Intelligent monitoring method and system for forging production

Abstract

The invention relates to the technical field of intelligent monitoring, and discloses an intelligent monitoring method and system for forging production. The method comprises the following steps: carrying out temperature data acquisition and polar coordinate conversion on the wind power flange forge piece to obtain polar coordinate temperature field data; inputting the polar coordinate temperature field data into an annular thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector; performing phase change critical temperature sensitivity enhancement on the temperature field feature vector to obtain a phase change sensitivity enhanced feature vector; on the basis of heat conduction partial differential equation constraints, submerged space reconstruction and interpolation are carried out on the phase change sensitivity enhancement feature vectors, and full-field temperature distribution data are obtained; and generating temperature field anomaly score and anomaly type probability distribution information according to the full-field temperature distribution data. The monitoring problem that the surface temperature is normal but the dangerous temperature gradient exists in the depth direction is effectively solved, and the production qualification rate of wind power flange forgings is increased.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring technology, and in particular to an intelligent monitoring method and system for forging production. Background Art

[0002] Traditional temperature field monitoring technology for wind turbine flange forgings during heat treatment primarily relies on distributed thermocouple arrays and simple interpolation algorithms for temperature data collection and processing. This approach employs a limited number of temperature sensors placed on and within the forging, using mathematical methods such as linear or spline interpolation to estimate the temperature distribution between the sparse measurement points. The heat treatment process is then judged as normal based on a preset temperature threshold. Existing monitoring systems typically utilize a Cartesian coordinate system for data processing and traditional convolutional neural networks or simple statistical analysis methods to detect temperature anomalies.

[0003] However, existing technologies have significant limitations when monitoring temperature anomalies in large annular heterogeneous structures. First, for large wind turbine flange forgings with flange diameters of 4-7 meters, the data sparsity problem caused by limited temperature measurement points is more prominent, making it impossible to effectively capture the dynamic characteristics of thermal flow and phase transformation within the forgings. Second, traditional monitoring methods lack sensitivity for detecting minor anomalies in the critical phase transformation temperature range of 590-720°C, making it difficult to promptly detect temperature anomalies that may lead to uneven internal structure, excessive residual stress, or even microcracks in the forgings. Finally, existing systems lack the ability to effectively monitor temperature gradients in the depth direction, making it impossible to identify temperature standing waves, where the surface temperature is normal but dangerous temperature gradients in the depth direction exist.

[0004] How to develop a more adaptive temperature field monitoring and data processing method for the annular geometric features of wind turbine flanges; how to achieve sensitive detection and enhancement of small anomalies in the critical temperature range of phase transformation under data-sparse conditions; how to combine physical constraints with deep learning to improve the accuracy and physical consistency of temperature field reconstruction; and how to establish a multi-scale anomaly feature extraction mechanism to identify hidden temperature anomaly patterns, including "standing temperature waves." These technical challenges make it difficult for traditional monitoring methods to detect potential defects in a timely manner. Problems are often discovered through quality inspection after heat treatment is completed, resulting in increased scrap rates for high-value forgings and reduced production efficiency. Summary of the Invention

[0005] The present invention provides an intelligent monitoring method and system for forging production, which effectively solves the monitoring problem of normal surface temperature but dangerous temperature gradient in depth direction, and improves the production qualification rate of wind power flange forgings.

[0006] In a first aspect, the present invention provides an intelligent monitoring method for forging production, the intelligent monitoring method for forging production comprising: Perform temperature data collection and polar coordinate conversion on wind turbine flange forgings to obtain polar coordinate temperature field data; Inputting the polar coordinate temperature field data into a circular thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector; Performing phase change critical temperature sensitivity enhancement on the temperature field characteristic vector to obtain a phase change sensitivity enhancement characteristic vector; Based on the constraints of the heat conduction partial differential equation, the phase change sensitive enhancement eigenvector is reconstructed and interpolated in the latent space to obtain the full-field temperature distribution data; A temperature field anomaly score and anomaly type probability distribution information are generated according to the full-field temperature distribution data.

[0007] Optionally, in a first implementation of the first aspect of the present invention, the temperature data of the wind turbine flange forging is collected and polar coordinate converted to obtain polar coordinate temperature field data, including: The wind turbine flange forging is radially divided into three regions including an outer ring, a middle ring, and an inner ring. Temperature measuring points are evenly arranged circumferentially in each region, and K-type thermocouples are installed at the temperature measuring points to obtain a multi-layer temperature measuring point structure. Performing real-time data acquisition on the K-type thermocouples in the multi-layer temperature measurement point structure to obtain raw temperature data, and performing noise elimination and smoothing processing on the raw temperature data to obtain filtered temperature data; Converting the filtered temperature data from a Cartesian coordinate system to a polar coordinate system to obtain a temperature field matrix represented by polar coordinates; Bicubic spline interpolation and standardization are performed on the temperature field matrix represented by polar coordinates to obtain polar coordinate temperature field data.

[0008] Optionally, in a second implementation of the first aspect of the present invention, inputting the polar coordinate temperature field data into a ring thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector includes: Inputting the polar coordinate temperature field data into the input layer of the annular thermal field gradient decomposition network for feature encoding to obtain an input feature vector; The input feature vector is sequentially processed through four series-connected polar coordinate convolution modules in the annular thermal field gradient decomposition network to obtain a multi-scale feature map, wherein each polar coordinate convolution module includes two polar coordinate convolution layers, a batch normalization layer, and a LeakyReLU activation function; Inputting the multi-scale feature map into the annular self-attention module in the annular thermal field gradient decomposition network for processing to obtain an enhanced feature map; Fusing the enhanced feature map with the feature map of the corresponding layer in the multi-scale feature map to obtain a fused multi-scale temperature gradient feature; The fused multi-scale temperature gradient features are input into the global polar coordinate pooling layer in the annular thermal field gradient decomposition network for processing to obtain a temperature field feature vector.

[0009] Optionally, in a third implementation of the first aspect of the present invention, performing phase transition critical temperature sensitivity enhancement on the temperature field eigenvector to obtain the phase transition sensitivity enhancement eigenvector includes: Converting the temperature field characteristic vector into a temperature domain space to obtain reconstructed temperature field data, and constructing a temperature sensitivity weight function based on the reconstructed temperature field data and the phase change temperature range of the wind power flange forging; Calculating the proximity between the temperature value of each point in the reconstructed temperature field data and the phase change critical temperature based on the temperature sensitivity weight function to generate a sensitivity map, and performing enhancement processing on the sensitivity map to obtain an enhanced sensitivity map; Multi-scale temperature gradient calculation and fusion are performed on the reconstructed temperature field data by using polar coordinate gradient operators with different kernel sizes in a multi-scale temperature gradient pyramid to obtain a comprehensive gradient feature map; A pixel-level product operation is performed on the enhanced sensitivity map and the comprehensive gradient feature map to obtain a phase change sensitive gradient map, and the phase change sensitive gradient map and the enhanced sensitivity map are fused to obtain a phase change sensitive enhanced feature vector.

