Extension temperature field reconstruction method based on neural network mapping

Abstract

The invention relates to the technical field of material forming, and discloses a pull-out temperature field reconstruction method based on neural network mapping. According to the method, a mapping relation between drawing process parameters and three-dimensional temperature field distribution data is established, and the three-dimensional temperature field distribution data is obtained through shell area equal-interval grid division and unit center point temperature value extraction; determining a reference measuring point layout scheme based on the temperature gradient distribution and extracting a temperature value of the reference measuring point layout scheme; calculating a temperature difference value between a non-measuring point area and the reference measuring point, and constructing a measuring point coordinate-temperature difference value field sample pair; processing the sample pair by using a multi-layer perceptron model, inputting the sample pair as a joint vector of a measuring point coordinate and a temperature value, and outputting the sample pair as a predicted temperature difference field; superposing the prediction difference value field and the reference measurement point temperature value to obtain a complete three-dimensional temperature field; a two-dimensional heat distribution diagram is generated through geometric mapping, the vertical coordinate is the axial length, and the horizontal coordinate is the circumferential expansion angle. According to the invention, gradient features and key temperature information in a temperature field can be accurately captured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metal drawing process, in particular to a drawing temperature field reconstruction method based on neural network mapping. BACKGROUND

[0002] In the field of metal processing, the drawing process is widely used in the production of pipe and profile products by applying axial tension and radial pressure to the blank to produce plastic deformation and obtain products with specific shape and size. The temperature field, as a key physical field in the drawing process, directly affects the plastic flow of the material, the evolution of the microstructure and the mechanical properties of the final product. Therefore, accurately obtaining the three-dimensional temperature field distribution during the drawing process is of great significance for optimizing process parameters and ensuring product quality.

[0003] Existing temperature field acquisition methods have obvious limitations. Contact measurement such as thermocouple temperature measurement requires the sensor to be implanted inside or in contact with the surface of the blank. Although local point temperature data can be obtained, the number of measurement points is limited, making it difficult to cover the complex three-dimensional space. The presence of the sensor may interfere with the temperature field distribution, and it is also prone to damage in high-temperature and high-speed drawing environments, affecting the continuity of the measurement. Non-contact measurement such as infrared thermal imaging technology can achieve rapid acquisition of surface temperature, but the measurement accuracy is limited due to factors such as material surface emissivity, oxide layer and environmental interference. Moreover, it is difficult to obtain internal temperature information of the blank, making it difficult to reflect the overall three-dimensional temperature field.

[0004] Numerical simulation methods such as finite element analysis calculate the temperature field by establishing heat conduction equations and boundary conditions. Although three-dimensional temperature distribution data can be provided, complex meshing, material parameter setting and iterative calculation are required, which is time-consuming. Especially in the scene of multi-condition parameter debugging or real-time process monitoring, it is difficult to meet the efficiency requirements. Traditional temperature field reconstruction methods often rely on empirical interpolation or simple fitting, often ignoring the nonlinear relationship between process parameters such as initial blank temperature, loading speed and friction factor and the temperature field. When the process conditions change, the reconstruction results are prone to large deviations.

[0005] The existing technology lacks targeted measurement point arrangement and does not fully consider the temperature gradient distribution characteristics, resulting in insufficient information capture in key temperature regions. After temperature field reconstruction, there is a lack of effective visualization conversion methods, and three-dimensional data is difficult to directly apply to process analysis and adjustment in the production site, further limiting the guiding value of temperature field data in actual production. Therefore, there is an urgent need for a drawing temperature field reconstruction method that can consider the influence of process parameters, combine accurate measurement point information and be efficient and comprehensive. SUMMARY

[0006] The present application aims to provide a drawing temperature field reconstruction method based on neural network mapping to solve the problems raised in the background.

[0007] To achieve the above object, the application provides a drawing-out temperature field reconstruction method based on neural network mapping, which comprises the following steps:

[0008] A mapping relationship between drawing-out process parameters and three-dimensional temperature field distribution data is established, wherein the drawing-out process parameters include initial blank temperature, loading speed, friction factor and boundary heat conduction condition, and the three-dimensional temperature field distribution data is obtained by equally spacing grid division of a blank area and extraction of unit center point temperature values;

[0009] A reference measuring point layout scheme is determined based on temperature gradient distribution, and reference measuring point temperature values are extracted from the three-dimensional temperature field distribution data;

[0010] Temperature difference values of non-measuring point regions and the reference measuring point temperature values are calculated, and a measuring point coordinate-temperature difference field sample pair is constructed;

[0011] The measuring point coordinate-temperature difference field sample pair is processed through a multilayer perception machine model, wherein the input of the multilayer perception machine model is a measuring point coordinate and temperature value joint vector, and the output is a predicted temperature difference field;

[0012] The predicted temperature difference field is superimposed to the reference measuring point temperature values, and a complete three-dimensional temperature field is reconstructed;

[0013] Geometric mapping processing is performed on the complete three-dimensional temperature field, and a two-dimensional heat distribution map is generated, wherein the longitudinal coordinate of the two-dimensional heat distribution map is an axial length, and the transverse coordinate is a circumferential development angle.

[0014] Preferably, the establishment of the mapping relationship between the drawing-out process parameters and the three-dimensional temperature field distribution data comprises:

[0015] A plurality of process vector groups are configured, and each process vector group contains independently configured initial blank temperature parameters, loading speed parameters, friction factor parameters and boundary heat conduction condition parameters;

[0016] Thermal-mechanical coupling finite element simulation is performed on each process vector group, and blank temperature field data at a plurality of simulation time points is output;

[0017] The blank surface and internal area are equally spaced grid divided in the axial and circumferential directions, and the spatial coordinates and corresponding temperature values of the grid unit center points are extracted;

[0018] The spatial coordinates and corresponding temperature values are stored in sequence according to the simulation time points, and a process parameter-temperature field mapping database is constructed.

[0019] Preferably, the determination of the reference measuring point layout scheme based on the temperature gradient distribution comprises:

[0020] extracting temperature field data of each simulation time from the process parameter-temperature field mapping database;

[0021] calculating a weighted sum of the axial temperature change rate and the circumferential temperature change rate in the temperature field data to generate a temperature gradient distribution map;

[0022] identifying a continuous region in the temperature gradient distribution map where the gradient value is greater than a preset threshold, and marking it as a high gradient region;

[0023] According to the spatial distribution density of the high gradient region and the equipment accessibility constraint, the installation position coordinate set of the infrared thermometer is configured.

