is Hadamard.

[0025] Preferably, by introducing a gated recurrent unit GRU by resetting the gate and update gate The GRU mechanism dynamically adjusts the update of hidden states, avoids the problem of information forgetting or excessive accumulation, and can effectively process historical data with long time spans in thermal pipeline life prediction. Compared with traditional recurrent neural networks, GRU has fewer parameters and higher training efficiency. It can also better handle long-term dependency problems and enhance the ability to capture long-term trends and short-term changes, thereby improving the robustness and accuracy of pipeline life prediction.

[0026] Preferably, in step S4, first construct The feature capture feed-forward neural network at each angle is used to obtain the feature representation of each data at different angles. The specific calculation process is as follows:

[0027] ;

[0028] In the formula For the Type of data Time passes The feature representation obtained by the feed-forward neural network is captured by the feature capture of the angles. To capture the A network with angle features, is the dimension of the data, and then the weight of each angle is calculated, and the feature representations of different angles are dynamically integrated. The specific calculation process is as follows:

[0029] ;

[0030] ;

[0031] In the formula For the The data in The weight distribution of each angle at the moment, is the weight parameter, For the The data in Moment The weight of the angle, For the The data in The final fusion representation of multiple angles at the moment.

[0032] Preferably, a feedforward neural network is used to extract different feature representations of data from multiple angles, which can fully capture the multi-dimensional information in each type of data. Then, the features of different angles are weighted and fused by dynamically calculating the weight of each angle, ensuring that the features of key angles receive more attention in the final representation. This method not only enhances the model's ability to understand complex data structures, but also can adaptively select and combine information from different feature angles, thereby improving the model's prediction accuracy and robustness in different data scenarios.

[0033] Preferably, in step S5, a hybrid computing module is constructed to capture context information and process the commonalities between features at different angles, extract the global feature representation and combine it with the feature representation after multi-angle fusion to obtain the final feature representation. The specific computing process is as follows:

[0034] ;

[0035] ;

[0036] ;

[0037] In the formula For the The data in Global feature representation shared at all times, is a feed-forward neural network, is the weight of the global feature representation, is the weight parameter, For the The data in The final feature representation of the moment.

[0038] Preferably, the global features in the data are effectively captured through the hybrid computing module, which can fully reflect the common information between features from different angles. The introduction of global features provides a macro perspective for the model and helps to capture contextual information. In addition, the Sigmoid function is used to dynamically adjust the weights of global features to ensure that the model can automatically assign the importance of global features according to the data, thereby achieving weighted fusion of global features and multi-angle features, and improving the prediction accuracy and robustness of the model in complex scenarios.

[0039] Preferably, in step S6, firstly, the feature vectors of each data at all time points are weightedly fused to obtain the overall feature vector representation of each data, and then, the feature vectors of all data types are further weightedly fused to obtain the final feature vector for thermal pipeline life prediction. The specific calculation process is as follows:

[0040] ;

[0041] ;

[0042] In the formula is the feature vector used for thermal pipeline life prediction. For the Time series data of the data type, For the Type of data The feature vector at time, is the number of data types, is the time step, For the The time series data weight of the data, For the Type of data The weight of the moment, is the weight parameter.

[0043] Preferably, this step enables the model to capture important time point information in the time series by weighted fusion of feature vectors in the time dimension, and normalizes the time dimension through the Softmax function. The model can dynamically adjust the contribution of each time point, thereby highlighting the data characteristics of key time points, ensuring that information at key moments has a greater influence on model predictions, and helping the model to comprehensively process multi-source data and improve the accuracy and robustness of thermal pipeline life prediction.

[0044] Preferably, in step S7, Through the fully connected neural network, the final predicted value of thermal pipeline life is calculated:

[0045] ;

[0046] In the formula is a fully connected neural network, It is the final predicted value of thermal pipeline life.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention effectively captures dynamic information in the time dimension through the gated recurrent unit GRU, thereby enhancing the model's modeling ability for long-term dependencies in time series; secondly, by extracting potential information in each data from multiple angles, an adaptive weight selection mechanism is introduced to weightedly fuse features from different angles, thereby adaptively highlighting important information; subsequently, shared information between different angles is further extracted through a hybrid computing module, thereby enhancing the model's ability to recognize global data patterns; finally, the model fuses the features of all data to generate a feature vector containing rich contextual information, thereby significantly improving the accuracy and robustness of thermal pipeline life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a step diagram of a thermal pipeline life prediction method.

