Creep life damage prediction method based on ResNet-LSTM neural network

Abstract

The invention discloses a main steam pipeline creep life damage prediction method based on a ResNet-LSTM neural network. The method comprises the steps that the ResNet-LSTM neural network is constructed based on LSTM basic units and ResNet basic units; the method comprises the following steps: dividing input features and output features from a data set for predicting the service life of a main steam pipeline, analyzing and screening out main input features through a Pearson correlation coefficient, training a ResNet-LSTM neural network by using the screened main input features and output features, and performing stress prediction; establishing a nearest neighbor interpolation stress-life damage mapping relation to obtain a life damage fitting curved surface, and substituting the stress prediction value into the fitting curved surface to obtain a life damage prediction value; and evaluating the life damage predicted value through an average percentage error. The operation parameter-life damage prediction model is established, and accurate prediction of the life of the main steam pipeline can be achieved.

Description

Technical Field

[0001] The present invention belongs to the field of big data technology and relates to a creep life damage prediction method based on a ResNet-LSTM neural network. Background Art

[0002] Currently, most thermal power units operate in a flexible mode. Under flexible load variations, the stress and lifespan of thick-walled boiler components, especially the main steam pipe, vary significantly compared to design conditions. Therefore, it is important to study the factors affecting the lifespan of main steam pipes under flexible operation and to predict lifespan damage based on unit operating conditions.

[0003] At present, the stress calculation of specific structural components mostly adopts the finite element method, but the finite element calculation speed is slow and it is impossible to obtain results in real time according to the changes in the operating status of the unit. The load of the flexible operating unit and the main steam temperature, pressure and other parameters change rapidly over time. If the finite element method is used to obtain stress, it cannot meet the requirements of real-time prediction. In terms of creep life prediction, existing technologies mainly rely on the results of material experiments, which are time-consuming and cannot meet the requirements of real-time prediction. Flexible peak-shaving units hope to quickly estimate the life of the pipeline based on load changes and main steam parameters, but the life prediction formulas obtained from material experiments mostly rely on stress and strain, such as the Larson-Miller relationship and the Wilshire relationship, which cannot meet this requirement. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a creep life damage prediction method based on the ResNet-LSTM neural network, and to establish an operating parameter-life damage prediction model by combining the ResNet-LSTM network with the nearest neighbor interpolation method, that is, to predict the creep life damage in the variable load process by using the unit load and main steam parameters.

[0005] Technical solution: The present invention provides a creep life damage prediction method based on a ResNet-LSTM neural network, comprising the following steps:

[0006] The LSTM basic unit is used as the first layer, the ResNet basic unit is used as the second layer, and the first and second layers are connected with the traditional neural network to construct a ResNet-LSTM neural network for stress prediction.

[0007] The input features and output features are divided from the data set used for main steam pipe life prediction. The main input features are screened out through Pearson correlation coefficient analysis. The selected main input features and output features are used to train the ResNet-LSTM neural network for stress prediction.

[0008] The nearest neighbor interpolation stress-life damage mapping relationship is established based on the three characteristics of pipeline stress, creep life damage and adjustment time, and the life damage fitting surface is obtained. The stress prediction value and adjustment time obtained by the ResNet-LSTM neural network are substituted into the established fitting surface to obtain the life damage prediction value; and the life damage prediction value is evaluated by the mean percentage error (MAPE).

[0009] Furthermore, the ResNet basic unit is composed of a convolutional layer Conv, a normalization layer BN, and a nonlinear activation function layer ReLU. Let the input of the i-th ResNet basic unit be x i , then the output of the ResNet basic unit is:

[0010] x i+1 =f(x i +F(x i ,W i ))

[0011] Among them, x i+1 is the output of the i+1th ResNet basic unit, F is the residual function; W i is the weight corresponding to the i-th ResNet basic unit; f is the activation function ReLU. The form of the residual function F is not unique, and its stacked content and number of layers are selected and adjusted as needed.

