A method for obtaining the surface temperature field of an object
By introducing batch normalization and forgetting layers into the PINN model and combining them with L2 regularization, the weights of the loss function are adaptively adjusted, which solves the instability and overfitting problems of the classic PINN model when solving complex physical models, and achieves higher accuracy and more stable prediction of object surface temperature fields.
Patent Information
- Application Number
- CN202411123843.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-08-16
AI Technical Summary
The classic PINN model in the existing technology performs well when calculating simple equations, but the results are unstable when solving more complex physical models. It is also prone to overfitting on small datasets and requires manual adjustment of the weights of the model loss term and the data loss term, which affects the prediction accuracy.
A model for obtaining the surface temperature field of an object is constructed by using a weighted network and a physical computation network with a fully connected network structure, adding batch normalization layers and forgetting layers. Overfitting is suppressed by L2 regularization, and the weighted network is used to adaptively adjust the weights of the loss function to enhance the accuracy and stability of the prediction results.
It improves the prediction accuracy and stability of the model when solving complex physical models, enhances the model's generalization ability, optimizes the prediction effect, and solves the instability and overfitting problems of the classic PINN model when solving complex models.
Smart Images

Figure CN118657060B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of obtaining object surface temperature field information, and in particular to a method for obtaining an object surface temperature field. Background Art
[0002] The surface temperature field of an object is an important infrared characteristic physical quantity of the object's surface. With the known surface temperature field information, people can directly predict and track the behavior of the object, or use the temperature field as an intermediate variable to calculate physical quantities such as the radiation brightness and radiation intensity of the object's surface based on Planck's law, and calculate and predict the behavior of the object through physical quantities such as radiation brightness and radiation intensity.
[0003] The method of obtaining the surface temperature field of an object includes using the physical knowledge neural network (PINN) algorithm, integrating the physical equations and solution conditions into the loss function, and using the physical model to supervise the learning process of the network, which can achieve high-precision numerical calculations when there is limited or no data. Although the classic PINN algorithm has achieved good results in scientific computing problems of some physical models, the following problems still exist: 1. The classic PINN model based on the basic MLP architecture works well when calculating simple equations, but the results are unstable when solving more complex physical models; 2. The classic PINN model still has overfitting when facing small data sets; 3. The classic PINN model requires manual adjustment of the weight values of the model loss term and the data loss term, which affects the prediction accuracy of PINN.
[0004] In summary, there is an urgent need to provide a method for acquiring the surface temperature field of an object to solve the technical problems existing in the prior art. Summary of the invention
[0005] The purpose of the present invention is to provide a method for obtaining the surface temperature field of an object to solve the technical problems existing in the prior art. The specific technical solution is as follows:
[0006] A method for obtaining a surface temperature field of an object comprises the following steps:
[0007] Step S1, obtaining parameters that affect the heat distribution and temperature field value on the surface of the object;
[0008] Step S2, constructing an object surface heat transfer model and an object surface temperature field acquisition model; the object surface heat transfer model is a physical model based on the object surface heat transfer equation; the object surface temperature field acquisition model includes a weight network and a physical computing network, the weight network adopts a fully connected network structure, and is used to output the weight values of the model loss term and the data loss term; the physical computing network adds a batch normalization layer and a forgetting layer on the basis of the fully connected network structure, and the physical computing network embeds the object surface heat transfer model into the loss function to generate the value of the object surface temperature field;
[0009] Step S3, training the object surface temperature field acquisition model;
[0010] Step S4: input the parameters obtained in step S1 into the trained object surface temperature field acquisition model to obtain object surface temperature field data.