[0024] Preferably, the constructed sample pair of measurement point coordinates-temperature difference field includes:

[0025] reading the grid integration point coordinate set and the corresponding temperature value of the target simulation time from the process parameter-temperature field mapping database;

[0026] extracting the reference measurement point temperature value corresponding to the infrared thermometer installation position coordinate set;

[0027] calculating the real-time temperature difference between each grid integration point temperature value and the nearest reference measurement point temperature value;

[0028] combining the reference measurement point coordinates, the reference measurement point temperature value and the grid integration point coordinates into an input vector, and setting the corresponding real-time temperature difference as an output label;

[0029] store the input vector and the output label in order according to the grid integration point number to form a temperature difference training sample set.

[0030] Preferably, the method further comprises processing the sample pair of measurement point coordinates-temperature difference field by a multilayer perception machine model, which includes:

[0031] constructing a fully connected network structure including an input layer, at least three hidden layers and an output layer;

[0032] inputting the input vector into the input layer and performing nonlinear transformation through the hidden layer nodes;

[0033] generating a predicted temperature difference value at the output layer;

[0034] The hidden layer adopts a rectified linear unit activation function, and the output layer adopts a linear activation function.

[0035] Preferably, the method further comprises performing supervised training on the multilayer perception machine model:

[0036] splitting the temperature difference training sample set into a training subset and a validation subset according to a preset proportion;

[0037] Inputting the training subset input vector as the model input and the training subset output label as the supervision signal;

[0038] Calculating the mean square error of the predicted temperature difference value and the output label;

[0039] Updating the network weight parameters through the stochastic gradient descent algorithm;

[0040] When the mean square error of the validation subset does not decrease for three consecutive iterations, stop training and save the network parameters.

[0041] Preferably, the method further comprises:

[0042] Collecting the measured temperature value corresponding to the coordinate set of the installation position of the infrared temperature measuring instrument at the current time;

[0043] Combining the measured temperature value with the corresponding coordinates to form a current time measurement point data set;

[0044] Inputting the current time measurement point data set into the trained multilayer perceptron model to obtain the predicted temperature difference field of the non-measurement point region output by the model.

[0045] Preferably, the reconstruction of the complete three-dimensional temperature field comprises:

[0046] Mapping the predicted temperature difference field to the three-dimensional space according to the grid integral point coordinates;

[0047] Performing temperature superposition on each grid integral point in the three-dimensional space: taking the measured temperature value of the nearest reference measurement point as the base value, and superimposing the predicted temperature difference value corresponding to the integral point;

[0048] Performing spatial interpolation smoothing processing on the superposition result to generate a continuously distributed three-dimensional temperature field.

[0049] Preferably, the generation of the two-dimensional heat distribution map comprises:

[0050] Establishing an axial-circumferential two-dimensional coordinate system, the axial coordinate being the projection length of the billet center axis, and the circumferential coordinate being the unwound circular arc angle;

[0051] Mapping the center point temperature value of each grid element in the three-dimensional temperature field to the two-dimensional coordinate system;

[0052] Filling the temperature blank area of the two-dimensional coordinate system using a bilinear interpolation algorithm, and drawing temperature contour lines and point cloud distribution maps according to the interpolation results.

[0053] Preferably, the method further comprises a dynamic optimization step:

[0054] Monitoring the change amplitude of the predicted temperature difference field in the last five reconstruction cycles;

[0055] When the change amplitude exceeds the preset temperature difference threshold, a new measurement point arrangement instruction is triggered;

[0056] According to the spatial coordinates of the region with the maximum change amplitude, the reference measurement point arrangement scheme is updated again, and the measured data after the new measurement points are added is used to retrain the multi-layer perception machine model.

[0057] Beneficial effects: The temperature field reconstruction method based on neural network mapping has significant advantages in temperature field representation and application through multi-link cooperation. In the measurement point arrangement link, the reference measurement point is determined based on the temperature gradient distribution, which can make the measurement point position more suitable for the key change region of the temperature field. Compared with the traditional uniform arrangement or experience arrangement method, the gradient characteristics and key temperature information in the temperature field can be more accurately captured, providing more representative basic data for subsequent temperature field reconstruction. By extracting the temperature value of the reference measurement point, the local accurate measurement data is used as the anchor point for reconstruction, reducing the overall deviation caused by the lack of global measurement.

[0058] The way of constructing the sample pair of measurement point coordinates-temperature difference field focuses on the relative temperature change of the non-measurement point region and the reference measurement point, rather than directly predicting the absolute temperature value. This processing method can reduce the influence of absolute temperature errors caused by environmental interference, measurement equipment precision limitations and other factors on the overall reconstruction result, making the model more focused on capturing the spatial distribution law of the temperature field and enhancing the stability of the temperature field reconstruction. At the same time, the measurement point coordinate information is integrated into the sample pair, which enables the model to learn the correlation characteristics of temperature difference and spatial position, laying a foundation for subsequent difference field prediction.

[0059] The application of the multi-layer perception machine model fully utilizes its ability to handle nonlinear mapping relationships. During the drawing and stretching process, there is a complex coupling relationship between process parameters such as initial billet temperature, loading speed, friction factor and temperature field distribution, which is difficult to accurately describe by traditional linear models or simple fitting methods. The multi-layer perception machine can effectively fit the complex mapping between process parameters, spatial coordinates and temperature field through nonlinear transformation of multiple neurons, making the predicted temperature difference field more consistent with the actual working condition, and being more adaptable to temperature field changes under different process parameter combinations than traditional methods.

[0060] Superimposing the predicted temperature difference field on the reference measurement point temperature value to reconstruct the complete three-dimensional temperature field realizes the organic combination of local accurate measurement and global distribution prediction. The reference measurement point provides a reliable local temperature reference, and the predicted difference field supplements the temperature information of the non-measurement point region. The superposition of the two not only retains the measurement accuracy of the key position, but also expands the spatial coverage of the temperature field, making the reconstructed three-dimensional temperature field not only reflect the local details, but also present the overall distribution characteristics, improving the comprehensiveness of temperature field representation.

[0061] The geometric mapping generates a two-dimensional heat distribution map, converts complex three-dimensional temperature field data into a two-dimensional form with axial length and circumferential expansion angle as coordinates, and makes the temperature distribution information more intuitive and easy to understand. In actual production, the operator can quickly observe the temperature change trend at different axial positions and circumferential angles through the two-dimensional heat map, facilitate the identification and analysis of the temperature abnormal area in the drawing process, enhance the practical value of the temperature field data in the production site, and provide a more easily interpreted visual basis for process adjustment. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 The working principle diagram of the drawing temperature field reconstruction method based on neural network mapping is described in the present application.