[0012] Furthermore, the basic unit of the long short-term memory neural network LSTM contains three inputs, which are the input vector X at the current moment t , the unit state C at the previous moment t-1 And the hidden layer state h at the previous moment t-1 ; The unit contains two outputs, which are the current unit state C t and the current hidden layer state h t ; The abandonment and inheritance functions of its internal control information are controlled by the forget gate f t , input gate i t and output gate o t To achieve this; the output of the three gates is expressed as:

[0013] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0014] i t =σ(W i ·[h t-1 ,X t ]+b i )

[0015] o t =σ(W o ·[h t-1 ,X t ]+b o )

[0016] Among them, f t 、i t 、o t are the outputs of the forget gate, input gate, and output gate respectively, W f 、W i 、W o are the weight matrices of the forget gate, input gate, and output gate, respectively, b f 、b i 、b o are the bias variables of the forget gate, input gate, and output gate respectively, σ is the sigmoid activation function; the current unit state C t The status is:

[0017]

[0018] in, intermediate quantities calculated for the unit;

[0019] The current hidden layer state h t for:

[0020] h t =o t *tanh(C t ).

[0021] Furthermore, the features in the dataset used for main steam pipeline life prediction include unit load, regulation time, main steam temperature, main steam pressure, main steam flow, pipeline temperature, pipeline stress, and creep life damage; among them, creep life damage is calculated from pipeline temperature and pipeline stress:

[0022]

[0023] Among them, σ m is the pipeline stress; R m is the ultimate tensile stress of the material; k . and v are parameters obtained by fitting material experiments; t0 is the creep life; is the creep activation energy; R is the molar gas constant; T is the pipe temperature; D3 is the creep life damage; t ； is the duration of the load; t0 is the creep life;

[0024] The input features divided for stress prediction include unit load, adjustment time, main steam temperature, main steam pressure, and main steam flow; the output feature is pipeline stress.

[0025] Furthermore, the main input features are screened out through Pearson correlation coefficient analysis, which specifically involves calculating the correlation coefficients between all features and filtering out features whose correlation with the output features is less than a threshold. The calculation formula is:

[0026]

[0027] Among them, cov(X,Y) is the covariance of the two feature sequences X and Y, σ > and σ Y is the standard deviation of the two feature sequences X and Y, x i and y i are the i-th data in the feature sequences X and Y respectively, i = 1, 2, ..., n, and n is the number of data in each feature sequence;

[0028] After screening, the main input characteristics include unit load, regulation time, main steam pressure and main steam flow.

[0029] Furthermore, the main input features and output features after screening are divided into training sets and test sets. The training set is used to train the ResNet-LSTM neural network, and the test set is used to test the prediction ability of the neural network to verify the stress prediction results.

[0030] Furthermore, the nearest neighbor interpolation stress-life damage mapping relationship is established based on the three characteristics of pipeline stress, creep life damage, and adjustment time, and the life damage fitting surface is obtained, specifically:

[0031] Assume that the average pipe stress of the i-th data in the data set is a 1i , the adjustment time is a .i , creep life damage is Z i =g(a 1i ,a .i ); If the average pipeline stress and adjustment time of a data point to be interpolated are c1 and c . , then the nearest neighbor interpolation result Z is:

[0032] Z=g(D miG )

[0033]

[0034] Among them, D i is the distance between the data point to be interpolated and the data point i, D miG For all D i The minimum value of g is the interpolation function, which is determined by the three-dimensional coordinate point (a 1i ,a .i ,Z i), where i = 1, 2, …, n, and n is the number of data in each feature sequence;

[0035] The stress prediction value and adjustment time are substituted into the established life damage fitting surface to obtain the life damage prediction value of the main steam pipeline.

[0036] The system corresponding to the method includes:

[0037] The ResNet-LSTM neural network building unit is used to use the LSTM basic unit as the first layer and the ResNet basic unit as the second layer, and connect the first and second layers with the traditional neural network to build a ResNet-LSTM neural network for stress prediction;

[0038] The neural network training and prediction unit is used to divide the input features and output features from the data set used for main steam pipe life prediction, and screen out the main input features through Pearson correlation coefficient analysis. The selected main input features and output features are used to train the ResNet-LSTM neural network for stress prediction;

[0039] The life damage prediction and evaluation unit is used to establish a nearest neighbor interpolation stress-life damage mapping relationship based on the three characteristics of pipeline stress, creep life damage, and adjustment time, obtain a life damage fitting surface, substitute the stress prediction value and adjustment time obtained by the ResNet-LSTM neural network into the established fitting surface to obtain the life damage prediction value; and evaluate the life damage prediction value using the mean percentage error (MAPE).