[0011] Furthermore, the parameters that affect the heat distribution and temperature field value on the surface of the object include the internal heat transfer power , External heat transfer power , Environmental heat transfer power And the heat dissipation power of the object surface to the outside world ;
[0012] The internal heat transfer power Affected by the difference between the internal temperature and the surface temperature of the object, the internal heat transfer power The expression is:
[0013] 1);
[0014] In formula 1), is the surface area of the object, is the internal temperature of the object, is the surface temperature of the object, is the thickness of the outermost material of the object, is the thermal conductivity of the outermost material of the object, is the thickness of the insulation material inside the object, is the thermal conductivity of the insulation material inside the object, is the internal flow heat transfer coefficient;
[0015] The external heat transfer power Affected by the relative speed of the object and the wind speed, the external ambient temperature and the surface temperature of the object, the external heat transfer power The expression is:
[0016] 2);
[0017] 3);
[0018] 4);
[0019] 5);
[0020] 6);
[0021] In equation 2) to equation 6), is the external ambient temperature; is the relative speed between wind speed and object speed, represents the tangential component of the relative velocity, represents the normal component of the relative velocity; is the wind speed, represents the tangential component of wind speed, represents the normal component of wind speed; is the object speed, represents the tangential component of the object's velocity, represents the normal component of the object's velocity;
[0022] The ambient heat transfer power Affected by the solar zenith angle and the external ambient temperature, the ambient heat transfer power The expression is:
[0023] 7);
[0024] 8);
[0025] 9);
[0026] 10);
[0027] In equation 7) to equation 10), Indicates the direct solar energy. represents the scattered energy of the sun, represents the atmospheric radiation energy, is the solar zenith angle, is the solar radiation constant, is the atmospheric optical thickness, is the blackbody radiation constant, is the angular coefficient between the object and the sky, is the emissivity of the atmosphere, It is the serial number of a certain day in a year. represents a natural constant;
[0028] The heat dissipation power of the object surface to the outside world Affected by the surface temperature of the object, its expression is:
[0029] 11);
[0030] In formula 11), is the emissivity of the object surface.
[0031] Furthermore, in step S2, the heat transfer equation in the object surface heat transfer model is:
[0032] 12);
[0033] In formula 12), The total heat of the surface of the object. is the density of the surface material of the object, is the surface temperature of the object, For time, is the specific heat capacity of the surface material of the object;
[0034] The initial conditions are:
[0035] 13);
[0036] In formula 13), is the initial temperature, The temperature at 0 o'clock every day;
[0037] The expression of the loss function is:
[0038] 14);
[0039] 15);
[0040] 16);
[0041] 17);
[0042] In equation 14) to equation 17), is the total loss, is the differential equation loss, is the initial loss, It is data loss. is the model loss term weight, is the weight of the data loss term, is the differential equation sampling number, is the number of initial condition samples, is the number of data samples, is the predicted value of the initial temperature, is the temperature at 0 o'clock every day, represents the predicted temperature value at a certain moment, Indicates the actual temperature value at a certain moment, is the sequence number of the vector element;
[0043] The model loss term weight and data loss term weight satisfy the following relationship:
[0044] 18);
[0045] In formula 18), , , is the model loss term weight, is the weight of the data loss term.
[0046] Furthermore, in the weight network, the expression for obtaining the output weight vector is:
[0047] 19);
[0048] in, The input weight vector includes the initialization model loss term weight and initialize the data loss weights Two elements, is the output weight vector of each layer of the weight network, and the output weight vector includes the model loss term weight and data loss weight Two elements, is the trainable weight of the hidden layer of the weight network, is the bias of the hidden layer of the weight network, is the activation function of the weight network.
[0049] Furthermore, the physical computing network is based on a fully connected network structure, and a batch normalization layer and a forgetting layer are added before the linear layer of the hidden layer of the physical computing network. The batch normalization layer, the forgetting layer and the linear layer are connected in sequence; the input layer and the hidden layer of the physical computing network use different activation functions, and the input data of the input layer is output after passing through the batch normalization layer, the forgetting layer and the linear layer, which is the network prediction data of the surface temperature field of the object.
[0050] Furthermore, the batch normalization layer is used to solve the problem of output data distribution changes in the high-level network after the low-level network is trained;
[0051] Before each layer of the physical computing network starts, the input data is normalized. The result of normalization is:
[0052] 20);
[0053] In formula 20), is the input data for each layer of the physical computing network, To normalize the input data, represents the mean of the input data, Represents the variance of the input data;
[0054] Add two learnable parameters to each neuron in the batch normalization layer and , the output data of the batch normalization layer for:
[0055] twenty one);
[0056] In formula 21), represents the variance learned by the batch normalization layer, Represents the mean learned by the batch normalization layer.