[0063] Figure 2 The flowchart for establishing the mapping relationship between the drawing process parameters and the three-dimensional temperature field distribution data is described.

[0064] Figure 3 The flowchart for constructing the sample pair of the measurement point coordinate temperature difference field is described.

[0065] Figure 4 The flowchart for supervised training of the multi-layer perception model is described. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0067] Please refer to Figure 1 The present application provides a drawing temperature field reconstruction method based on neural network mapping, which comprises:

[0068] By establishing the mapping relationship between the drawing process parameters and the three-dimensional temperature field distribution data, the temperature field data is generated by finite element simulation, and the reference measurement point layout scheme is determined based on the temperature gradient distribution. By extracting the temperature value of the reference measurement point, the temperature difference of the non-measurement point area is calculated, the sample pair of the measurement point coordinate-temperature difference field is constructed, the temperature difference field is predicted by using the multi-layer perception model, and finally the complete three-dimensional temperature field is reconstructed by superimposing the temperature value of the reference measurement point, and a two-dimensional heat distribution map is generated. This method can efficiently and accurately reconstruct the temperature field distribution in the drawing process, and provide data support for process optimization.

[0069] Example 1: Please refer to Figure 2When establishing the mapping relationship between the drawing process parameters and the three-dimensional temperature field distribution data, multiple sets of process vector groups are configured. Each set of process vector groups includes independently set initial blank temperature parameters, loading speed parameters, friction factor parameters, and boundary heat conduction condition parameters. The initial blank temperature parameters are selected within a specific range, the loading speed parameters are selected within another specific range, the friction factor parameters are configured within a specific interval, and the boundary heat conduction condition parameters are selected according to the actual working conditions as natural convection or forced cooling mode. For each set of process vector groups, perform a thermal-mechanical coupling finite element simulation operation, and complete the simulation process using a commercial finite element software platform. During the simulation process, set the material physical property parameters, define the boundary constraint conditions, and configure the loading path parameters. The simulation time step is set to a specific numerical range according to the process requirements. The simulation process outputs the shell temperature field data at multiple time nodes, including surface temperature distribution and internal temperature distribution information. The shell geometry region is equally spaced grid divided in the axial and circumferential directions, and the grid element size is set to a specific value according to the accuracy requirements. The three-dimensional space coordinate data and the corresponding temperature values of the center point of each grid element are extracted. All space coordinates and temperature values are stored in sequence according to the simulation time series to form the mapping database of process parameters and temperature field data.

[0070] When determining the reference measurement point layout scheme based on the temperature gradient distribution, the temperature field data at each time node is extracted from the process parameter-temperature field mapping database. The temperature change rate of adjacent grid elements in the axial direction of the temperature field data is calculated, and the temperature change rate of adjacent grid elements in the circumferential direction is also calculated. The change rates in the two directions are weighted and summed to generate a global temperature gradient distribution map. The continuous spatial region with a gradient value exceeding a set threshold value in the temperature gradient distribution map is identified, and the set threshold value is determined according to the specific value range of the process characteristics. The identified high gradient region is spatially marked. According to the spatial distribution density characteristics of the high gradient region, combined with the spatial accessibility restriction conditions of the on-site equipment installation, the spatial layout position coordinate set of the infrared temperature measuring instrument is determined. The coordinate set needs to cover the main high gradient region while meeting the physical installation constraint conditions of the temperature measuring equipment. The three-dimensional coordinate data of each layout point is accurately recorded and stored to form a reference measurement point spatial position database.

[0071] In the temperature gradient calculation process, the numerical difference method is used to process the grid cell temperature data. The temperature change rate in the axial direction is obtained by calculating the ratio of the temperature difference of adjacent grid cells to the axial distance of the cell. The temperature change rate in the circumferential direction is obtained by calculating the ratio of the temperature difference of adjacent grid cells to the circumferential arc length of the cell. The weighting coefficients of the change rates in the two directions are assigned specific values according to the process characteristics. The temperature gradient distribution map is presented in the form of a spatial cloud chart, and different color areas represent different gradient intensity levels. The identification of high gradient areas uses the region growing algorithm in image processing technology to automatically mark the connected areas with gradient values exceeding the threshold value. The selection of the installation position of the infrared thermometer follows the principles of spatial uniformity and gradient coverage, maximizing the coverage of high gradient areas within the range that the equipment can reach. Each selected installation position outputs three-dimensional coordinate data and records the grid cell number information corresponding to the position.

[0072] In the process parameter configuration stage, the initial billet temperature parameters are selected as discrete value points within a specific high temperature range. The loading speed parameters are selected as multiple discrete values within a specific speed range. The friction factor parameters are selected as different values within a specific coefficient range. The boundary heat conduction condition parameters are set according to the cooling medium type to set the convective heat transfer coefficient value range. Each parameter in each process vector group is independently combined to form a complete parameter configuration scheme. When the finite element simulation is executed, the material model selects the elastic-plastic constitutive relationship to define the temperature-dependent material property parameters. The loading process uses displacement control mode and sets specific loading rate values. The thermal boundary conditions include environmental temperature parameters and heat transfer coefficient parameters. The simulation output includes temperature field data files at multiple time steps, each file recording the temperature values of all grid nodes at that time. The structured grid generation technology is used for grid division, with grid nodes uniformly distributed in the axial and circumferential directions. The cell center point coordinates are obtained by calculating the average of the cell node coordinates. The temperature values are directly obtained from the output results of the finite element solver. The mapping database uses a time series database structure to store data, with each time step corresponding to an independent data table, recording grid coordinates and temperature value fields.

[0073] In the temperature gradient analysis stage, the spatial gradient of the temperature field data at each time step is calculated. The calculation process traverses all internal grid cells, and the axial and circumferential temperature change rates of each cell are calculated. The weighted summation coefficient is set to a fixed proportion according to process experience. The generated temperature gradient distribution map is stored in matrix form, and the matrix row and column correspond to the grid position, and the element value is the gradient intensity. The high gradient area identification uses a threshold segmentation algorithm, and after setting the gradient threshold, the connected region is automatically extracted. The region marking result is stored in the form of a binary image, and the high gradient area is marked as 1, and the remaining areas are marked as 0. When the measuring point layout scheme is generated, candidate points are selected in the high gradient marked area according to the spatial density requirement. The candidate points need to meet the spatial constraint conditions of equipment installation, such as avoiding moving parts and meeting the minimum installation distance. The final determined measuring point coordinate set contains position number and three-dimensional coordinate data, and records the grid cell information of each measuring point. The coordinate set is used as the basis for the spatial position of the subsequent temperature field reconstruction reference points.