[0040] An electronic device for storing and executing the method, the device comprising:

[0041] a memory storing executable program code;

[0042] a processor coupled to the memory;

[0043] The processor calls the executable program code stored in the memory to execute the steps of the creep life damage prediction method based on the ResNet-LSTM neural network.

[0044] A computer-readable storage medium for storing and executing the method, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the steps of the creep life damage prediction method based on the ResNet-LSTM neural network.

[0045] Beneficial Effects: Compared with existing technologies, this invention demonstrates significant advantages: It adds LSTM and ResNet layers before traditional neural network layers to construct a ResNet-LSTM neural network. It also utilizes a ResNet-LSTM network combined with nearest neighbor interpolation to establish an operating parameter-life damage prediction model. This model predicts creep life damage during load variations using unit load and main steam parameters, enabling accurate prediction of main steam pipe life. Compared with conventional neural network predictions, this method reduces prediction error from 3.94% to 0.57%. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 1 is a structural diagram of the basic unit of the neural network of the present invention, wherein (a) is the ResNet basic unit structure, and (b) is the LSTM basic unit structure;

[0048] Figure 3 This is the ResNet-LSTM neural network structure diagram of the present invention;

[0049] Figure 4 It is the Pearson correlation coefficient analysis heat map of the present invention;

[0050] Figure 5 is the stress prediction result of the present invention, where (a) is the training set result and (b) is the test set result;

[0051] Figure 6 is the neighbor interpolation surface of the present invention;

[0052] Figure 7 It is a comparison chart of the prediction results of the present invention and the common neural network, wherein (a) is the life damage prediction value of the present invention, and (b) is the life damage prediction value of the common neural network. DETAILED DESCRIPTION

[0053] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] In order to meet the needs of fast adjustment and real-time prediction under flexible operation, a neural network method can be used to predict creep life. As the number of stacked layers of a deep neural network increases, the performance of the model will increase accordingly, but at the same time, the problem of gradient vanishing or gradient exploding will occur, which will make it difficult for the deep neural network to converge during training. The Residual Neural Network (ResNet) can accelerate the training of deep neural networks and significantly improve their accuracy. In addition, ResNet can largely avoid the problem of gradient vanishing or gradient exploding as the number of network layers increases, which makes it possible to train extremely deep networks.

[0055] Long short-term memory neural networks (LSTM) can overcome the problem of vanishing gradients in traditional recurrent neural networks, where all hidden layer states before a certain moment do not affect the update of weights during training. LSTM has been widely used in various fields in recent years.

[0056] The Pearson correlation coefficient reflects the degree of linear correlation between two series, with values ranging from -1 to 1. The closer the Pearson coefficient is to 1, the closer the two series are to a positive linear correlation; the closer it is to -1, the closer the series are to a negative linear correlation; and the closer it is to 0, the weaker the linear correlation. This method is often used to screen for highly correlated feature parameters in a dataset.

[0057] The nearest neighbor interpolation method is a commonly used interpolation method. Its core lies in that the value of the interpolation point can be replaced by the value of the point closest to it, thereby fitting the unknown coordinate points, so that the stress prediction results can be mapped to the life damage prediction results.

[0058] Therefore, in order to solve the problem that the life of the main steam pipeline of the flexible thermal power unit deviates from the set value, it is necessary to quickly predict the pipeline life online. The present invention constructs a ResNet-LSTM neural network structure by adding an LSTM layer and a ResNet layer before the neural network layer; through Pearson correlation coefficient analysis, the main input features are screened out from all input features, and the ResNet-LSTM neural network is trained using the main input features and output features to perform stress prediction; the nearest neighbor interpolation stress-life damage mapping relationship (NNI) is established based on the three features of pipeline stress, creep life damage and adjustment time to obtain a life damage fitting surface, and the stress prediction value and adjustment time are substituted into the established fitting surface to obtain the pipeline life damage prediction value.