[0057] Furthermore, the forgetting layer is used to freeze the neurons in the linear layer with probability. The forgetting probability vector of the neurons in the linear layer is All elements in satisfy the Bernoulli distribution, that is,
[0058] twenty two);
[0059] In formula 22), represents the forgetting probability distribution of the linear layer neurons, Represents the forgetting probability of the linear layer neuron;
[0060] Input the output data of the batch normalization layer into the forgetting layer, and the output data of the forgetting layer is:
[0061] twenty three);
[0062] The output data of the forgetting layer is input into the linear layer, then the output data of the linear layer for:
[0063] twenty four);
[0064] In formula 24), is the output data vector of each forgetting layer, is the output data of the forgetting layer, is the trainable weight vector of neurons in each layer, are the trainable weights of the neurons, represents the bias vector of each linear layer, is the bias vector of the linear layer, is the activation function of the physical computing network.
[0065] Furthermore, an L2 regularization term is added to the loss function to correct the loss function. The value of the regularization term is set to , then we have the following expression:
[0066] 25);
[0067] 26);
[0068] In formula 25), represents the hyperparameters to be learned.
[0069] Furthermore, step S3 is specifically as follows:
[0070] Step S3.1, adjust the network state of the object surface temperature field acquisition model to the training state, in which the weight network is turned on, the batch normalization layer and the forgetting layer of the physical calculation network are turned on, and the training threshold is set;
[0071] Step S3.2, initialize the structure and parameters of the weight network and the physical computing network; the activation function of the input layer of the physical computing network is , the activation function of the first hidden layer of the physical computing network is , the activation function of the last hidden layer is , the activation functions of the remaining hidden layers are ;
[0072] Step S3.3, input the time, wind speed, solar zenith angle, date, external environment temperature and internal temperature of the object obtained in step S1, and generate network prediction data of temperature field by the physical computing network;
[0073] Step S3.4, using time, wind speed, solar zenith angle, date, external ambient temperature, internal temperature of the object and network prediction data generated by the physical computing network, respectively calculate the time differential term, internal heat transfer power, external heat transfer power, ambient heat transfer power and external heat dissipation power;
[0074] Step S3.5, calculating the differential equation loss according to the time differential term, internal heat transfer power, external heat transfer power, ambient heat transfer power and external heat dissipation power obtained in step S3.4, calculating the initial loss and data loss using network prediction data; generating the model loss term weight and data loss term weight using the weight network;
[0075] Step S3.6, calculate the total loss value based on the result of step S3.5, and compare the total loss value with the set training threshold; if the total loss value is less than or equal to the training threshold, output the training result; if the total loss value is greater than the training threshold, repeat steps S3.2 to S3.6 until the total loss value is less than or equal to the training threshold;
[0076] Step S3.7, save the trained object surface temperature field acquisition model.
[0077] Furthermore, step S4 is specifically as follows:
[0078] Step S4.1, adjusting the network state of the object surface temperature field acquisition model to a test state, in which the batch normalization layer and the forgetting layer of the physical computing network are turned off;
[0079] Step S4.2: Input the time, wind speed, solar zenith angle, date, external ambient temperature and internal temperature of the object obtained in step S1 into the trained object surface transmission model obtained in step S3.7 to obtain the object surface temperature field data.
[0080] The application of the technical solution of the present invention has the following beneficial effects:
[0081] The invention provides a method for acquiring a surface temperature field of an object, and proposes a surface temperature field acquisition model for an object, wherein the model comprises a weight network and a physical computing network, and the model adds a forgetting layer (dropout layer) and a batch normalization layer (BN layer) in a network architecture on the basis of a classical physical knowledge neural network (PINN) model, so as to enhance the accuracy and stability of prediction results; each layer adopts a different activation function, so as to enhance the nonlinear expression ability of the model, and enable the model to better solve and fit target variables, thereby increasing the generalization ability of the model; an L2 regularization technology is used in model training to further suppress overfitting; a weight network is used to adaptively adjust the weight of a loss function, so as to enable the model to find a weight value that balances the model loss and the data loss, which is conducive to further optimizing the prediction effect of the model.
[0082] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0083] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:
[0084] Figure 1 It is a flow chart of the method for acquiring the surface temperature field of an object in the present invention. DETAILED DESCRIPTION
[0085] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. However, the present invention can be implemented in many different ways as defined and covered.