[0010] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing latent space reconstruction and interpolation on the phase change sensitive enhancement eigenvector based on the heat conduction partial differential equation constraint to obtain full-field temperature distribution data includes: Inputting the phase change sensitive enhanced feature vector into an encoder in a latent space reconstruction algorithm for dimensionality reduction processing to obtain a low-dimensional latent space variable; Upsampling and reconstructing the low-dimensional latent space variables by a decoder in the latent space reconstruction algorithm to obtain first temperature field data; Applying a heat conduction partial differential equation constraint to perform an iterative calculation of minimizing a reconstruction error and a physical constraint error on the first temperature field data to obtain second temperature field data; Perform space filling and sparse region completion processing on the second temperature field data to obtain full-field temperature distribution data.

[0011] Optionally, in a fifth implementation of the first aspect of the present invention, applying the heat conduction partial differential equation constraint to perform iterative calculations to minimize reconstruction error and physical constraint error on the first temperature field data to obtain the second temperature field data includes: Calculating a temperature time derivative and a space derivative based on the first temperature field data, and generating temperature field partial derivative data based on the temperature time derivative and the space derivative; The material composition vector and local temperature information of the wind turbine flange forging are input into the thermal diffusion coefficient network for processing to obtain the thermal diffusion coefficient in spatial variational form. Inputting the first temperature field data into a heat source function network for processing to obtain an internal heat source function; Applying heat conduction partial differential equation constraints to construct a physical constraint loss function according to the temperature field partial derivative data, the thermal diffusion coefficient and the internal heat source function; The physical constraint loss function, the reconstruction loss function and the regularization term are combined into a total loss function, and the total loss function is minimized and iteratively optimized using a gradient descent method to obtain second temperature field data.

[0012] Optionally, in a sixth implementation of the first aspect of the present invention, generating the temperature field anomaly score and anomaly type probability distribution information based on the full-field temperature distribution data includes: Performing feature calculations on the full-field temperature distribution data at three scales, namely, macroscopic, mesoscopic, and microscopic, calculating the global temperature standard deviation and radial temperature gradient at the macroscopic scale, calculating the local temperature mean, standard deviation, and skewness at the mesoscopic scale, and calculating the temperature change rate and temperature acceleration at the microscopic scale, to obtain a multi-scale temperature feature set; Extracting a temperature profile along the depth direction of the full-field temperature distribution data, and calculating the depth direction temperature gradient and depth gradient variance based on the temperature profile, identifying the temperature standing wave phenomenon where the surface temperature is normal but a dangerous temperature gradient exists in the depth direction, and obtaining the depth anomaly feature; Calculating a temperature field anomaly score for each spatial point based on the multi-scale temperature feature set and the depth anomaly feature; The area where the temperature field anomaly score exceeds the anomaly warning threshold is segmented to obtain segmented areas, and the target feature vector of each segmented area is calculated. By comparing and analyzing the target feature vector with the historical defect database, temperature field anomaly type probability distribution information including anomaly type and probability distribution is generated.

[0013] In a second aspect, the present invention provides an intelligent monitoring system for forging production, the intelligent monitoring system for forging production comprising: The acquisition module is used to collect temperature data of wind power flange forgings and perform polar coordinate conversion to obtain polar coordinate temperature field data; An extraction module is used to input the polar coordinate temperature field data into a ring thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector; An enhancement module, configured to perform phase change critical temperature sensitivity enhancement on the temperature field characteristic vector to obtain a phase change sensitivity enhancement characteristic vector; A reconstruction module is used to perform latent space reconstruction and interpolation on the phase change sensitive enhancement eigenvector based on the constraints of the heat conduction partial differential equation to obtain full-field temperature distribution data; A generation module is used to generate temperature field anomaly scores and anomaly type probability distribution information based on the full-field temperature distribution data.

[0014] The technical solution provided by this invention, through polar coordinate transformation and polar coordinate convolutional layer design, is specifically optimized for the annular geometry of wind turbine flanges. This effectively addresses the shortcomings of traditional convolutional networks in handling the continuity of temperature gradients between the inner and outer rings of wind turbine flanges, making temperature field monitoring more consistent with the characteristics of annular structures. Using a phase transition critical temperature sensitivity enhancement module, the invention focuses on monitoring and enhancing subtle anomalies within the critical phase transition temperature range of 590-720°C, enabling the system to more sensitively detect temperature anomalies during phase transitions. Using a latent space reconstruction algorithm constrained by partial differential equations of heat conduction, the invention organically combines deep learning with the laws of thermodynamics, ensuring that the reconstructed temperature field conforms to the laws of heat conduction and improving the accuracy of temperature prediction in data-sparse areas. By extracting anomaly features at three scales, namely macroscopic, mesoscopic, and microscopic, the invention comprehensively analyzes temperature field anomalies, particularly significantly improving the ability to identify "standing temperature waves," effectively solving the difficult problem of monitoring situations where surface temperatures are normal but dangerous temperature gradients exist at depth. This invention can identify potential defect areas at an early stage, providing operators with ample time for intervention, effectively preventing defects from forming and improving the production yield of wind turbine flange forgings. It can also precisely locate areas of temperature anomalies, especially those around bolt holes, to monitor even small temperature anomalies in critical structural areas, providing precise spatial positioning information for defect prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 A schematic diagram of an embodiment of an intelligent monitoring method for forging production according to an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of an intelligent monitoring system for forging production in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] An embodiment of the present invention provides an intelligent monitoring method and system for forging production. The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products or apparatus.

[0018] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 An embodiment of the intelligent monitoring method for forging production in the embodiment of the present invention includes: Step S101: collecting temperature data of a wind turbine flange forging and performing polar coordinate conversion to obtain polar coordinate temperature field data; It is understandable that the execution subject of the present invention can be an intelligent monitoring system for forging production, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0019] Specifically, a temperature monitoring structure was constructed based on a spatial layout. The forging was radially divided into three regions: the outer ring, the middle ring, and the inner ring. Temperature measurement points were evenly distributed around each region and stacked in layers along the thickness direction to form a three-dimensional temperature measurement point structure. K-type thermocouples were used as sensing elements, and raw temperature data was collected in real time at a sampling frequency of 5 Hz. The Savitzky-Golay filter algorithm was used to suppress noise and smooth the collected data, eliminating high-frequency disturbances while preserving the temperature trend. Considering the axisymmetric nature of the forging, the data was converted from a Cartesian coordinate system to a polar coordinate system, expressed in the form of T(r,θ,z,t), to better align the data structure with the flange geometry. Bicubic spline interpolation was used for spatial interpolation, increasing the angular resolution to 1° and the radial resolution to 10 mm. Finally, the temperature field data was normalized and mapped to the range 0 to 1, resulting in polar coordinate temperature field data that can be directly input into the deep learning model.