[0074] Example 2: see Figure 3 , when constructing the measuring point coordinate-temperature difference field sample pair, the data record of the target simulation time is selected from the process parameter-temperature field mapping database. The grid integration point coordinate set corresponding to this time is extracted completely, and each coordinate point contains three-dimensional spatial position information. At the same time, the temperature values corresponding to each coordinate point are read, which come from the temperature field results output by the finite element simulation. The infrared thermometer installation position coordinate set is used as the basis for the spatial position of the reference points, and the three-dimensional coordinate data of each measuring point is obtained from the set one by one. According to the coordinate matching principle, the grid cell closest to the reference point coordinate is located in the grid integration point set, and its temperature value is recorded as the reference point temperature value. This matching process uses a spatial distance minimization algorithm to realize, and the Euclidean distance between the center point of each grid cell and the reference point is calculated, and the temperature value of the cell with the smallest distance is selected as the representative temperature of the corresponding reference point.

[0075] For each grid integration point, the real-time temperature difference between its temperature value and the temperature value of the nearest reference measurement point is calculated. The determination of the nearest reference measurement point also adopts the principle of minimizing spatial distance, and the spatial distance between the current grid integration point and all reference measurement point coordinates is calculated, and the reference measurement point with the smallest distance is selected as the reference. The temperature difference value is obtained by subtracting the temperature value of the reference measurement point from the temperature value of the current grid integration point, and a positive value indicates that it is higher than the reference temperature, and a negative value indicates that it is lower than the reference temperature. The temperature difference value reflects the degree of temperature deviation of the local region relative to the reference measurement point. The reference measurement point coordinate data, reference measurement point temperature value data, and current grid integration point coordinate data are combined into a multi-dimensional input vector. The dimension of the input vector is determined by the number of reference measurement points and the coordinate dimension, and the coordinates and temperature values of each reference measurement point are included in the vector. The output label is set to the real-time temperature difference value corresponding to the current grid integration point. According to the numbering order of the grid integration points in the database, the input vector and the output label are paired and stored to form a temperature difference training sample set. The sample set size depends on the number of grid points and the number of time steps of the simulation model, and usually contains thousands to tens of thousands of sample data.

[0076] When processing the measurement point coordinate-temperature difference field sample pair through the multi-layer perception model, the designed neural network structure is as follows. The network adopts a fully connected architecture, including an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer strictly corresponds to the dimension of the input vector, ensuring that all input data can be completely received. The number of hidden layers is set according to the complexity of the problem, usually not less than three. The number of nodes in each hidden layer is configured within a certain range, which is determined according to the sample characteristics and computing resources. The nodes in the hidden layer are connected through weighted connections to realize information transmission, and the weight parameters are assigned with a specific distribution in the initialization stage. The activation function of the hidden layer is selected as the rectified linear unit type, which maintains linear characteristics in the positive interval and outputs zero in the negative interval, which can effectively avoid the problem of gradient disappearance. The number of nodes in the output layer is fixed to one, corresponding to the predicted temperature difference value. The output layer adopts a linear activation function to ensure that the output value range is not limited.

[0077] In the network parameter initialization process, the hidden layer weight matrix adopts a specific initialization method that considers the variance adjustment of the activation function characteristics. The bias vector is initially set to zero. During forward propagation, the input vector enters the network through the input layer and undergoes nonlinear transformation in the hidden layer. The output value of each hidden layer node is obtained by weighting and processing the input through the activation function, and the calculation result is used as the input of the next layer. The final output layer generates the predicted temperature difference value, which is a continuous scalar. Network training uses a supervised learning mode, with the difference between the predicted value and the true label as the optimization target. The loss function is defined as the mean squared error of the predicted temperature difference and the actual temperature difference, which can quantify the prediction accuracy of the model. During backpropagation, the gradient of the loss function with respect to the network parameters is calculated using automatic differentiation techniques. The weight update uses an improved version of the stochastic gradient descent algorithm, which introduces a momentum term and an adaptive learning rate mechanism. The learning rate parameter is set within a certain range and can be dynamically adjusted during training. The batch size of the training samples is set according to the memory capacity, usually using tens to hundreds of samples for each update. The number of training rounds is determined based on the convergence condition, and the training process is terminated when the validation set error tends to be stable.

[0078] In the sample construction phase, the grid integration point coordinate set covers the entire shell region, including surface and internal points. When extracting the temperature values of the reference measurement points, if the reference measurement point coordinates do not coincide with the grid nodes, the temperature values are obtained using the interpolation method of adjacent grid elements. In the temperature difference calculation process, the spatial distance calculation considers the actual geometric distance in the three-dimensional coordinate system. When constructing the input vector, the reference measurement point information is arranged in a fixed order to ensure data consistency. Sample storage uses a structured data format, with each record containing a complete input feature vector and output label. In the design of neural network structure, the number of hidden layer nodes can be set hierarchically, and different hidden layers can have different numbers of nodes. The selection of the activation function is based on the function characteristics and problem requirements, and the rectified linear unit can alleviate the training difficulties of deep networks. Linear activation of the output layer ensures that the temperature difference prediction is not limited by the value range. The parameter initialization method affects the training convergence speed, and a specific initialization strategy can accelerate the training process. The loss function calculation uses the statistical average of batch samples to balance the calculation efficiency and stability. The optimization algorithm configuration includes multiple hyperparameters, such as the initial learning rate and the momentum coefficient, which are set through empirical values. The training termination condition is determined based on the change in the validation set error, and the training is stopped when the error does not improve continuously to prevent overfitting.

[0079] The training sample generation process traverses all simulation time points and all grid cells to form a large-scale sample set. Each sample corresponds to a specific grid cell at a specific time, recording the temperature difference relationship between the cell and the reference measuring point. The reference measuring point information in the input vector contains complete data of all measuring points, enabling the model to perceive the overall measuring point distribution. The output label is only for the current grid cell, reflecting the local temperature difference characteristics. Before training the neural network, standardization processing is performed on the input data to scale the characteristics of each dimension to a similar numerical range. The connection between the hidden layers is in full connection mode, and the output of all nodes of the previous layer is used as the input of the nodes of the next layer. Forward propagation calculation is realized by matrix operation, and hardware acceleration is used to improve efficiency. During back propagation, the gradient calculation is based on the chain rule, and the error signal is transmitted from the output layer to the input layer layer by layer. The weight update formula contains the learning rate and the gradient direction, and the loss function value is gradually reduced through iterative optimization. During the training process, the loss curves of the training set and the validation set are monitored to observe the learning state of the model. The final saved model parameters include the weight matrix and bias vector of each layer, which are used for subsequent temperature difference prediction tasks.