[0086] In the description of the present invention, it should be noted that the terms "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", "front", "back", "lateral", "longitudinal" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.
[0087] Example:
[0088] See also Figure 1 This embodiment provides a method for obtaining the surface temperature field of an object, comprising the following steps:
[0089] Step S1, obtaining parameters that affect the heat distribution and temperature field value on the surface of the object;
[0090] Step S2, constructing an object surface heat transfer model and an object surface temperature field acquisition model; the object surface heat transfer model is a physical model based on the object surface heat transfer equation; the object surface temperature field acquisition model includes a weight network and a physical computing network, the weight network adopts a fully connected network structure, and is used to output the weight values of the model loss term and the data loss term; the physical computing network adds a batch normalization layer and a forgetting layer on the basis of the fully connected network structure, and the physical computing network embeds the object surface heat transfer model into the loss function to generate the value of the object surface temperature field;
[0091] Step S3, training the object surface temperature field acquisition model;
[0092] Step S4: input the parameters obtained in step S1 into the trained object surface temperature field acquisition model to obtain object surface temperature field data.
[0093] In step S1, the parameters that affect the heat distribution and temperature field value on the surface of the object include the internal heat transfer power , External heat transfer power , Environmental heat transfer power And the heat dissipation power of the object surface to the outside world ;
[0094] The internal heat transfer power Affected by the difference between the internal temperature and the surface temperature of the object, the internal heat transfer power The expression is:
[0095] 1);
[0096] In formula 1), is the surface area of the object, is the internal temperature of the object, is the surface temperature of the object, is the thickness of the outermost material of the object, is the thermal conductivity of the outermost material of the object, is the thickness of the insulation material inside the object, is the thermal conductivity of the insulation material inside the object, is the internal flow heat transfer coefficient;
[0097] The external heat transfer power Affected by the relative speed between the object speed and the wind speed, the external environment temperature and the surface temperature of the object, generally, the greater the tangential relative speed between the object speed and the wind speed, the greater the temperature difference between the external environment and the surface of the object, the greater the value of the external heat transfer power, and the external heat transfer power The expression is:
[0098] 2);
[0099] 3);
[0100] 4);
[0101] 5);
[0102] 6);
[0103] In equation 2) to equation 6), is the external ambient temperature; is the relative speed between wind speed and object speed, represents the tangential component of the relative velocity, represents the normal component of the relative velocity; is the wind speed, represents the tangential component of wind speed, represents the normal component of wind speed; is the object speed, represents the tangential component of the object's velocity, represents the normal component of the object's velocity;
[0104] The ambient heat transfer power Affected by the solar zenith angle and external ambient temperature, including direct solar energy , solar scattered energy , atmospheric radiation energy (On the water surface, the radiation of the water to the object must also be considered ); The size of the solar zenith angle mainly affects the amount of solar energy received by an object, and the external ambient temperature affects the heat exchange between an object and other media (such as the atmosphere);
[0105] Environmental heat transfer power The expression is:
[0106] 7);
[0107] 8);
[0108] 9);
[0109] 10);
[0110] In equation 7) to equation 10), Indicates the direct solar energy. represents the scattered energy of the sun, represents the atmospheric radiation energy, is the solar zenith angle, is the solar radiation constant, is the atmospheric optical thickness, is the blackbody radiation constant, is the angular coefficient between the object and the sky, is the emissivity of the atmosphere, It is the serial number of a certain day in a year. represents a natural constant;
[0111] The heat dissipation power of the object surface to the outside world Affected by the surface temperature of the object, its expression is:
[0112] 11);
[0113] In formula 11), is the emissivity of the object surface.