[0049] Figure 2 It is the GRU diagram of the gated recurrent unit.

[0050] Figure 3 Diagram for the adaptive weight selection mechanism.

[0051] Figure 4 This is a diagram of the hybrid computing module.

[0052] Figure 5 This is the effect diagram of thermal pipeline life prediction. DETAILED DESCRIPTION

[0053] The present invention proposes a method for predicting the life of a thermal pipeline. In this method, the gated recurrent unit GRU effectively captures dynamic information in the time dimension. Secondly, by extracting potential information in each data from multiple angles, the features of different angles are weighted and fused using an adaptive weight selection mechanism. Subsequently, the shared information between different angles is further extracted through a hybrid computing module. Finally, the model fuses the features of all data to generate a feature vector containing rich contextual information, and uses a fully connected network to achieve accurate prediction of the life of the thermal pipeline. The technical solution in the embodiment of the present invention is described in detail and completely below, specifically including the following steps: Figure 1 shown.

[0054] S1. Collect 90 days of thermal pipeline related data, including the following 9 types of data: temperature, humidity, pressure, pipeline inner surface corrosion degree, pipeline outer surface corrosion degree, pipeline material, operation time, cycle load, soil pH. For missing data values, select the data of the 5 days before and after for mean filling. If there are less than 5 days of data before and after under boundary conditions, use the mean of the available days for filling. Then divide the data set into training set and test set in a ratio of 8:2. Next, use the sliding window method. The input data of each day is composed of the thermal pipeline related data of the previous 14 days, which serves as the basis for model learning and prediction, ensuring that the model can capture dynamic changes in the time dimension.

[0055] S2. Using a point-by-point labeling method, a feedforward neural network is used to embed the observations in all the data, thereby mapping the input to a high-dimensional feature space.

[0056] Furthermore, in step S2, for each type of data , given a historical observation sequence ,in is the observation time step, For the Data at a point in time In order to convert the observation value of each time point into a high-dimensional vector that can be processed by the model, a point-by-point labeling method is adopted, and the conversion is realized by using a feedforward neural network. The specific calculation process is as follows:

[0057] ;

[0058] In the formula for The high-dimensional vector representation obtained by mapping is is the learnable weight parameter matrix, is the ReLU activation function.

[0059] S3. The temporal dependency of the data is further captured through the gated recurrent unit (GRU). GRU is an improved recurrent neural network that selectively retains or forgets information from long time series through the mechanism of resetting gates and updating gates, ensuring that in long time series data, the model can effectively process key historical information related to the prediction of thermal pipeline life.

[0060] Further, in step S3, if Figure 2 As shown, the gated recurrent unit GRU is used to capture the dependency of high-dimensional vectors in the time dimension. GRU selectively retains or forgets information by resetting the gate and updating the gate to enhance the model's ability to process time series. The specific calculation process is as follows:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] In the formula For the The reset gate for the time step, For the The update gate for the time step, For the Candidate hidden states for time steps, is the hidden state, , , , , , The parameters learned for model training, , , is the bias parameter, is the sigmoid function, is the Hadamard product, which is the element-wise product.

[0066] S4. Construct an adaptive feature selection mechanism. First, use multiple feedforward neural networks to capture data features from different angles, and use an adaptive weight allocation mechanism to weight the features extracted from each feedforward neural network to form the final fused feature representation.

[0067] Further, in step S4, if Figure 3 As shown, first build A feed-forward neural network to capture The characteristics of the angles are Take an angle as an example, the specific calculation process is as follows:

[0068] ;

[0069] In the formula To pass the The feed-forward neural network obtained by The data in The characteristic representation of the moment, is the dimension of the data, For the The feedforward neural network at each angle is then Generate weights at different angles and dynamically fuse feature representations at different angles. The specific calculation process is as follows:

[0070] ;

[0071] ;

[0072] In the formula For the The data in The weight distribution of each angle at the moment, is the weight parameter, For the The data in Moment The weight of the angle, For the The data in The final fusion representation of multiple angles at the moment.