[0080] In Example 3, referring to Figure 4 When performing supervised training on the multi-layer perceptron model, the temperature difference training sample set is divided into a training subset and a validation subset according to a preset ratio. The ratio is dynamically adjusted according to the total number of samples, and the training subset is used for model parameter optimization, and the validation subset is used for monitoring the training process. The input vector of the training subset is used as the model input data, and the input vector dimension is completely matched with the network input layer node number. The output label of the training subset is used as the supervision signal, which provides the corresponding true temperature difference value for each input sample. The model predicted temperature difference is obtained by forward propagation calculation, and compared with the output label to generate an error signal. The error calculation uses the mean square error function, which quantifies the deviation between the predicted value and the true value. The mean square error calculation process is as follows:

[0081]

[0082] Wherein: represents the loss function value, N represents the number of samples in the current batch, y i is the true temperature difference value of the i-th sample, is the temperature difference value predicted by the model, and θ represents all trainable parameters of the network. This loss function is sensitive to outliers and can effectively drive the model to focus on larger error samples.

[0083] The network weight parameter update uses the stochastic gradient descent algorithm, which gradually optimizes the model parameters through iteration. Each parameter update is based on the loss function gradient of the current batch, and the gradient calculation is realized by the back propagation algorithm. During back propagation, the error signal is transmitted from the output layer to the input layer layer by layer, and the parameter gradient of each layer is calculated by the chain rule. The weight update formula is as follows:

[0084]

[0085] where: θ t+1 denotes the updated parameter, θ t is the current parameter, η is the learning rate parameter, is the gradient of the loss function with respect to the parameter. The learning rate parameter controls the step size of each update, which is set to a large value at the beginning of training and gradually decays as the training progresses. The momentum term accelerates convergence and suppresses oscillations by accumulating the direction of historical gradients to enhance the stability of parameter updates.

[0086] An early stopping strategy is employed to monitor the validation set error during training. The validation set mean squared error is computed at the end of each training epoch, and its relative change from the previous epoch is recorded. When the validation error does not show a downward trend for three consecutive iterations, the model is considered to have reached optimal performance, and the training process is terminated. The saved network parameters at this time include the weight matrices and bias vectors of each hidden layer, as well as the connection weights of the output layer. The parameter saving format uses a structured data file, which contains the numerical matrices and configuration information of each layer's parameters.

[0087] The division of the training set and the validation set uses a random sampling strategy to ensure the consistency of the data distribution of the two subsets. The input vector of the training set enters the model through the input layer of the network, and after nonlinear transformation by the hidden layer, it reaches the output layer. The activation function of the hidden layer is processed using element-level operations, and the output of each node is calculated independently. The linear transformation of the output layer maps the features of the hidden layer to the temperature difference prediction, which is a continuous numerical value. When calculating the loss function, the predicted values and the true values of the batch samples are compared pair by pair, and the average of the squared errors is obtained as the current loss value.

[0088] During the backpropagation process, the output layer gradient is directly calculated from the prediction error. The hidden layer gradient is calculated by backpropagating the gradient from the upper layer, taking into account the derivative characteristics of the activation function. When updating the parameters, the weight matrix and the bias vector are adjusted according to the corresponding gradient. The learning rate parameter uses an adaptive adjustment strategy, with an initial value that is dynamically decayed according to the training situation. The momentum coefficient controls the influence of historical gradients, and this parameter remains fixed during the training process. The early stopping judgment is based on the moving average of the validation set error, which eliminates the misjudgment caused by random fluctuations.

[0089] In the implementation of the training process, the input data is first standardized for preprocessing. The mean and standard deviation of each feature dimension are calculated, and the original data is converted to a zero-mean unit-variance distribution. The standardization parameters are obtained from the training subset and are applied simultaneously to the validation subset. Network parameter initialization uses random generation with a specific distribution, considering the relationship between the input and output dimensions of each layer. Training iterations are organized in cycles, with each cycle traversing the training subset once. The batch size is set considering the memory capacity and computational efficiency, usually selected between tens and hundreds.

[0090] The training monitoring records the evolution curves of the loss function value and the validation error. The training set loss reflects the model fitting ability, and the validation error indicates the generalization performance. The synchronous decline of the two indicates that the training is effective, and if it deviates, overfitting may occur. The early stopping mechanism triggers when the validation error continues to deteriorate, and the current optimal model parameters are saved. In the model evaluation stage, the final performance is based on the validation error at the time of termination, which reflects the model's prediction ability on unseen data. The saved model parameters contain complete network structure information and can be directly loaded for prediction tasks.

[0091] In the training algorithm implementation, batch samples are organized into tensor structures to accelerate computation using matrix operations. Forward propagation transforms input tensors layer by layer, and each layer operation is executed in parallel. The activation function application maintains the dimension unchanged and only performs nonlinear processing on element values. Loss calculation compares the predicted tensor with the label tensor element by element and calculates the batch average error. Backpropagation starts from the output layer, and the computation graph automatically records the operation dependency relationship. Gradient calculation is achieved through the chain rule, and the gradients of each layer are accumulated and propagated. Parameter update applies the scaled gradient to the current parameter value by the learning rate, and the momentum term considers the historical update direction.

[0092] The learning rate scheduling strategy dynamically adjusts the update step size during training. The initial learning rate is set high to promote rapid convergence. As the training progresses, the learning rate is gradually decayed to improve parameter optimization accuracy. The decay strategy can be stepwise or continuous, and the adjustment frequency is related to the training progress. The momentum coefficient remains stable, balancing the influence of the current gradient and the historical direction. Early stopping is based on the error trend in the sliding window, and the window size matches the training stability. Model saving records training configuration parameters, including learning rate strategy, momentum coefficient, and other hyperparameters.

[0093] The training process visualization tool displays the loss curve and validation error in real time. The horizontal axis is the number of training periods, and the vertical axis is the error value. The training set loss curve usually monotonically decreases, and the validation error curve may fluctuate. The relative position of the two curves reflects the model fitting state, and a large gap indicates the risk of overfitting. The early stopping point is selected at the lowest validation error, and the model parameters at this time are saved. The final model performance evaluation uses an independent test set that is completely excluded from any parameter adjustment during the training process. The test set error serves as an objective indicator of the model's true generalization ability.