[0114] In step S2, the heat transfer equation in the object surface heat transfer model is:
[0115] 12);
[0116] In formula 12), The total heat of the surface of the object. is the density of the surface material of the object, is the surface temperature of the object, For time, is the specific heat capacity of the surface material of the object;
[0117] The initial conditions are:
[0118] 13);
[0119] In formula 13), is the initial temperature, The temperature at 0 o'clock every day;
[0120] The expression of the loss function is:
[0121] 14);
[0122] 15);
[0123] 16);
[0124] 17);
[0125] In equation 14) to equation 17), is the total loss, is the differential equation loss, is the initial loss, It is data loss. is the model loss term weight, is the weight of the data loss term, is the differential equation sampling number, is the number of initial condition samples, is the number of data samples, is the predicted value of the initial temperature, is the temperature at 0 o'clock every day, represents the predicted temperature value at a certain moment, Indicates the actual temperature value at a certain moment, is the sequence number of the vector element;
[0126] The model loss term weight and data loss term weight satisfy the following relationship:
[0127] 18);
[0128] In formula 18), , , is the model loss term weight, is the weight of the data loss term.
[0129] In this embodiment, the weight network adopts the MLP (i.e., fully connected network) architecture, and the weight network is mainly used to generate loss weights. In the weight network, the expression for obtaining the output weight vector is:
[0130] 19);
[0131] in, The input weight vector includes the initialization model loss term weight and initialize the data loss weights Two elements, is the output weight vector of each layer of the weight network, and the output weight vector includes the model loss term weight and data loss weight Two elements, is the trainable weight of the hidden layer of the weight network, is the bias of the hidden layer of the weight network, is the activation function of the weight network.
[0132] In this embodiment, the physical computing network is based on a fully connected network structure, and a batch normalization layer (BN layer) and a forgetting layer (dropout layer) are added before the linear layer of the hidden layer of the physical computing network. The batch normalization layer, the forgetting layer and the linear layer are connected in sequence; the input layer and the hidden layer of the physical computing network use different activation functions, and the input data of the input layer is output after the batch normalization layer, the forgetting layer and the linear layer, which is the network prediction data of the surface temperature field of the object. The activation function of the input layer of the physical computing network is , the activation function of the first hidden layer of the physical computing network is , the activation function of the last hidden layer is , the activation functions of the remaining hidden layers are .
[0133] In this embodiment, the batch normalization layer is used to solve the problem of output data distribution changes in the high-level network after the low-level network is trained;
[0134] Specifically, before each layer of the physical computing network starts, the input data is normalized. The result of normalization is:
[0135] 20);
[0136] In formula 20), is the input data for each layer of the physical computing network, To normalize the input data, represents the mean of the input data, Represents the variance of the input data;
[0137] In order not to destroy the features learned in the previous layer, two learning parameters are added to each neuron in the batch normalization layer. and , the output data of the batch normalization layer for:
[0138] twenty one);
[0139] In formula 21), represents the variance learned by the batch normalization layer, represents the mean learned by the batch normalization layer. At this time, we only need to , , the features learned in the previous layer can be restored.
[0140] In this embodiment, the forgetting layer is used to freeze the neurons in the linear layer with probability. The forgetting probability vector of the neurons in the linear layer is All elements in satisfy the Bernoulli distribution, that is,
[0141] twenty two);
[0142] In formula 22), represents the forgetting probability distribution of the linear layer neurons, Represents the forgetting probability of the linear layer neuron;
[0143] Input the output data of the batch normalization layer into the forgetting layer, and the output data of the forgetting layer is:
[0144] twenty three);
[0145] The linear layer is composed of several neurons represented as linear functions. The output data of the forgetting layer is input into the linear layer. The output data of the linear layer (that is, the output data of the entire physical computing network) for:
[0146] twenty four);
[0147] In formula 24), is the output data vector of each forgetting layer, is the output data of the forgetting layer, is the trainable weight vector of neurons in each layer, are the trainable weights of the neurons, represents the bias vector of each linear layer, is the bias vector of the linear layer, is the activation function of the physical computing network.
[0148] In order to further suppress the overfitting phenomenon, this embodiment adds an L2 regularization term to the loss function to correct the loss function. The value of the regularization term is set to , then we have the following expression:
[0149] 25);
[0150] 26);
[0151] In formula 25), Represents the hyperparameter to be learned, which is used to implement weight decay. In the deep learning framework, hyperparameter Will be integrated into the optimizer as a weight decay parameter.
[0152] The core of L2 regularization is to reduce the generalization error through weight decay, thereby suppressing overfitting. In actual training, setting the weight decay parameter in the Adam optimizer is equivalent to adding the L2 loss term to the loss function.