[0059] The present invention takes the actual operation data of the main steam pipeline of a power plant as an example to further explain in detail:

[0060] like Figure 1As shown, the creep life damage prediction method based on ResNet-LSTM neural network described in the present invention includes the following steps:

[0061] Step 1: Build a ResNet-LSTM neural network;

[0062] To address issues such as vanishing and exploding gradients in traditional deep neural networks, this paper adds LSTM and ResNet layers before the traditional neural network layer to construct a ResNet-LSTM neural network for stress prediction. The structure of the ResNet-LSTM neural network is shown in Table 1. The ResNet-LSTM neural network consists of three LSTM layers, eight ResNet layers, and 11 fully connected neural network layers. Each neural network layer is composed of a certain number of basic structural connections.

[0063] Table 1 ResNet-LSTM neural network structure

[0064]

[0065] Dense is a fully connected layer, and the number of layers in parentheses after it is the number of layers. ReLU is a nonlinear activation function, and Linear is a linear activation function. This network architecture can overcome the problems of vanishing and exploding gradients, improve the network's learning ability, and enhance prediction results.

[0066] Construct a ResNet basic unit. The ResNet basic unit consists of a convolutional layer Conv, a normalization layer BN, and a nonlinear activation function layer ReLU. Let the input of the i-th ResNet basic unit be x i , then the output of the i-th ResNet basic unit is:

[0067] x i+1 =f(x i +F(x i ,W i ))

[0068] Among them, x i+1 is the output of the i+1th ResNet basic unit, F is the residual function; W i is the weight corresponding to the i-th ResNet basic unit; f is the activation function ReLU, the form of the residual function F is not unique, and its stacked content and number of layers are selected and adjusted as needed. Figure 2 The residual function F shown in (a) is shown in Figure 2.

[0069] Construct the basic unit of long short-term memory neural network (LSTM), such as Figure 2As shown in (b), the LSTM basic unit contains three inputs, which are the input vector X at the current moment t , the unit state C at the previous moment t-1 And the hidden layer state h at the previous moment t-1 ; The unit contains two outputs, which are the current unit state C t and the current hidden layer state h t ; The abandonment and inheritance functions of its internal control information are controlled by the forget gate f t , input gate i t and output gate o t To achieve this; the output of the three gates is expressed as:

[0070] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0071] i t =σ(W i ·[h t-1 ,X t ]+b i )

[0072] o t =σ(W o ·[h t-1 ,X t ]+b o )

[0073] Among them, f t 、i t 、o t are the outputs of the forget gate, input gate, and output gate respectively, W f 、W i 、W o are the weight matrices of the forget gate, input gate, and output gate, respectively, b f 、b i 、b o are the bias variables of the forget gate, input gate, and output gate respectively, σ is the sigmoid activation function; the current unit state C t The status is:

[0074]

[0075] in, Intermediate quantities calculated for the element.

[0076] The current hidden layer state h t for:

[0077] h t =ot *tanh(C t )

[0078] A certain number of LSTM basic units are connected as the first layer, a certain number of ResNet basic units are connected as the second layer, and these two layers are connected with the traditional neural network to construct the ResNet-LSTM neural network. The neural network structure is as follows Figure 3 shown.

[0079] The constructed ResNet-LSTM neural network will be used in the subsequent stress prediction in step 2.

[0080] Step 2: The input and output features of the main steam pipeline life prediction dataset were divided and the key input features were selected through Pearson correlation coefficient analysis. The selected key input features were used to train the ResNet-LSTM neural network for stress prediction. The input and output features included in this dataset include unit load, adjustment time, main steam temperature, main steam pressure, main steam flow, pipeline temperature, pipeline stress, and creep life damage. Creep life damage is calculated from pipeline temperature and pipeline stress:

[0081]

[0082] Among them, σ m is the pipeline stress, MPa; R m is the ultimate tensile stress of the material, MPa; k . and v are parameters obtained by fitting material experiments; t0 is the creep life, h; is the creep activation energy, kJ / mol; R is the molar gas constant, 8.314J / (mol*K); T is the pipe temperature, K. D3 is the creep life damage, t ； is the duration of the load, h, and t0 is the creep life, h.