[0094] Infrared temperature measuring instrument installation position coordinate set corresponding to the measured temperature value at the current time, deploy high-precision infrared temperature measuring device at a specific position on the surface of the shell. Take a certain time as an example, when the process is carried out to the 120th second, the system automatically triggers the temperature collection instruction. Four infrared temperature measuring instruments installed at coordinate points A(152.3, 86.7, 0), B(152.3, 263.1, 0), C(457.8, 86.7, 0), and D(457.8, 263.1, 0) synchronously collect surface temperature data, and the sampling frequency is set to 20 Hz. After filtering processing, the median value of the continuous 5 sampling points is taken as the stable temperature value at this time. The measurement data is transmitted to the processing unit through the industrial bus to form the current time measurement point data group. The data group contains three core fields of measurement point number, three-dimensional coordinates and temperature value, and the data structure is shown in Table 1.

[0095] Table 1: Measured reference measurement point temperature data record table.

[0096] Measurement point number X coordinate (mm) Y coordinate (mm) Z coordinate (mm) Temperature value (°C) A001 152.3 86.7 0.0 843.2 A002 152.3 263.1 0.0 796.5 A003 457.8 86.7 0.0 812.7 A004 457.8 263.1 0.0 778.9

[0097] The current time measurement point data group is input into the pre-trained multi-layer perception model. The model input layer receives a joint vector composed of measurement point coordinates and temperature values, and the vector dimension is 4 measurement points x (3-dimensional coordinates + 1-dimensional temperature value) = 16 dimensions. During the forward propagation of the model, the data is subjected to nonlinear transformation by three hidden layers, and the number of hidden layer nodes is 128, 64, and 32, respectively. The final output layer generates a predicted temperature difference field of the non-measurement point region, which covers the surface and interior of the shell. Take an internal grid point P(305.5, 175.2, 15.8) as an example, the model outputs a predicted temperature difference of +28.6°C relative to the nearest reference measurement point A002, indicating that the temperature of this point is expected to be 28.6°C higher than the temperature of measurement point A002.

[0098] When reconstructing the complete three-dimensional temperature field, first, the predicted temperature difference field is mapped to the three-dimensional spatial coordinate system according to the grid topology structure. For each grid integral point, perform temperature superposition calculation: take the measured temperature value of the nearest reference measurement point as the base value, and superimpose the predicted temperature difference corresponding to the integral point. The three-dimensional Euclidean distance formula is used for distance calculation, and the nearest neighbor search is accelerated by the spatial index tree. For example, the nearest reference measurement point of grid point Q(280.1, 200.5, 10.2) is A002(152.3, 263.1, 0), the measured temperature is 796.5°C, the predicted temperature difference is +15.3°C, and the superimposed temperature value is 811.8°C.

[0099] In the spatial interpolation and smoothing process of the superimposed results, an interpolation method based on radial basis functions is adopted. A Gaussian kernel function is selected as the basis function, and the kernel radius parameter is set to 1.5 times the average size of the grid. The interpolation process traverses all grid cell vertices, and the vertex temperature is calculated according to the temperature values of the adjacent integral points. For example, in the eight vertices of a certain hexahedral element, the interpolation calculation of vertex V1(290.0, 180.0, 0.0) involves 12 surrounding integral points, each point is given a distance inverse weight. After interpolation, a continuous distributed three-dimensional temperature field is generated, and the temperature data is stored in a three-dimensional array in floating point format, with the same dimension as the original grid division.

[0100] In the temperature field visualization stage, the three-dimensional array data is converted into an isothermal surface rendering model. The temperature gradient is set to 10℃, and different temperature intervals are assigned different color coding. For example, the 800-810℃ interval is displayed in orange-red, and the 810-820℃ interval is displayed in bright yellow. Through the acceleration of the graphics processor, a rotatable and scalable three-dimensional temperature field model is generated. The model also marks the positions of the reference measurement points and their measured temperature values, which are used to verify the local consistency of the reconstruction results.

[0101] In the data acquisition link, the installation position of the infrared thermometer strictly follows the coordinate set determined in Example 1. Each measurement point is equipped with an independent temperature calibration module to compensate for environmental radiation interference in real time. The temperature acquisition timestamp is synchronized with the process control system to ensure time consistency. The data transmission uses a redundancy check mechanism to avoid communication errors affecting data quality. When building the measurement point data group, the coordinate values are kept to millimeter accuracy, and the temperature values are kept to one decimal place.

[0102] The model input vector construction follows a fixed order: the three-dimensional coordinates and temperature values of each measurement point are stored continuously in ascending order of measurement point number. The input vector is standardized before entering the model, which subtracts the mean value of the training data and divides by the standard deviation. The model inference process is executed on a dedicated computing device, and the single prediction time is controlled within milliseconds. The predicted temperature difference field output is an N×1-dimensional vector, where N represents the total number of grid integral points, and the vector elements correspond to the grid number order.

[0103] The temperature superposition operation is performed point by point in the spatial grid system, and the nearest reference measurement point index is quickly obtained through the pre-constructed spatial position mapping table, which is generated during the grid division stage. The superposition calculation uses parallel processing, processing tens of thousands of grid point data simultaneously. In the spatial interpolation stage, first, detect the temperature mutation area (the temperature difference between adjacent cells exceeds 50℃), and automatically reduce the interpolation kernel radius in these areas to maintain the temperature gradient characteristics. The final three-dimensional temperature field data is stored as a hierarchical structure, with each layer corresponding to a specific Z-axis coordinate plane, facilitating the generation of subsequent two-dimensional heat maps.

[0104] The reconstruction result verification link is arranged at a non-measurement point position to perform sampling detection. For example, a K-type thermocouple installed at a coordinate (350.0, 150.0, 5.0) measures an instantaneous temperature of 817.4°C, and the reconstruction model outputs a temperature of 815.9°C at the position, and the difference between the two is within the measurement error range. Such sampling verification points are uniformly distributed in different regions of the shell, and the verification data are independently stored and do not participate in model training. The dynamic updating frequency of the three-dimensional temperature field is synchronized with the infrared temperature acquisition frequency, so that the temperature field is reconstructed in real time. After each reconstruction is completed, the system automatically records key parameters: the average temperature of the reference measurement point, the coordinate of the highest temperature position, the maximum temperature gradient, and other process characteristic quantities.