[0153] In order to facilitate training and using the model, the established heat transfer equation should be rewritten as:
[0154] 27);
[0155] The initial conditions are rewritten as:
[0156] 28);
[0157] In order to evaluate the quality of the model during training, this embodiment sets the average relative error between the predicted value and the actual value of the object surface temperature as the training threshold. Assume that the relative error vector between the predicted value and the actual value of the object surface temperature is , Indicates the relative error between a predicted value and the actual value. represents the average relative error value, is the number of elements contained in the relative error vector. The temperature prediction value at a certain moment is , the actual temperature at a certain moment is , then:
[0158] 29);
[0159] 30);
[0160] Usually, in an experimental environment, the average relative error should be controlled within 5%, that is, ; In general industrial production, the range of average relative error can be appropriately increased, but it should also be controlled within 10%, that is, .
[0161] In this embodiment, step S3 is specifically as follows:
[0162] Step S3.1, adjust the network state of the object surface temperature field acquisition model to the training state. In the training state, the weight network is turned on, the batch normalization layer and the forgetting layer of the physical calculation network are turned on, and the training threshold is set. (i.e. the average relative error between the predicted surface temperature and the actual temperature value of the object does not exceed 5%)
[0163] Step S3.2, initialize the structure and parameters of the weight network and the physical computing network; the activation function of the input layer of the physical computing network is , the activation function of the first hidden layer of the physical computing network is , the activation function of the last hidden layer is , the activation functions of the remaining hidden layers are ;
[0164] Step S3.3, input the time, wind speed, solar zenith angle, date, external environment temperature and internal temperature of the object obtained in step S1, and generate network prediction data of temperature field by the physical computing network;
[0165] Step S3.4, using time, wind speed, solar zenith angle, date, external ambient temperature, internal temperature of the object and network prediction data generated by the physical computing network, respectively calculate the time differential term, internal heat transfer power, external heat transfer power, ambient heat transfer power and external heat dissipation power;
[0166] Step S3.5, calculating the differential equation loss according to the time differential term, internal heat transfer power, external heat transfer power, ambient heat transfer power and external heat dissipation power obtained in step S3.4, calculating the initial loss and data loss using network prediction data; generating the model loss term weight and data loss term weight using the weight network;
[0167] Step S3.6, calculate the total loss value based on the result of step S3.5, and compare the total loss value with the set training threshold; if the total loss value is less than or equal to the training threshold, output the training result; if the total loss value is greater than the training threshold, repeat steps S3.2 to S3.6 until the total loss value is less than or equal to the training threshold;
[0168] Step S3.7, save the trained object surface temperature field acquisition model.
[0169] In this embodiment, step S4 is specifically as follows:
[0170] Step S4.1, adjusting the network state of the object surface temperature field acquisition model to a test state, in which the batch normalization layer and the forgetting layer of the physical computing network are turned off;
[0171] Step S4.2: Input the time, wind speed, solar zenith angle, date, external ambient temperature and internal temperature of the object obtained in step S1 into the trained object surface transmission model obtained in step S3.7 to obtain the object surface temperature field data.