[0073] S5, build hybrid computing module, use A feedforward neural network captures the commonalities between features at different angles to obtain the global features between features at different angles, so that the model can consider both global and local features at the same time.

[0074] Further, in step S5, if Figure 4 As shown in the figure, a hybrid computing module is constructed to extract global features and dynamically weight them with feature representations fused from different angles. First, a feedforward neural network is used to extract global features, and then an adaptive weight is generated to control the importance of global features in the final feature fusion. The specific calculation process is as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] In the formula For the The data in Global feature representation shared at all times, is the weight of the global feature representation, For the The data in The final feature representation of the moment, is the weight parameter, It is a feed-forward neural network.

[0079] S6. Integrate multiple data information, comprehensively consider the feature information of different data sources, and generate a unique feature representation based on the importance of each data feature for thermal pipeline life prediction.

[0080] Furthermore, in step S6, the weight of the time series feature vector corresponding to each data type is first calculated, and the time series feature vectors are weighted and fused according to the weights. After the fusion of a single data type is completed, the feature vectors of all data types are further weighted and fused, and finally the final feature vector for thermal pipeline life prediction is obtained. The specific calculation process is as follows:

[0081] ;

[0082] ;

[0083] In the formula For the The time series data weight of the data, For the Type of data The weight of the moment, is the weight parameter, is the feature vector used for thermal pipeline life prediction. For the Time series data of the data type, For the Type of data The feature vector at time, is the number of data types, is the time step.

[0084] S7. The weighted fused feature vector is further calculated and processed through a fully connected neural network to generate a prediction result for the thermal pipeline life.

[0085] Furthermore, in step S7, Through the fully connected neural network, the final predicted value of thermal pipeline life is calculated:

[0086] ;

[0087] In the formula is a fully connected neural network, It is the final predicted value of thermal pipeline life.

[0088] Furthermore, the method is implemented in Python 3.8. The model is built and trained based on the PyTorch framework and runs in the CUDA11 environment to make full use of GPU accelerated computing. The entire training process runs on the NVIDIA RTX 3090 GPU. The batch size set during training is 64 and the learning rate is ,use Optimizer.

[0089] Furthermore, the method predicts the effect as Figure 5 As shown, the ordinate is the thermal pipeline life (days), the abscissa is time (days), the gray straight line represents the actual thermal pipeline life, and the black straight line represents the predicted value of the thermal pipeline life. It can be seen from the figure that the prediction result is highly consistent with the true value, indicating that this method has a high degree of fit and accuracy in the prediction of thermal pipeline life, verifying its effectiveness.