[0020] Step S102: inputting polar coordinate temperature field data into an annular thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector; Specifically, the polar coordinate temperature field tensor is input into the input layer of the network for initial feature encoding processing. This input layer converts the five-dimensional input tensor (batch size, number of channels, radial dimension, angular dimension, and depth dimension) into a unified input feature vector form, enabling the subsequent convolution module to perform spatial structure perception and local pattern extraction. After completing the preliminary encoding, the temperature field data is sequentially processed through the main network structure consisting of four serially connected polar coordinate convolution modules. Each convolution module contains two polar coordinate convolution layers, a batch normalization layer, and a LeakyReLU activation function. The convolution layer uses kernel weights designed in polar coordinate format to adapt to the periodicity of the ring structure and ensure feature continuity between 0° and 360° in the angular dimension; the batch normalization layer is used to stabilize the distribution differences between batches and improve the convergence speed of network training; the LeakyReLU activation function introduces a non-zero slope when processing negative interval data to alleviate the gradient vanishing problem. In terms of channel count, the first polar convolution module has 32 output channels, and the output channels of each subsequent module are doubled to 64, 128, and 256, respectively, forming a multi-scale feature map structure. To improve the model's sensitivity to key regions of the temperature field and its ability to express long-range dependencies, the multi-scale feature map is fed into a circular self-attention module for global enhancement. This module, through three 1×1 convolutional layers, maps the query Q, key K, and value V matrices, respectively. An attention mechanism builds relationships between spatial locations, enabling the network to generate more significant responses to temperature anomalies. The enhanced feature map is fused with the original feature map from each layer of the multi-scale convolution module. Feature concatenation is performed using skip connections, effectively combining low-level details with high-level abstract semantics, improving the model's ability to resolve complex temperature gradient fields at multiple scales. Finally, the fused multi-scale temperature gradient features are fed into a global polar pooling layer, which performs a weighted average of the three-dimensional spatial structure using polar coordinates, outputting a temperature field feature vector with the dimensions (batch size, 256, 1, 1, 1).

[0021] Step S103, performing phase change critical temperature sensitivity enhancement on the temperature field eigenvector to obtain a phase change sensitivity enhancement eigenvector; Specifically, the temperature field feature vectors are mapped back to the temperature domain through a fully connected neural network layer, restoring their spatial structure and numerical physical meaning, and outputting reconstructed temperature field data with the same spatial resolution as the original input. Based on the reconstructed temperature field data and the typical phase transition temperature ranges of wind turbine flange steel (590°C, 650°C, and 720°C), a Gaussian-distributed temperature sensitivity weight function is constructed. This function calculates the exponential distance between the temperature value of each spatial point and the phase transition critical temperature. A width control parameter σ is introduced to adjust the distribution range of the response curve. The response is calculated for each spatial location in the reconstructed temperature field data, resulting in an original sensitivity map with a range between 0 and 1. A larger value indicates that the temperature at that point is closer to the phase transition critical temperature. To enhance the model's ability to discern subtle fluctuations near key points, a hyperbolic tangent activation function is used to perform nonlinear amplification on the sensitivity map, with the slope parameter β controlling the amplification amplitude. A multi-scale temperature gradient pyramid structure is also constructed. Four polar coordinate gradient operators with different kernel sizes (3×3, 5×5, 7×7, and 9×9) are sequentially applied to the reconstructed temperature field to extract gradient information at different scales. These gradient results are normalized and then fed into a feature pyramid fusion network. Upsampling and concatenation create a comprehensive gradient feature map, covering temperature distribution patterns from micro-local perturbations to macro-gradient anomalies. The enhanced sensitivity map is pixel-wise multiplied with the comprehensive gradient feature map to generate a phase change sensitivity gradient map, achieving dual coupling of temperature change intensity and critical response weights. Finally, the phase change sensitivity gradient map and the enhanced sensitivity map are channel-wise fused, and a channel-weighted strategy is employed to obtain the final phase change sensitivity enhancement feature vector.

[0022] Step S104: Based on the constraints of the heat conduction partial differential equation, the phase change sensitive enhancement feature vector is reconstructed and interpolated in the latent space to obtain the full-field temperature distribution data; Specifically, the phase change sensitive enhancement feature vector is input into the encoder module of the latent space reconstruction algorithm for dimensionality compression. The encoder uses a multi-layer polar coordinate convolution structure to extract deep semantic information layer by layer through a downsampling operation with a step size of 2. While maintaining spatial structural features, it compresses the data dimension and maps the input high-dimensional features to a low-dimensional latent space with a dimension of 128. This latent variable retains important features such as temperature gradient, sensitivity enhancement, and structural boundaries. The low-dimensional latent space variable is input into the decoder module and reversely expanded through a multi-layer upsampling operation with a symmetrical structure. The decoder uses transposed polar coordinate convolution combined with instance normalization and PReLU activation function to gradually reconstruct the spatial distribution features and restore the first temperature field data consistent with the original spatial resolution.

[0023] The first temperature field data is iteratively calculated using the partial differential equation for heat conduction to minimize the reconstruction error and the physical constraint error. This optimization process, based on the governing heat conduction equation, approximates the diffusion evolution of the temperature field using finite differences, including forward and central differences. The model introduces a spatially varying thermal diffusion coefficient function, which is inferred by a perceptron network based on local temperature and material chemical composition parameters (C%, Si%, Mn%, Cr%, Mo%, and Ni%) to reflect the variations in thermal diffusion rates across different regions. The internal heat source term is modeled using a neural network based on a U-Net architecture, which outputs a spatial heat source distribution based on the historical temperature sequence and the current temperature state. By alternately updating the reconstruction results and the equation constraint objectives, the model iteratively generates second temperature field data that conforms to the observed data fit and the physical laws of heat conduction.

[0024] Based on the first temperature field data, partial derivatives are calculated. The temperature-time derivative is calculated using forward differencing. Second-order derivatives of the spatial three-dimensional grid are calculated using a central difference scheme to construct the temperature field partial derivative data. The material composition vector and local temperature information are input into the thermal diffusion coefficient network, which uses a multi-layer perceptron structure to output the thermal diffusion coefficient in spatial variational form. The first temperature field data is input into the heat source function network, which uses a U-Net architecture to fuse the current temperature state with the historical temperature evolution sequence, outputting an internal heat source function with spatiotemporal correlation characteristics. The temperature partial derivative data, thermal diffusion coefficient, and internal heat source function are introduced into the partial differential equation for heat conduction, constructing a physical constraint loss function. This function is combined with the reconstruction loss function and an L1 regularization term to form the total loss function. Iterative optimization using gradient descent minimizes the model, ensuring convergence towards both improved reconstruction accuracy and enhanced physical consistency, ultimately obtaining the second temperature field data. Based on the second temperature field data, the model performs space filling and sparse region completion, combining latent space interpolation capabilities with spatial continuity learning mechanisms to ultimately output the full-field temperature distribution data.