[0105] An abnormality detection mechanism is set during the reconstruction process. When the predicted temperature difference exceeds the range of the training data (such as more than ±60°C), the abnormal grid region is automatically marked. The spatial coordinates of the abnormal region are fed back to the control system in real time to trigger process parameter adjustment or equipment inspection instructions. At the same time, the reconstruction result is compared with the predicted value of the finite element simulation in terms of trend, and if a significant difference is found in the overall temperature distribution pattern, a model retraining suggestion is generated. All reconstruction data are stored in time sequence to form a complete temperature field evolution database during the drawing and stretching process.

[0106] In example 5, an axial-circumferential two-dimensional coordinate system is established when generating a two-dimensional heat distribution map. The axial coordinate is defined as the projection length of the shell center axis, and the numerical range is the actual geometric length from the starting end to the ending end. The circumferential coordinate uses the developed circular arc angle to represent, and the angle range is from zero to three hundred and sixty degrees, which is continuously distributed and corresponds to the complete development of the shell circumference. The center point data of each grid element in the three-dimensional temperature field is mapped to the two-dimensional coordinate system. The mapping process directly takes the axial coordinate of the three-dimensional point as the two-dimensional vertical coordinate and the circumferential angle of the three-dimensional point as the two-dimensional horizontal coordinate according to the spatial geometric relationship. The temperature value of the center point of each grid element is assigned to the corresponding two-dimensional coordinate position to form a discrete temperature data point set.

[0107] A bilinear interpolation algorithm is used to process the temperature blank area of the two-dimensional coordinate system. A local rectangular influence area is constructed with each known temperature data point as the center. The temperature value of any blank point in the area is calculated by the weighted average of the four adjacent data points. The weight distribution is determined according to the coordinate distance between the blank point and each data point, and the closer the distance, the higher the weight. The interpolation process traverses the entire two-dimensional coordinate system plane to generate a dense grid temperature distribution matrix. The number of rows corresponds to the axial resolution, and the number of columns corresponds to the circumferential resolution, and the matrix elements store the temperature values.

[0108] Temperature contours are drawn according to the interpolation results. A fixed temperature interval value is set, and the coordinate points with the same temperature value are automatically connected to form closed curves. The contour drawing uses a layer-by-layer tracking algorithm to generate from the highest temperature area to the low temperature area. Each contour is marked with the corresponding temperature value, and different temperature intervals use different line styles. The point cloud distribution map is generated synchronously, and each interpolation grid point is mapped to a specific color according to the temperature value. The color coding uses a continuous color spectrum scheme, with high temperature areas corresponding to warm colors and low temperature areas corresponding to cold colors. The color spectrum range covers the entire temperature interval, and each color accurately corresponds to a temperature value interval.

[0109] In the dynamic optimization step, the predicted temperature difference field changes in the last five reconstruction cycles are continuously monitored. Each reconstruction cycle corresponds to a fixed time interval, which is synchronized with the process rhythm. The change amplitude calculation uses a full-field temperature difference data comparison method to perform absolute difference operations on the temperature difference values of adjacent grid points in adjacent cycles. The full-field change amplitude is quantified as the average of the absolute differences of all grid points. The preset temperature difference threshold is set to a fixed value according to the process stability requirements. When the system detects that the average change amplitude exceeds this threshold, it automatically triggers the new measurement point layout instruction.

[0110] The determination of the position of the new measurement point is based on the spatial distribution analysis of the change amplitude. The largest continuous spatial region of the change amplitude is identified, and the three-dimensional geometric center coordinates of this region are calculated. Based on the center coordinates, the optimal position of the new measurement point is determined in combination with the device installation constraints. The number of new measurement points is dynamically configured according to the size of the change region, usually one to three measurement points. The coordinates of the new measurement points are added to the original reference measurement point set to form an updated list of infrared thermometry instrument installation positions.

[0111] The multilayer perceptron model is retrained using the measured temperature data of the new measurement points. In the training data preparation stage, sample data containing the coordinates of the new measurement points are extracted from the process parameter-temperature field mapping database. The sample construction rules remain the same as the initial training, and the input vector dimension is expanded with the increase in the number of measurement points. The model structure is adjusted to moderately adjust the number of hidden layer nodes according to the information amount of the new measurement points. The training process executes the same supervised learning process, including data standardization, forward propagation, loss calculation, back propagation, and parameter update steps. The training termination condition follows the judgment criterion that the validation set error does not decrease for three consecutive times. The model parameters after retraining replace the original parameters and are used for subsequent temperature field reconstruction tasks.

[0112] In the two-dimensional heat distribution map generation process, the axial coordinate scale is marked with actual length units. The circumferential coordinate scale is marked with angle values, and main scale lines are set every thirty degrees. The temperature contour line is marked with text set in the horizontal direction. The point cloud color rendering uses a smooth gradient technique to eliminate color step phenomena. The legend area clearly shows the correspondence between temperature values and colors. The chart title contains process timestamp information, which facilitates historical data tracing.

[0113] At the end of each reconstruction cycle, the temperature difference field change analysis program is automatically started. The change amplitude calculation uses an incremental storage strategy, retaining only the temperature difference field data from the last five cycles. The preset temperature difference threshold supports online adjustment, adapting to the needs of different process stages. After the addition of a new measurement point instruction, the system automatically generates a measurement point installation guide, marking the three-dimensional coordinates and installation azimuth angle of the new measurement point.

[0114] When the model retraining process is started, the original model parameters are automatically backed up. In the training data preparation stage, historical simulation data related to the location of the new measurement points are selected. When the input vector is reconstructed, the measurement point information is arranged in numerical order, with the data of the new measurement points appended after the original data. The number of hidden layer nodes is adjusted according to a specific proportion rule, usually expanded by the square root of the measurement point quantity increase. The training process log records the hyperparameter configuration and error convergence curve in detail. After training is completed, verification testing is automatically performed to ensure the effective integration of the new measurement point information.

[0115] The two-dimensional graph output module supports real-time refresh functionality, automatically updating the heat distribution map data source after each temperature field reconstruction is completed. The view interaction function allows operators to freely scale and pan the observation area. The temperature profile extraction tool supports temperature curve drawing for any axial and circumferential positions. All graphical elements are rendered in vector format, ensuring display quality at different scales.

[0116] The dynamic optimization mechanism includes a self-check program that automatically evaluates the optimization effect after each new measurement point is deployed. The evaluation method is to compare the temperature reconstruction stability in the change area before and after the addition of the new measurement point. The evaluation results are recorded in the optimization history database to guide subsequent optimization strategy adjustments. When the expected effect is not achieved, the system automatically starts a secondary optimization process, including measurement point position fine-tuning or model structure adjustment.