[0090] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the life of a thermal pipeline, characterized in that: The following steps are involved: S1. Collect relevant data of thermal pipelines, including 9 types of data: temperature, humidity, pressure, corrosion degree of pipeline inner surface, corrosion degree of pipeline outer surface, pipeline material, operation time, cycle load, soil pH, and pre-process the collected data to provide a reliable data basis for subsequent thermal pipeline life prediction; S2. Use a feedforward neural network to embed features of the preprocessed data by point-by-point labeling, and map all data observations to a high-dimensional feature space. The specific steps are as follows: Input all types of data, taking the i-th data as an example, its historical observation sequence of T time steps is the observed value of the i-th data at the time t∈[1, T]. Through the feedforward neural network, the observed value at each time point is converted into a high-dimensional vector that can be processed by the model by using a point-by-point labeling method. The specific calculation process is as follows: in i，t =ReLU(Wx i，t ); In the formula For x i，t The high-dimensional vector representation obtained by mapping is is the learnable weight parameter matrix, ReLU(·) is the ReLU activation function, and H is the dimension of the data; S3. The long time series data is processed by the gated recurrent unit GRU. GRU uses the reset gate and update gate mechanism to selectively retain or forget information from the long time series, retain important historical information related to the life prediction of thermal pipelines, and accurately model the time dynamic characteristics. The specific steps are as follows: The gated recurrent unit GRU is used to capture the dependency of high-dimensional vectors in the time dimension and model the temporal dynamic features. First, the reset gate and update gate are used to selectively retain or forget information from long time series. The specific calculation process is as follows: r i，t =σ(v i，t W vr +h i，t-1 W hr +b r ); z i，t =σ(v i，t W vz +h i，t-1 W hz +b z ); In the formula is the reset gate at the tth time step, is the update gate for the tth time step, is the weight parameter, b r , is the bias parameter, σ is the sigmoid function, and then the hidden vector of the gated recurrent unit GRU is calculated according to the reset gate and the update gate. The specific calculation process is as follows: Mode is the candidate hidden state at the tth time step, is the hidden state at the tth time step, W vh , is the weight parameter, is the bias parameter, σ is the sigmoid function, ⊙ is the Hadamard product; S4. Build an adaptive feature selection mechanism, use multiple feedforward neural networks to capture the features of data at different angles, and then learn the weights of each angle to fuse the data features at different angles to ensure that the impact of key variables on pipeline life is fully modeled. The specific steps are: First, a feature capture feedforward neural network of N angles is constructed to obtain the feature representation of each data at different angles. Then, the weight of each angle is calculated and the feature representation of different angles is dynamically integrated. The specific calculation process is as follows: u i，t,j =FFN j (h i，t ); s i，t =softmax(W s h i，t ); In the formula FFN is the feature representation obtained by the feature capture network of the jth angle for the i-th data at the tth time. j (·) is the network used to capture the j-th angle feature, D is the dimension of the data, is the weight distribution of each angle of the i-th data at the t-th moment, is the weight parameter, is the weight of the i-th data at the j-th angle at the t-th time, is the final fusion representation of the i-th data at multiple angles at time t; S5, construct a hybrid computing module, use a feedforward neural network to capture the commonalities between features at different angles, so as to obtain the global features between features at different angles, and add the weighted global features to the data features fused at different angles, so that the model can consider both global and local features at the same time. The specific steps are as follows: A hybrid computing module is constructed to capture contextual information and process the global feature representation between features at different angles, and dynamically weighted fuse it with the feature representation after fusion at different angles to obtain the representation of the global and local features of this function. The specific calculation process is as follows: g i,t =FFN(h i，t ); s i，t,g =Sigmoid(W g h i，t ); y i,t =s i，t,g g i,t +f i，t ; In the formula is the global feature representation of the i-th data at time t, FFN(·) is a feedforward neural network, is the weight of the global feature representation, is the weight parameter, is the final feature representation of the i-th data at time t; S6. Fusion of multiple data information, comprehensive consideration of the importance of each data feature, and generation of a final unique feature representation for predicting the life of thermal pipelines. The specific steps are: First, all the final feature representations of the time series data of all data are weighted and fused to obtain the overall feature vector representation of each data. Then, the overall feature vector representations of all data are further fused to obtain a comprehensive feature representation. The specific calculation process is as follows: w i =softmax(y i W i ); In the formula is the time series data weight of the i-th data, is the weight of the i-th data at the t-th moment, is the weight parameter, is the feature vector used for thermal pipeline life prediction. is the time series data of the i-th data, is the feature vector of the i-th data at the t-th time, M is the number of data types, and T is the time step; S7. The weighted fused feature representation is further calculated through a fully connected neural network, and the thermal pipeline life prediction result is generated.

2. A thermal pipeline life prediction method according to claim 1, characterized in that: In order to predict the life of thermal pipelines, relevant data of thermal pipelines are collected, including temperature, humidity, pressure, degree of corrosion of the inner surface of the pipeline, degree of corrosion of the outer surface of the pipeline, pipeline material, operating time, cycle load, and soil pH. When processing missing data, the data of the 5 days before and after are used for mean filling. If the data of the 5 days before and after are less than 5 days under the boundary conditions, the mean of the available days is used for filling. Then the processed data set is divided into training set and validation set as the basis for model learning and thermal pipeline life prediction.

Images