[0025] Step S105: Generate temperature field anomaly scores and anomaly type probability distribution information based on the full-field temperature distribution data.

[0026] Specifically, a multi-scale spatial analysis of the full-field temperature distribution data reveals temperature anomaly characteristics at different levels. At the macroscale, the overall temperature standard deviation is calculated to reflect the uniformity of the thermal field, and the radial temperature gradient is calculated to capture the spatial offset trend of large-scale heat distribution. At the mesoscale, the sliding window method is used to scan the temperature of local areas, and the temperature mean, standard deviation, and skewness of each local area are calculated to capture regional thermal anomalies such as local overheating or uneven cooling. At the microscale, the temperature change rate and temperature acceleration of each spatial point are analyzed to identify abnormal temperature fluctuations or short-term drastic changes, forming a multi-scale temperature feature set.

[0027] To identify "standing temperature waves" (normal surface temperature but abrupt temperature changes at depth), we extract temperature profiles at multiple radial and angular locations from the full-field temperature distribution data. We analyze the temperature variation through the thickness and calculate the temperature gradient and its variance along the depth. When the surface temperature is stable but a sharp temperature change occurs below, we identify it as a deep anomaly.

[0028] Based on a multi-scale temperature feature set and deep anomaly features, a multivariate anomaly assessment model is used to calculate a temperature field anomaly score for each spatial point. A larger value indicates more abnormal temperature behavior. A density clustering algorithm is applied to points with anomaly scores above the warning threshold to segment the abnormal region. Each abnormal region is encoded as a target feature vector containing parameters such as the average anomaly score, maximum anomaly value, area range, maximum temperature deviation, temperature gradient amplitude, and depth gradient statistics. By performing similarity calculation and multi-dimensional matching with a historical defect database, a specific anomaly type label is assigned to each abnormal region. A probability distribution is generated based on the matching confidence, forming a probability distribution information for the temperature field anomaly type.

[0029] In a specific embodiment, the process of executing step S101 may specifically include the following steps: The wind turbine flange forging is divided into three areas along the radial direction, including the outer ring, the middle ring, and the inner ring. Temperature measuring points are evenly arranged circumferentially in each area, and K-type thermocouples are installed at the temperature measuring points to obtain a multi-layer temperature measuring point structure. Real-time data acquisition is performed on the K-type thermocouples in the multi-layer temperature measurement point structure to obtain the original temperature data, and the original temperature data is subjected to noise elimination and smoothing processing to obtain the filtered temperature data; The filtered temperature data is converted from the Cartesian coordinate system to the polar coordinate system to obtain the temperature field matrix represented by polar coordinates; The temperature field matrix represented by polar coordinates is subjected to bicubic spline interpolation and normalization to obtain polar coordinate temperature field data.

[0030] Specifically, based on the annular axisymmetric characteristics of wind turbine flange forgings, radial zoning is used to divide them into three thermal characteristic zones: outer ring, middle ring, and inner ring. Temperature measuring points are evenly arranged at equal angles in the circumferential direction of each ring to ensure consistent angular data acquisition density. At the same time, temperature measuring points are laid layer by layer at fixed intervals in the thickness direction to form a three-dimensional multi-layer temperature measuring point system in the radial, angular, and depth directions. A K-type thermocouple is installed at each node as a temperature sensing unit. K-type thermocouples have good high-temperature adaptability, stability, and anti-interference capabilities, and are suitable for accurate temperature measurement at different positions, depths, and heating stages of large-diameter, heavy-mass, and high-heat-capacity forgings such as wind turbine flanges.

[0031] All K-type thermocouples are connected to a high-frequency data acquisition system with a sampling frequency set to 5Hz, enabling tracking of temperature changes with sub-second resolution. Because the measurement environment is affected by thermal radiation disturbances, electromagnetic interference, conductor thermal inertia, and thermocouple contact errors, the raw temperature data contains high-frequency noise and localized jitter. Therefore, a Savitzky-Golay filter algorithm is used for smoothing. Polynomial fitting within a local sliding window eliminates short-term fluctuations while preserving the overall trend.

[0032] Taking into account the axisymmetric characteristics of wind turbine flanges, the filtered three-dimensional temperature measurement data is converted from the Cartesian coordinate system (x, y, z) to the polar coordinate system (r, θ, z), where r is the radial position, θ is the angular position, and z is the thickness direction. Bicubic spline interpolation is performed on the polar coordinate temperature field matrix, and the continuous variation relationship between data points is reconstructed within the local plane through bidirectional cubic polynomial fitting. This improves the angular temperature distribution accuracy to 1° and the radial accuracy to 10mm. Finally, the interpolated high-resolution polar coordinate temperature field matrix is ​​normalized using a linear mapping to transform the temperature values ​​to the range of 0 to 1. The maximum value of the historical sample data is used as the normalization boundary to improve training stability and model generalization ability, and the polar coordinate temperature field data is output.

[0033] In a specific embodiment, the process of executing step S102 may specifically include the following steps: Input the polar coordinate temperature field data into the input layer of the annular thermal field gradient decomposition network for feature encoding to obtain the input feature vector; The input feature vector is processed sequentially through four series-connected polar convolution modules in the annular thermal field gradient decomposition network to obtain a multi-scale feature map, where each polar convolution module contains two polar convolution layers, a batch normalization layer, and a LeakyReLU activation function; The multi-scale feature map is input into the ring self-attention module in the ring thermal field gradient decomposition network for processing to obtain the enhanced feature map; The enhanced feature map is fused with the feature map of the corresponding layer in the multi-scale feature map to obtain the fused multi-scale temperature gradient feature; The fused multi-scale temperature gradient features are input into the global polar coordinate pooling layer in the annular thermal field gradient decomposition network for processing to obtain the temperature field feature vector.

[0034] Specifically, the polar coordinate temperature field data is input into the annular thermal field gradient decomposition network as a five-dimensional tensor (batch_size, 1, Hr, Hθ, Hz), where 1 represents a single-channel temperature value, and Hr, Hθ, and Hz represent the spatial dimensions in the radial, angular, and depth directions, respectively. The input layer extracts low-level features from the raw temperature field through a polar coordinate convolution operation, encoding the input feature vector from pure temperature values ​​into a directional gradient and local response capability.