[0083] The main input features for stress prediction are divided into unit load, adjustment time, main steam temperature, main steam pressure, and main steam flow; the output feature is pipeline stress.

[0084] The Pearson correlation coefficient can reflect the degree of linear correlation between two sequences, and its value is between -1 and 1. The closer the Pearson coefficient is to 1, the closer the two sequences are to linear positive correlation; the closer it is to -1, the closer the sequences are to linear negative correlation; the closer it is to 0, the weaker the linear correlation between the two sequences. If X and Y are two feature sequences containing n data, then the Pearson correlation coefficient ρ between the two features is >,Y It can be calculated by the following formula:

[0085]

[0086] Among them, cov(X,Y) is the covariance of the two feature sequences X and Y, σ > and σ Y is the standard deviation of the two feature sequences X and Y, x i 、y i are the i-th data in the feature sequences x and Y respectively, i = 1, 2,…, n, and n is the number of data in each feature sequence.

[0087] Calculate the Pearson correlation coefficient between all features in the dataset and plot the results as a correlation heat map, such as Figure 4 As shown in the figure, the horizontal and vertical axes represent each feature, and the number at the intersection of the axes is the Pearson correlation coefficient between the features. The darker the color, the stronger the correlation.

[0088] Depend on Figure 4 It can be seen that the characteristics that are strongly correlated with creep damage are: initial load (0.34), adjustment time (0.81), average steam pressure (0.34), average steam flow (0.33), and average pipe stress (0.34).

[0089] After screening, the main input features include unit load, adjustment time, main steam pressure and main steam flow, and pipeline stress is used as the output feature. The first 85% of the data is taken as the training set for training the neural network, and the last 15% of the data is used as the test set to verify the stress prediction effect of the model.

[0090] The ResNet-LSTM neural network constructed in step 1 is trained using the training set, and the prediction ability of the neural network is tested using the test set to verify the stress prediction results. The stress prediction values on the training set and the test set are as follows: Figure 5 As shown in (a) and (b).

[0091] Step 3: A nearest neighbor interpolation stress-life damage mapping (NNI) is established based on the pipeline stress, creep life damage, and adjustment time characteristics. This results in a life damage fitting surface. The stress prediction values and adjustment time from the ResNet-LSTM neural network are substituted into this fitting surface to obtain the life damage prediction value. The prediction results are evaluated using the mean average percentage error (MAPE).

[0092] The nearest neighbor interpolation method is a commonly used interpolation method. Assume that the average pipeline stress of the i-th data in the data set is a 1i , the adjustment time is a .i , creep life damage is Z i =g(a 1i ,a .i). If the average pipeline stress and adjustment time of a data point to be interpolated are c1 and c . , then the nearest neighbor interpolation result Z is:

[0093] Z=g(D miG )

[0094]

[0095] Among them, D i is the distance between the data point to be interpolated and the data point i, D miG For all D i The minimum value of g is the interpolation function, which is determined by the three-dimensional coordinate point (a 1i ,a .i ,Z i ), where i = 1, 2, …, n, and n is the number of data in each feature sequence.

[0096] The life damage fitting surface obtained by fitting the pipeline stress, creep life damage and adjustment time characteristics is as follows: Figure 6 shown.

[0097] The obtained main steam pipeline life damage prediction value is evaluated by MAPE and compared with the life damage prediction value of the ordinary neural network under the same data set. The prediction results of the two are as follows: Figure 7 (a) and (b) are shown in Figure 2, and the MAPE is shown in Table 2. It can be seen from the results that the MAPE of the method proposed in this invention is significantly smaller than that of the common neural network method, which proves the effectiveness of the method proposed in this invention.

[0098] Table 2 MAPE of this method and ordinary neural network prediction

[0099]

[0100] In summary, the ResNet-LSTM neural network-based main steam pipeline creep life damage prediction method described in this paper can establish an operating parameter-life damage prediction model based on the stress prediction model structure of the ResNet-LSTM network and the nearest neighbor interpolation (NNI) method, achieving accurate prediction of the main steam pipeline life. Compared with conventional neural network prediction, the life damage prediction value of this method has a smaller MAPE and better prediction effect.