[0117] The temperature field reconstruction results drive monitoring and analysis, the monitoring results trigger optimization instructions, and the optimization results are fed back to the reconstruction system. The cycle period matches the process time scale, ensuring that the temperature field reconstruction accuracy adapts to the process dynamic change characteristics. All optimization operations are recorded with detailed timestamps and operation parameters, forming a complete temperature field reconstruction optimization audit trail.

[0118] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0119] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for reconstructing the stretching temperature field based on neural network mapping, characterized in that, include: A mapping relationship between drawing process parameters and three-dimensional temperature field distribution data is established. The drawing process parameters include initial billet temperature, loading speed, friction factor, and boundary heat conduction conditions. The three-dimensional temperature field distribution data is obtained by dividing the billet into equally spaced grids and extracting the temperature value of the unit center point. The layout scheme of the reference measuring points is determined based on the temperature gradient distribution, and the temperature values ​​of the reference measuring points are extracted from the three-dimensional temperature field distribution data. Calculate the temperature difference between the non-measuring point area and the temperature value of the reference measuring point, and construct a measuring point coordinate-temperature difference field sample pair; The sample pairs of measuring point coordinates and temperature difference fields are processed by a multilayer perceptron model. The input of the multilayer perceptron model is the joint vector of measuring point coordinates and temperature values, and the output is the predicted temperature difference field. The predicted temperature difference field is superimposed on the reference measurement point temperature value to reconstruct a complete three-dimensional temperature field. A geometric mapping process is performed on the complete three-dimensional temperature field to generate a two-dimensional thermal distribution map. The vertical axis of the two-dimensional thermal distribution map is the axial length, and the horizontal axis is the circumferential unfolding angle.

2. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 1, characterized in that, The establishment of the mapping relationship between drawing process parameters and three-dimensional temperature field distribution data includes: Configure multiple process vector groups, each containing independently configured initial billet temperature parameters, loading speed parameters, friction factor parameters, and boundary heat conduction condition parameters; Perform thermo-mechanical coupled finite element simulation on each set of process vectors and output billet temperature field data at multiple simulation times; The surface and internal area of ​​the billet are divided into equally spaced grids along the axial and circumferential directions, and the spatial coordinates and corresponding temperature values ​​of the center points of the grid units are extracted. The spatial coordinates and corresponding temperature values ​​are stored according to the simulation time sequence to construct a process parameter-temperature field mapping database.

3. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 2, characterized in that, The scheme for determining the layout of reference measuring points based on temperature gradient distribution includes: Extract temperature field data for each simulation moment from the process parameter-temperature field mapping database; Calculate the weighted sum of the axial and circumferential temperature change rates in the temperature field data to generate a temperature gradient distribution map; Identify continuous regions in the temperature gradient distribution map where the gradient value is greater than a preset threshold and mark them as high gradient regions; Based on the spatial distribution density of high gradient regions and the constraints of equipment accessibility, a set of coordinates for the installation location of infrared thermometers is configured.

4. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 3, characterized in that, The constructed measurement point coordinate-temperature difference field sample pairs include: Read the set of grid integration point coordinates and corresponding temperature values ​​at the target simulation time from the process parameter-temperature field mapping database; Extract the temperature values ​​of the reference measuring points corresponding to the set of coordinates of the infrared thermometer's installation location; Calculate the real-time temperature difference between the temperature value at each grid integration point and the temperature value at the nearest reference measurement point; Combine the coordinates of the reference measuring point, the temperature value of the reference measuring point, and the coordinates of the grid integration point into an input vector, and set the corresponding real-time temperature difference as the output label. The input vector and output label are stored in the order of grid integration point numbers to form a temperature difference training sample set.

5. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 4, characterized in that, The process of processing the measurement point coordinate-temperature difference field sample pairs using a multilayer perceptron model includes: Construct a fully connected network structure that includes an input layer, at least three hidden layers, and an output layer; The input vector is input to the input layer and then undergoes a nonlinear transformation through the hidden layer nodes. Generate predicted temperature differences in the output layer; The hidden layer uses a modified linear unit activation function, and the output layer uses a linear activation function.

6. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 5, characterized in that, It also includes performing supervised training on the multilayer perceptron model: The temperature difference training sample set is split into a training subset and a validation subset according to a preset ratio; The training subset input vectors are used as model inputs, and the training subset output labels are used as supervision signals. Calculate the mean square error between the predicted temperature difference and the output label; Update network weight parameters using the stochastic gradient descent algorithm; If the mean squared error of the validation subset fails to decrease for three consecutive iterations, training is stopped and the network parameters are saved.

7. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 6, characterized in that, Also includes: Collect the measured temperature value corresponding to the set of coordinates of the infrared thermometer's installation location at the current moment; Combine the measured temperature value with the corresponding coordinates to form the measurement point data set for the current time. Input the current measurement point data set into the trained multilayer perceptron model to obtain the predicted temperature difference field of the non-measurement point area output by the model.

8. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 7, characterized in that, The reconstructed complete three-dimensional temperature field includes: The predicted temperature difference field is mapped to three-dimensional space according to the coordinates of the grid integration points; Temperature superposition is performed on each grid integration point in three-dimensional space: the predicted temperature difference corresponding to the integration point is superimposed on the measured temperature value of the nearest benchmark measurement point as the base value. Spatial interpolation smoothing is performed on the superposition results to generate a continuously distributed three-dimensional temperature field.

9. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 8, characterized in that, The generation of the two-dimensional heat map includes: Establish an axial-circumferential two-dimensional coordinate system, where the axial coordinate is the projected length of the billet's central axis and the circumferential coordinate is the unfolded arc angle; Map the temperature value at the center point of each grid cell in the three-dimensional temperature field to the two-dimensional coordinate system; The temperature blank area in the two-dimensional coordinate system is filled by bilinear interpolation algorithm, and temperature contour lines and point cloud distribution map are drawn based on the interpolation results.

10. The method for reconstructing the stretching temperature field based on neural network mapping as described in claim 9, characterized in that, It also includes dynamic optimization steps: Monitor the magnitude of changes in the predicted temperature difference field over five consecutive reconstruction cycles; When the change exceeds the preset temperature difference threshold, a command to add new measuring points is triggered. Based on the spatial coordinates of the area with the largest change, the baseline measurement point layout scheme is updated again, and the multilayer perceptron model is retrained using the measured data after the addition of measurement points.

Images