[0035] The input feature vector is processed sequentially through four serially connected polar convolution modules. Each module consists of two polar convolution layers, a batch normalization layer, and a LeakyReLU activation function. Polar convolution uses weighted kernels that conform to polar geometry. A circular padding strategy is used in the angular dimension to preserve periodic continuity in the θ direction, while reflective padding is used in the radial dimension to prevent loss of boundary gradients. The first convolution module has 32 output channels, which are then doubled with each subsequent layer, forming a progressively enhanced structure of 64, 128, and 256 channels, generating a multi-scale feature atlas.

[0036] To enhance the model's ability to identify local thermal anomalies and model long-range dependencies, multi-scale feature maps are fed into a circular self-attention module. This module generates query Q, key K, and value V matrices through three independent 1×1 convolutional layers. It then performs the core computations of the attention mechanism: matrix dot products capture similarity relationships, softmax normalization obtains attention weights, and multiplication with the value matrix yields an enhanced feature map. Because of its polar coordinate feature structure, the attention mechanism naturally preserves angular periodicity and radial distribution patterns, enabling the enhanced feature map to precisely focus on localized areas of abnormal thermal gradients.

[0037] The network design employs a multi-scale skip connection fusion mechanism to fuse the enhanced feature maps with the corresponding level feature maps output by the four convolutional modules. A concatenation operation is performed on the channel dimension, and fusion weights are assigned in conjunction with a channel attention modulation strategy to construct a fused multi-scale temperature gradient feature set with global vision and local recognition accuracy. Finally, the fused feature set is input into a global polar coordinate pooling layer, which uses a polar coordinate weighting mechanism to perform weighted compression on the entire spatial domain feature tensor, outputting a temperature field feature vector with the dimension (batch_size, 256, 1, 1, 1).

[0038] In a specific embodiment, the process of executing step S103 may specifically include the following steps: The temperature field eigenvector is converted into the temperature domain space to obtain the reconstructed temperature field data, and a temperature sensitivity weight function is constructed based on the reconstructed temperature field data and the phase change temperature range of the wind turbine flange forging. Based on the temperature sensitivity weight function, the proximity between the temperature value of each point in the reconstructed temperature field data and the critical temperature of the phase change is calculated to generate a sensitivity map, and the sensitivity map is enhanced to obtain an enhanced sensitivity map; Multi-scale temperature gradient calculation and fusion are performed on the reconstructed temperature field data through polar coordinate gradient operators with different kernel sizes in the multi-scale temperature gradient pyramid to obtain a comprehensive gradient feature map. The enhanced sensitivity map is multiplied pixel-wise with the comprehensive gradient feature map to obtain a phase change sensitive gradient map, and the phase change sensitive gradient map is fused with the enhanced sensitivity map to obtain a phase change sensitive enhanced feature vector.

[0039] Specifically, the temperature field eigenvectors are remapped back to the actual temperature domain space, restoring their physical meaning and constructing a feature representation that can directly perceive the critical state of the material. This mapping process is implemented using a fully connected mapping network. The temperature field eigenvectors are transformed into a spatial data tensor of the same dimension as the original polar coordinate temperature field through multiple neural layers, resulting in physically meaningful reconstructed temperature field data. To enable the model to focus on the phase transformation temperature range unique to the heat treatment of wind turbine flanges—that is, the key temperature points corresponding to the structural evolution stages of steel such as austenitization and ferrite-pearlite transformation, such as 590°C, 650°C, and 720°C—a set of temperature sensitivity weight functions centered on the phase transformation critical temperature is constructed. This function is modeled using a Gaussian distribution. By setting a temperature center point and a standard deviation width parameter σ, the closer the temperature of any spatial point is to the critical temperature, the higher its sensitivity value is, while the weight decreases rapidly as the temperature deviates from the center point. The numerical calculation result of the temperature sensitivity weight function falls between 0 and 1, forming a smooth response curve with a peak response at the critical temperature and a rapid decay in the noncritical region. This function is applied to each spatial location in the reconstructed temperature field data, calculating the proximity of its temperature value to the critical temperature of each phase transition point. This generates a sensitivity map, which, based on spatial points, indicates which regions within the entire heat treatment domain are most likely to be at the critical edge of a structural phase transition. While the basic sensitivity map is numerically responsive to the critical temperature, its numerical changes may still be suppressed by the smoothing effect of the temperature gradient, resulting in a less sensitive response to small fluctuations near the critical temperature. Therefore, further enhancement is performed. A nonlinear activation strategy is used to amplify the sensitivity map, using the hyperbolic tangent function as the primary nonlinear enhancement. By adjusting the function slope parameter, the response of sensitive regions is amplified, resulting in an enhanced sensitivity map. In this map, spatial points close to the critical temperature of the phase transition receive a significantly amplified response, resulting in higher numerical discrimination and enabling the model to more accurately identify small temperature fluctuations associated with phase transitions during subsequent processing. To enhance the model's ability to analyze temperature variations, a multi-scale temperature gradient pyramid structure is introduced to capture gradient information about temperature distribution changes at different scales. This pyramid model applies polar gradient operators with different kernel sizes, applying 3×3, 5×5, 7×7, and 9×9 convolution kernels to the reconstructed temperature field data for gradient calculation. Each polar gradient operator, based on its kernel center, estimates the directional gradient of the temperature variation in the local neighborhood while preserving angular periodicity and radial boundary symmetry. The gradient maps output by each layer reflect the temperature variation trends at different scales. Small-scale kernels can perceive local disturbances, while large-scale kernels can capture the global temperature difference distribution.Gradient maps from different scales are input to a feature fusion module. This module, through upsampling and channel-by-channel weighting, merges the multi-layer outputs into a comprehensive gradient feature map. This map spatially reflects both the microscopic and macroscopic thermal distributions and possesses high structural expressiveness. A pixel-by-pixel product operation is performed on the enhanced sensitivity map and the comprehensive gradient feature map to construct a phase transition sensitive gradient map. This operation essentially couples the spatial temperature change rate with the response to the critical temperature of the phase transition. This results in spatial points with strong gradient changes and within the critical temperature range receiving higher response values ​​in the map. Non-critical regions, even with temperature changes, are not highlighted, thus focusing the model's attention on locations with the greatest evolutionary risk. To integrate this highly physically sensitive response feature back into the main feature pathway, the phase transition sensitive gradient map and the enhanced sensitivity map are combined via a channel fusion mechanism. This fusion is achieved through channel-weighted superposition, preserving the global contour information of the critical response in the sensitivity map while incorporating directional information of the change trend in the gradient map to form the final output phase transition sensitive enhanced feature vector.