[0172] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for obtaining the surface temperature field of an object, characterized in that: The following steps are involved: Step S1, obtaining parameters that affect the heat distribution and temperature field value on the surface of the object; Step S2, constructing an object surface heat transfer model and an object surface temperature field acquisition model; the object surface heat transfer model is a physical model based on the object surface heat transfer equation; the object surface temperature field acquisition model includes a weight network and a physical calculation network, and the weight network adopts a fully connected network structure to output the weight values of the model loss term and the data loss term; The physical computing network adds a batch normalization layer and a forgetting layer on the basis of a fully connected network structure, and the physical computing network embeds a heat transfer model of the object surface into a loss function to generate a numerical value of the temperature field on the object surface; Step S3, training the object surface temperature field acquisition model; Step S4, inputting the parameters obtained in step S1 into the trained object surface temperature field acquisition model to obtain the object surface temperature field data; The parameters that affect the heat distribution and temperature field value on the surface of an object include the internal heat transfer power Q in , External heat transfer power Q out , Environmental heat transfer power Q en And the heat dissipation power Q of the object surface to the outside world sf ; The internal heat transfer power Q in Affected by the difference between the internal temperature and the surface temperature of the object, the internal heat transfer power Q in The expression is: In formula 1), A is the surface area of the object, T in is the internal temperature of the object, T is the surface temperature of the object, d1 is the thickness of the outermost material of the object, k1 is the thermal conductivity of the outermost material of the object, d2 is the thickness of the insulation layer inside the object, k2 is the thermal conductivity of the insulation layer inside the object, and h in is the internal flow heat transfer coefficient; The external heat transfer power Q out Affected by the relative speed of the object and the wind speed, the external ambient temperature and the surface temperature of the object, the external heat transfer power Q out The expression is: Q out =3.65A·|v rp | 0.6 (T out -T) Formula 2); In formula 2) to formula 6), T out is the external ambient temperature; is the relative speed between wind speed and object speed, v rp represents the tangential component of the relative velocity, v rn represents the normal component of the relative velocity; is the wind speed, v wp represents the tangential component of wind speed, v wn represents the normal component of wind speed; is the object velocity, v op represents the tangential component of the object's velocity, v on represents the normal component of the object's velocity; The ambient heat transfer power Q en Affected by the solar zenith angle and the external ambient temperature, the ambient heat transfer power Q en The expression is: Q en = (E dir +q scatter +q sky )A formula 7); In equations 7) to 10), E dir represents the direct solar energy, q scatter represents the solar scattered energy, q sky represents the atmospheric radiation energy, θ is the solar zenith angle, E0 is the solar radiation constant, τ0 is the atmospheric optical thickness, σ is the blackbody radiation constant, is the angular coefficient between the object and the sky, ε sky is the emissivity of the atmosphere, N is the number of a day in a year, and e represents a natural constant; The object surface dissipates heat to the outside world Q sf Affected by the surface temperature of the object, its expression is: Q sf = AεσT 4 Equation (11); In formula (11), ε is the emissivity of the object surface; In step S2, the heat transfer equation in the object surface heat transfer model is: In formula 12), represents the total heat on the surface of the object, ρ is the density of the surface material of the object, T is the surface temperature of the object, t is the time, and c is the specific heat capacity of the surface material of the object; The initial conditions are: T0=T 0:00 Formula 13); In formula (13), T0 is the initial temperature, T 0:00 The temperature at 0 o'clock every day; The expression of the loss function is: L = λ model (L pde + L0) + λ p L p Equation 14); In equations 14) to 17), L is the total loss, L pde is the differential equation loss, L0 is the initial loss, L p is the data loss, λ model is the model loss term weight, λ p is the weight of the data loss term, N pde is the differential equation sampling number, N0 is the initial condition sampling number, N p is the number of data samples, T net (0) is the predicted value of the initial temperature, T 0:00 is the temperature at 0 o'clock every day, T net (t i ) represents the predicted temperature value at a certain moment, T true (t i ) represents the actual temperature value at a certain moment, and i is the sequence number of the vector element; The model loss term weight and data loss term weight satisfy the following relationship: λ model +λ p = 1 Equation (18); In formula (18), λ model >0,λ p >0,λ model is the model loss term weight, λ p is the weight of the data loss term.
2. The method for obtaining the surface temperature field of an object according to claim 1, characterized in that: In the weight network, the expression for obtaining the output weight vector is: λ = f weight (ω weight ·λ0 + b weight ) Equation (19); Among them, λ0 is the input weight vector, which includes the initialization model loss term weight λ model_0 and initialize the data loss weight λ p_0 Two elements, λ is the output weight vector of each layer of the weight network, and the output weight vector includes the model loss term weight λ model and data loss weight λ p Two elements, ω weight is the trainable weight of the hidden layer of the weight network, b weight is the bias of the hidden layer of the weight network, f weight (·) is the activation function of the weight network.
3. The method for acquiring the surface temperature field of an object according to claim 2, characterized in that: The physical computing network is based on a fully connected network structure. A batch normalization layer and a forgetting layer are added before the linear layer of the hidden layer of the physical computing network. The batch normalization layer, the forgetting layer and the linear layer are connected in sequence. The input layer and the hidden layer of the physical computing network use different activation functions. The input data of the input layer are output after the batch normalization layer, the forgetting layer and the linear layer, which are the network prediction data of the surface temperature field of the object.