[0101] The present invention also provides a creep life damage prediction system based on a ResNet-LSTM neural network, comprising:

[0102] The ResNet-LSTM neural network building unit is used to use the LSTM basic unit as the first layer and the ResNet basic unit as the second layer. The first and second layers are connected with the traditional neural network to build a ResNet-LSTM neural network for stress prediction and life prediction;

[0103] The neural network training and prediction unit is used to divide the input features and output features from the data set used for main steam pipe life prediction, and screen out the main input features through Pearson correlation coefficient analysis. The selected main input features and output features are used to train the ResNet-LSTM neural network for stress prediction;

[0104] The life damage prediction and evaluation unit is used to establish a nearest neighbor interpolation stress-life damage mapping relationship based on the three characteristics of pipeline stress, creep life damage, and adjustment time, obtain a life damage fitting surface, substitute the stress prediction value obtained by the ResNet-LSTM neural network into the established fitting surface to obtain the life damage prediction value; and evaluate the life damage prediction value using the mean percentage error (MAPE).

[0105] The present invention further provides an electronic device, comprising:

[0106] a memory storing executable program code;

[0107] a processor coupled to the memory;

[0108] The processor calls the executable program code stored in the memory to execute the steps of the creep life damage prediction method based on the ResNet-LSTM neural network.

[0109] The present invention also provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute the steps of the creep life damage prediction method based on the ResNet-LSTM neural network.

Claims

1. A creep life damage prediction method based on ResNet-LSTM neural network, characterized in that: The following steps are involved: The LSTM basic unit is used as the first layer, the ResNet basic unit is used as the second layer, and the first and second layers are connected with the traditional neural network to construct a ResNet-LSTM neural network for stress prediction. The input features and output features are divided from the data set used for main steam pipe life prediction. The main input features are screened out through Pearson correlation coefficient analysis. The selected main input features and output features are used to train the ResNet-LSTM neural network for stress prediction. A nearest neighbor interpolation stress-life damage mapping relationship is established based on the three characteristics of pipeline stress, creep life damage, and adjustment time, and a life damage fitting surface is obtained. The stress prediction value and adjustment time obtained by the ResNet-LSTM neural network are substituted into the established fitting surface to obtain the life damage prediction value. The lifetime damage prediction value is evaluated by the mean percentage error (MAPE).

2. The creep life damage prediction method based on ResNet-LSTM neural network according to claim 1, characterized in that: The ResNet basic unit consists of a convolutional layer Conv, a normalization layer BN, and a nonlinear activation function layer ReLU. Suppose the input of the i-th ResNet basic unit is x i , then the output of the ResNet basic unit is: x i+1 =f(x i +F(x i ,W i )) Among them, x i+1 is the output of the i+1th ResNet basic unit, F is the residual function; W i is the weight corresponding to the i-th ResNet basic unit; f is the activation function ReLU. The form of the residual function F is not unique, and its stacked content and number of layers are selected and adjusted as needed.

3. The creep life damage prediction method based on ResNet-LSTM neural network according to claim 1, characterized in that: The basic unit of the long short-term memory neural network LSTM contains three inputs, namely the input vector X at the current moment t , the unit state C at the previous moment t-1 And the hidden layer state h at the previous moment t-1 ; The unit contains two outputs, which are the current unit state C t and the current hidden layer state h t ; The abandonment and inheritance functions of its internal control information are controlled by the forget gate f t , input gate i t and output gate o t To achieve this; the output of the three gates is expressed as: f t =σ(W f ·[h t-1 ,X t ]+b f ) i t =σ(W i ·[h t-1 ,X t ]+b i ) the t =σ(W o ·[h t-1 ,X t ]+b o ) Among them, f t 、i t 、o t are the outputs of the forget gate, input gate, and output gate respectively, W f 、W i 、W o are the weight matrices of the forget gate, input gate, and output gate, respectively, b f 、b i 、b o are the bias variables of the forget gate, input gate, and output gate respectively, σ is the sigmoid activation function; the current unit state C t The status is: in, intermediate quantities calculated for the unit; The current hidden layer state h t for: h t =o t *tanh(C t )。 4. The creep life damage prediction method based on ResNet-LSTM neural network according to claim 1, characterized in that: The features in the dataset used for main steam pipeline life prediction include unit load, regulation time, main steam temperature, main steam pressure, main steam flow, pipeline temperature, pipeline stress, and creep life damage; the creep life damage is calculated from pipeline temperature and pipeline stress: Among them, σ m is the pipeline stress; R m is the ultimate tensile stress of the material; k . and v are parameters obtained by fitting material experiments; t0 is the creep life; is the creep activation energy; R is the molar gas constant; T is the pipe temperature; D3 is the creep life damage; t ； is the duration of the load; t0 is the creep life; The input features divided for stress prediction include unit load, adjustment time, main steam temperature and main steam pressure, and main steam flow; the output feature is pipeline stress.