[0040] In a specific embodiment, the process of executing step S104 may specifically include the following steps: The phase change sensitive enhanced feature vector is input into the encoder of the latent space reconstruction algorithm for dimensionality reduction processing to obtain a low-dimensional latent space variable; Upsampling and reconstructing the low-dimensional latent space variables through a decoder in the latent space reconstruction algorithm to obtain first temperature field data; Applying the heat conduction partial differential equation constraint to perform iterative calculation of minimizing reconstruction error and physical constraint error on the first temperature field data to obtain second temperature field data; The second temperature field data is processed by space filling and sparse region completion to obtain full-field temperature distribution data.

[0041] Specifically, the phase-change-sensitive enhanced feature vector is input into the encoder portion of the latent space structure. The encoder employs a multi-layered polar convolutional neural network (PCNN) to progressively compress the spatial and channel dimensions of the input features through successive convolutions, normalizations, and nonlinear activations, while preserving their gradient characteristics, spatial periodicity, and phase-change sensitivity information. This process maps the high-dimensional tensor into a low-dimensional latent space variable. The dimensions of this latent space are set between 128 and 256. The compressed tensor contains the spatial layout of the temperature field, as well as key representations of the material's thermophysical properties, temperature gradient information, and phase-change region responses, thereby effectively transitioning from physical perception to structural abstraction. In the decoder portion, a set of polar transposed convolutional networks, symmetrical to the encoder, is used to upsample the low-dimensional latent space variables and reconstruct the temperature field. Transposed convolutions are used to gradually restore the compressed information to the spatial scale of the original temperature field. Instance normalization and the Prelude (PReLU) activation function are used to amplify spatial information while preserving gradient continuity and smooth temperature variations. This decoding process not only restores dimensionality but also maps and restores the physical state, as the features encoded in the latent space have been imbued with abstract rules for temperature evolution and material behavior by network training. The resulting first temperature field data formally restores the complete three-dimensional spatial structure and effectively predicts the overall trend of the actual temperature evolution process. However, because this prediction process relies primarily on feature reconstruction and is not constrained to adhere to the partial differential equation model of heat conduction, its details exhibit local fluctuations that violate physical laws, particularly near thermal diffusion boundaries, heat source interference points, or areas with sparse temperature measurements. To improve the physical consistency of the reconstructed data, the first temperature field data is used as the initial solution input into a physically constrained optimization module based on the partial differential equation of heat conduction. A joint loss function is constructed that comprehensively considers the reconstruction error term, the residual term of the physical partial derivative, and the latent space regularization term. The physical constraint component discretizes the governing heat conduction equation, calculating the first-order derivative of the temperature field with respect to time and the second-order derivative with respect to space (radial, angular, and depth). The residual term is then constructed by combining the material thermal diffusivity function and the internal heat source term. The thermal diffusion coefficient is modeled as a spatial variation, dynamically output via a perceptron network based on local temperature and material composition parameters. The internal heat source function is generated by inputting the historical temperature sequence and the current temperature state into a U-Net structure, capturing the non-uniform energy input during actual heat treatment. These partial derivatives are all implemented via finite difference approximations, enabling the model to directly numerically optimize the residuals of the partial differential equations within the training framework. During the optimization process, the model parameters are iteratively updated using gradient descent, targeting the total loss function. This ensures that the generated secondary temperature field data fits the original observations while satisfying the physical constraints of the heat conduction equation as closely as possible.The essence of this stage is to guide the model to find the temperature field evolution trajectory that is most consistent with the observation in the solution space of physical laws. Its output results show higher physical rationality in terms of local continuity, heat conduction propagation direction, and abnormal point diffusion path, thereby overcoming the limitations of simple data-driven models in non-uniform materials, complex boundaries, and transient dynamic responses. The second temperature field data is spatially filled and sparse regions are completed. The heat diffusion law and neighborhood relationship learned by the heat conduction constraint model in the latent space are comprehensively utilized, and the structure-preserving interpolation method is used to fill in the missing areas. The specific method includes propagating and extending the gradient direction of the adjacent known areas, and smoothly fitting the reconstructed boundaries through spatial consistency rules, so that the interpolation results are not only numerically continuous, but also conform to the propagation trend of the natural boundary in terms of heat diffusion behavior, and finally the full-field temperature distribution data is obtained.

[0042] In a specific embodiment, the step of applying the heat conduction partial differential equation constraint to perform iterative calculations to minimize the reconstruction error and the physical constraint error on the first temperature field data to obtain the second temperature field data may specifically include the following steps: Calculating a temperature time derivative and a space derivative based on the first temperature field data, and generating temperature field partial derivative data based on the temperature time derivative and the space derivative; The material composition vector and local temperature information of the wind turbine flange forging are input into the thermal diffusion coefficient network for processing to obtain the thermal diffusion coefficient in spatial variational form. Inputting the first temperature field data into the heat source function network for processing to obtain the internal heat source function; Apply the heat conduction partial differential equation constraint to construct a physical constraint loss function based on the temperature field partial derivative data, thermal diffusion coefficient and internal heat source function; The physical constraint loss function, the reconstruction loss function and the regularization term are combined into a total loss function, which is then iteratively optimized by minimizing the total loss function using the gradient descent method to obtain the second temperature field data.