4. The method for acquiring the surface temperature field of an object according to claim 3, characterized in that: The batch normalization layer is used to solve the problem of output data distribution changes in the high-level network after the low-level network is trained; Before each layer of the physical computing network starts, the input data is normalized. The result of normalization is: In formula 20), x i is the input data for each layer of the physical computing network, To normalize the input data, represents the mean of the input data, and S represents the variance of the input data; Add two learning parameters S to each neuron in the batch normalization layer bn and Output data of batch normalization layer for: In formula 21), S bn represents the variance learned by the batch normalization layer, Represents the mean learned by the batch normalization layer.
5. The method for acquiring the surface temperature field of an object according to claim 4, characterized in that: The function of the forgetting layer is to freeze the neurons in the linear layer with probability. The forgetting probability vector of the neurons in the linear layer is R = {r1, r2, ..., r i All elements in} satisfy the Bernoulli distribution, that is, r i ~ Bernoulli (p) formula 22); In formula 22), r i represents the forgetting probability distribution of the linear layer neurons, and p represents the forgetting probability of the linear layer neurons; Input the output data of the batch normalization layer into the forgetting layer, and the output data of the forgetting layer is: The output data of the forgetting layer is input into the linear layer, and the output data Y of the linear layer is: In formula 24), is the output data vector of each forgetting layer, is the output data of the forgetting layer, ω physics =[ω1,ω2,...,ω i ] is the trainable weight vector of neurons in each layer, ω i is the trainable weight of the neuron, b physics =[b1,b2,...,b i ] represents the bias vector of each linear layer, b i is the bias vector of the linear layer, f physics (·) is the activation function of the physical computing network.
6. The method for acquiring the surface temperature field of an object according to claim 5, characterized in that: Add the L2 regularization term to the loss function to correct the loss function. Assuming the value of the regularization term is L2, the following expression is obtained: L = λ model (L pde + L0) + λ p L p + L2 Equation 26); In Equation 25), α represents the hyperparameter to be learned.
7. The method for acquiring the surface temperature field of an object according to claim 6, characterized in that: Step S3 is specifically as follows: Step S3.1, adjust the network state of the object surface temperature field acquisition model to the training state, in which the weight network is turned on, the batch normalization layer and the forgetting layer of the physical calculation network are turned on, and the training threshold is set; Step S3.2, initialize the structure and parameters of the weight network and the physical computing network; the activation function of the input layer of the physical computing network is sin(·), the activation function of the first hidden layer of the physical computing network is tanh(·), the activation function of the last hidden layer is rrelu(·), and the activation function of the remaining hidden layers is relu(·); Step S3.3, input the time, wind speed, solar zenith angle, date, external environment temperature and internal temperature of the object obtained in step S1, and generate network prediction data of temperature field by the physical computing network; Step S3.4, using time, wind speed, solar zenith angle, date, external ambient temperature, internal temperature of the object and network prediction data generated by the physical computing network, respectively calculate the time differential term, internal heat transfer power, external heat transfer power, ambient heat transfer power and external heat dissipation power; Step S3.5, calculating the differential equation loss according to the time differential term, internal heat transfer power, external heat transfer power, ambient heat transfer power and external heat dissipation power obtained in step S3.4, calculating the initial loss and data loss using network prediction data; generating the model loss term weight and data loss term weight using the weight network; Step S3.6, calculating the total loss value based on the result of step S3.5, and comparing the total loss value with the set training threshold; if the total loss value is less than or equal to the training threshold, outputting the training result; If the total loss value is greater than the training threshold, repeat steps S3.2 to S3.6 until the total loss value is less than or equal to the training threshold; Step S3.7, save the trained object surface temperature field acquisition model.
8. The method for acquiring the surface temperature field of an object according to claim 7, characterized in that: Step S4 is specifically as follows: Step S4.1, adjusting the network state of the object surface temperature field acquisition model to a test state, in which the batch normalization layer and the forgetting layer of the physical computing network are turned off; Step S4.2: Input the time, wind speed, solar zenith angle, date, external ambient temperature and internal temperature of the object obtained in step S1 into the trained object surface transmission model obtained in step S3.7 to obtain the object surface temperature field data.