5. The creep life damage prediction method based on ResNet-LSTM neural network according to claim 1, characterized in that: The main input features are screened out through Pearson correlation coefficient analysis, which specifically involves calculating the correlation coefficients between all features and filtering out features whose correlation with the output features is less than a threshold; The calculation formula is: Among them, cov(X,Y) is the covariance of the two feature sequences X and Y, σ > and σ Y is the standard deviation of the two feature sequences X and Y, x i and y i are the i-th data in the feature sequences X and Y respectively, i = 1, 2, ..., n, and n is the number of data in each feature sequence; After screening, the main input characteristics include unit load, regulation time, main steam pressure and main steam flow.

6. The creep life damage prediction method based on ResNet-LSTM neural network according to claim 1, characterized in that: The filtered main input features and output features are divided into training sets and test sets. The training set is used to train the ResNet-LSTM neural network, and the test set is used to test the prediction ability of the neural network and verify the stress prediction results.

7. The creep life damage prediction method based on ResNet-LSTM neural network according to claim 1, characterized in that: The nearest neighbor interpolation stress-life damage mapping relationship is established based on the three characteristics of pipeline stress, creep life damage, and adjustment time, and the life damage fitting surface is obtained, specifically: Assume that the average pipe stress of the i-th data in the data set is a 1i , the adjustment time is a .i , creep life damage is Z i =g(a 1i ,a .i ); If the average pipeline stress and adjustment time of a data point to be interpolated are c1 and c . , then the nearest neighbor interpolation result Z is: Z=g(D miG ) Among them, D i is the distance between the data point to be interpolated and the data point i, D miG For all D i The minimum value of g is the interpolation function, which is determined by the three-dimensional coordinate point (a 1i ,a .i ,Z i ), where i = 1, 2, …, n, and n is the number of data in each feature sequence; The stress prediction value and adjustment time are substituted into the established life damage fitting surface to obtain the life damage prediction value of the main steam pipeline.

8. A creep life damage prediction system based on ResNet-LSTM neural network, characterized in that: include: The ResNet-LSTM neural network building unit is used to use the LSTM basic unit as the first layer and the ResNet basic unit as the second layer, and connect the first and second layers with the traditional neural network to build a ResNet-LSTM neural network for stress prediction; The neural network training and prediction unit is used to divide the input features and output features from the data set used for main steam pipe life prediction, and screen out the main input features through Pearson correlation coefficient analysis. The selected main input features and output features are used to train the ResNet-LSTM neural network for stress prediction; The life damage prediction and assessment unit is used to establish a nearest neighbor interpolation stress-life damage mapping relationship based on the three characteristics of pipeline stress, creep life damage, and adjustment time, obtain a life damage fitting surface, and substitute the stress prediction value and adjustment time obtained by the ResNet-LSTM neural network into the established fitting surface to obtain the life damage prediction value; The lifetime damage prediction value is evaluated by the mean percentage error (MAPE).

9. An electronic device, characterized in that: The device comprises: a memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the creep life damage prediction method based on the ResNet-LSTM neural network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, which, when called, are used to execute the steps of the creep life damage prediction method based on the ResNet-LSTM neural network as described in any one of claims 1 to 7.

Images