[0043] Specifically, key partial derivatives are calculated based on the first temperature field data: the first-order derivative of temperature with respect to time and the second-order derivatives with respect to each spatial dimension. The time derivative reflects the transient characteristics of the temperature field at different moments. Its value is approximated using forward differencing: the difference between the temperature values ​​of adjacent time frames divided by the time interval yields the instantaneous rate of change. The spatial derivative is numerically approximated using central differencing. The second-order derivatives in the radial, angular, and thickness directions describe the spatial diffusion behavior of heat. These spatial derivatives characterize the direction of heat conduction, changes in heat flux density, and temperature gradient trends in a local area, thus forming the temperature field partial derivative data set. The microscopic composition of the wind turbine flange material is considered, as the thermal diffusion capacity of a material not only varies spatially but also significantly depends on its composition and temperature state. A vector reflecting the material's chemical composition (e.g., the mass percentages of elements such as C%, Si%, Mn%, Cr%, Mo%, and Ni) and the temperature values ​​at each spatial point in the first temperature field are input into a pre-trained thermal diffusivity network. The network, based on a multi-layer perceptron architecture, outputs a spatially variational thermal diffusion coefficient, i.e., a coefficient that dynamically changes with spatial coordinates. This effectively reflects the thermal diffusion heterogeneity in complex forgings caused by compositional inhomogeneities, varying heat treatment histories, and microstructural differences. This output not only improves the ability to fit the thermal diffusion field distribution but also provides a stronger explanation of material behavior at the physical level. After determining that the thermal diffusion coefficient characterizes the regulatory effect of internal energy input on thermal behavior, a heat source function is constructed, representing the amount of heat generated per unit volume per unit time. Because this heat source term is not only dependent on time but also on historical temperature states, current process parameters, and the thermal environment within the furnace, a deep neural network constructed with a U-Net architecture is used as the heat source function prediction model. The input consists of the first temperature field distribution at the current time point and the historical temperature series from several previous time steps. The output is a spatially distributed heat source intensity tensor. This model preserves spatial distribution information through a downsampling encoding and upsampling decoding structure, and extracts multi-scale features through skip connections. This ensures that the output heat source function is not only spatially and temporally consistent but also reflects non-steady-state heat source variations caused by local temperature differences, thermal radiation interference, or phase change exotherm. Based on the governing partial differential equation for heat conduction, a residual expression is constructed, which is used to define a physical constraint loss function. In actual modeling, the combination of the partial derivatives, thermal diffusion coefficient, and internal heat source term is considered the desired physical relationship. This is then compared with the actual derivative characteristics exhibited by the first temperature field. The difference represents the degree of deviation of the model's current output from the heat conduction law. This deviation is spatially integrated to form a loss function. This physical constraint loss function measures the deviation between the reconstructed temperature field and the actual heat diffusion process. A smaller value indicates that the current temperature evolution is more consistent with the physical governing laws; a larger value indicates that the model contains serious violations of thermodynamic equilibrium or unreasonable regions of the heat diffusion process.To ensure the model maintains high accuracy in fitting the observed data despite physical constraints, a data reconstruction loss is incorporated into the overall loss function. This loss measures the numerical deviation between the first temperature field and the actual observed value. A latent space regularization term is also introduced to regularize the latent space variables output by the encoder, preventing overdispersion in the latent variable distribution and improving the model's convergence stability and generalization capabilities. These three components together constitute the overall loss function. After constructing the complete loss function, a gradient descent-based optimization method is used to backpropagate all learnable parameters of the neural network. By continuously iteratively updating the network weights, the model gradually converges to an optimal state that balances physical laws and observed data with each iteration. In this process, the model corrects local perturbations in the first temperature field that violate the laws of heat diffusion and optimizes the thermal gradient distribution across the entire spatial domain. This results in the final output of the second temperature field data exhibiting high consistency and physical interpretability in terms of temporal evolution, spatial propagation, and boundary continuity.

[0044] In a specific embodiment, the process of executing step S105 may specifically include the following steps: The full-field temperature distribution data is characterized by macroscopic, mesoscopic, and microscopic scales. The global temperature standard deviation and radial temperature gradient are calculated at the macroscopic scale, the local temperature mean, standard deviation, and skewness are calculated at the mesoscopic scale, and the temperature change rate and temperature acceleration are calculated at the microscopic scale to obtain a multi-scale temperature feature set. The temperature profile is extracted along the depth direction of the full-field temperature distribution data, and the depth direction temperature gradient and depth gradient variance are calculated based on the temperature profile. The temperature standing wave phenomenon where the surface temperature is normal but there is a dangerous temperature gradient in the depth direction is identified, and the depth anomaly characteristics are obtained; Based on the multi-scale temperature feature set and depth anomaly features, the temperature field anomaly score is calculated for each spatial point; The areas where the temperature field anomaly score exceeds the anomaly warning threshold are segmented to obtain segmented areas, and the target feature vector of each segmented area is calculated. By comparing and analyzing the target feature vector with the historical defect database, the temperature field anomaly type probability distribution information including the anomaly type and probability distribution is generated.

[0045] Specifically, at the macroscopic scale, the global temperature standard deviation is calculated to measure the degree of thermal field dispersion of the entire forging, and the temperature gradient is calculated in the radial direction. By extracting the rate of change of the temperature difference from the inner ring to the outer ring, a global gradient image reflecting the main diffusion path of the heat flow and the steepness of the gradient is obtained. At the mesoscopic scale, a sliding window mechanism (30° angle, 100mm radial, 50mm thickness) is introduced to extract local statistical features by region, including the local temperature mean, standard deviation, and skewness value, to describe the heating intensity, fluctuation amplitude, and distribution symmetry of the local area. At the microscopic scale, first-order and second-order difference operations are performed on the temperature time series of each spatial point to calculate the temperature change rate and temperature acceleration to characterize the short-term dynamic behavior during the heat treatment process.

[0046] Deep-depth anomaly identification is performed based on typical structural risk characteristics of wind turbine flanges. Starting at each angle θ and radial position r, a temperature profile is extracted along the thickness direction from the surface to the inner layer. The depth-directed temperature gradient and its variance are calculated as indicators of depth gradient nonuniformity. When the surface temperature is stable but abrupt temperature gradient changes and violent fluctuations occur deep within the forging, this phenomenon is labeled "temperature standing wave," indicating potential defects within the forging, such as stress concentration, inadequate material microstructure transformation, or impeded thermal diffusion.

[0047] Based on a complete temperature feature set consisting of macroscopic, mesoscopic, microscopic, and deep anomaly features, a weighted comprehensive approach is used to calculate the temperature field anomaly score for each spatial point. This process considers factors such as global statistical fluctuations, local structural disturbances, dynamic anomaly responses, and deep temperature changes. Using a feature weighting model, each feature is assigned different weights, and a normalized anomaly score is output. Points with anomaly scores above the warning threshold are spatially clustered using a density clustering algorithm (such as DBSCAN) to obtain a set of continuous anomaly regions.

[0048] Each abnormal region is structured and coded, and a target feature vector is calculated, including multi-dimensional attributes such as average anomaly score, maximum anomaly value, spatial area, maximum temperature deviation, maximum temperature gradient, maximum depth gradient, anomaly distribution directionality, and temporal dynamic characteristics. This vector is input into an anomaly type recognition system built on a historical defect database. The system matches the historical defect templates through methods such as K-nearest neighbor matching, distance function fitting, or graph similarity calculation. The anomaly type classification result and probability distribution information are output based on the matching strength.

[0049] The above describes the intelligent monitoring method for forging production in the embodiment of the present invention. The following describes the intelligent monitoring system for forging production in the embodiment of the present invention. Figure 2 An embodiment of the intelligent monitoring system for forging production in the embodiment of the present invention includes: The acquisition module 201 is used to collect temperature data of the wind turbine flange forging and perform polar coordinate conversion to obtain polar coordinate temperature field data; Extraction module 202, for inputting polar coordinate temperature field data into an annular thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector; An enhancement module 203 is used to perform phase change critical temperature sensitivity enhancement on the temperature field eigenvector to obtain a phase change sensitivity enhancement eigenvector; The reconstruction module 204 is used to perform latent space reconstruction and interpolation on the phase change sensitive enhancement eigenvector based on the heat conduction partial differential equation constraint to obtain full-field temperature distribution data; The generation module 205 is used to generate temperature field anomaly scores and anomaly type probability distribution information based on the full-field temperature distribution data.

[0050] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0051] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling an intelligent monitoring device for forging production (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0052] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent monitoring method for forging production, characterized in that: include: Perform temperature data collection and polar coordinate conversion on wind turbine flange forgings to obtain polar coordinate temperature field data; Inputting the polar coordinate temperature field data into a circular thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector; Performing phase change critical temperature sensitivity enhancement on the temperature field characteristic vector to obtain a phase change sensitivity enhancement characteristic vector; Based on the constraints of the heat conduction partial differential equation, the phase change sensitive enhancement eigenvector is reconstructed and interpolated in the latent space to obtain the full-field temperature distribution data; A temperature field anomaly score and anomaly type probability distribution information are generated according to the full-field temperature distribution data.

2. The intelligent monitoring method for forging production according to claim 1, characterized in that: The temperature data of the wind turbine flange forging is collected and polar coordinate converted to obtain polar coordinate temperature field data, including: The wind turbine flange forging is radially divided into three regions including an outer ring, a middle ring, and an inner ring. Temperature measuring points are evenly arranged circumferentially in each region, and K-type thermocouples are installed at the temperature measuring points to obtain a multi-layer temperature measuring point structure. Performing real-time data acquisition on the K-type thermocouples in the multi-layer temperature measurement point structure to obtain raw temperature data, and performing noise elimination and smoothing processing on the raw temperature data to obtain filtered temperature data; Converting the filtered temperature data from a Cartesian coordinate system to a polar coordinate system to obtain a temperature field matrix represented by polar coordinates; Bicubic spline interpolation and standardization are performed on the temperature field matrix represented by polar coordinates to obtain polar coordinate temperature field data.

3. The intelligent monitoring method for forging production according to claim 1, characterized in that: The polar coordinate temperature field data is input into the annular thermal field gradient decomposition network for feature extraction to obtain the temperature field feature vector, including: Inputting the polar coordinate temperature field data into the input layer of the annular thermal field gradient decomposition network for feature encoding to obtain an input feature vector; The input feature vector is sequentially processed through four series-connected polar coordinate convolution modules in the annular thermal field gradient decomposition network to obtain a multi-scale feature map, wherein each polar coordinate convolution module includes two polar coordinate convolution layers, a batch normalization layer, and a LeakyReLU activation function; Inputting the multi-scale feature map into the annular self-attention module in the annular thermal field gradient decomposition network for processing to obtain an enhanced feature map; Fusing the enhanced feature map with the feature map of the corresponding layer in the multi-scale feature map to obtain a fused multi-scale temperature gradient feature; The fused multi-scale temperature gradient features are input into the global polar coordinate pooling layer in the annular thermal field gradient decomposition network for processing to obtain a temperature field feature vector.

4. The intelligent monitoring method for forging production according to claim 1, characterized in that: The step of performing phase change critical temperature sensitivity enhancement on the temperature field characteristic vector to obtain the phase change sensitivity enhancement characteristic vector includes: Converting the temperature field characteristic vector into a temperature domain space to obtain reconstructed temperature field data, and constructing a temperature sensitivity weight function based on the reconstructed temperature field data and the phase change temperature range of the wind power flange forging; Calculating the proximity between the temperature value of each point in the reconstructed temperature field data and the phase change critical temperature based on the temperature sensitivity weight function to generate a sensitivity map, and performing enhancement processing on the sensitivity map to obtain an enhanced sensitivity map; Multi-scale temperature gradient calculation and fusion are performed on the reconstructed temperature field data by using polar coordinate gradient operators with different kernel sizes in a multi-scale temperature gradient pyramid to obtain a comprehensive gradient feature map; A pixel-level product operation is performed on the enhanced sensitivity map and the comprehensive gradient feature map to obtain a phase change sensitive gradient map, and the phase change sensitive gradient map and the enhanced sensitivity map are fused to obtain a phase change sensitive enhanced feature vector.

5. The intelligent monitoring method for forging production according to claim 1, characterized in that: Based on the constraints of the heat conduction partial differential equation, the phase change sensitive enhancement feature vector is reconstructed and interpolated in the latent space to obtain the full-field temperature distribution data, including: Inputting the phase change sensitive enhanced feature vector into an encoder in a latent space reconstruction algorithm for dimensionality reduction processing to obtain a low-dimensional latent space variable; Upsampling and reconstructing the low-dimensional latent space variables by a decoder in the latent space reconstruction algorithm to obtain first temperature field data; Applying a heat conduction partial differential equation constraint to perform an iterative calculation of minimizing a reconstruction error and a physical constraint error on the first temperature field data to obtain second temperature field data; Perform space filling and sparse region completion processing on the second temperature field data to obtain full-field temperature distribution data.

6. The intelligent monitoring method for forging production according to claim 5, characterized in that: The applying the heat conduction partial differential equation constraint to perform iterative calculation of minimizing reconstruction error and physical constraint error on the first temperature field data to obtain second temperature field data includes: Calculating a temperature time derivative and a spatial derivative based on the first temperature field data, and generating temperature field partial derivative data based on the temperature time derivative and the spatial derivative; The material composition vector and local temperature information of the wind turbine flange forging are input into the thermal diffusion coefficient network for processing to obtain the thermal diffusion coefficient in spatial variational form. Inputting the first temperature field data into a heat source function network for processing to obtain an internal heat source function; Applying heat conduction partial differential equation constraints to construct a physical constraint loss function according to the temperature field partial derivative data, the thermal diffusion coefficient and the internal heat source function; The physical constraint loss function, the reconstruction loss function and the regularization term are combined into a total loss function, and the total loss function is minimized and iteratively optimized using a gradient descent method to obtain second temperature field data.

7. The intelligent monitoring method for forging production according to claim 1, characterized in that: Generating the temperature field anomaly score and anomaly type probability distribution information according to the full-field temperature distribution data includes: Performing feature calculations on the full-field temperature distribution data at three scales, namely, macroscopic, mesoscopic, and microscopic, calculating the global temperature standard deviation and radial temperature gradient at the macroscopic scale, calculating the local temperature mean, standard deviation, and skewness at the mesoscopic scale, and calculating the temperature change rate and temperature acceleration at the microscopic scale, to obtain a multi-scale temperature feature set; Extracting a temperature profile along the depth direction of the full-field temperature distribution data, and calculating the depth direction temperature gradient and depth gradient variance based on the temperature profile, identifying the temperature standing wave phenomenon where the surface temperature is normal but a dangerous temperature gradient exists in the depth direction, and obtaining the depth anomaly feature; Calculating a temperature field anomaly score for each spatial point based on the multi-scale temperature feature set and the depth anomaly feature; The area where the temperature field anomaly score exceeds the anomaly warning threshold is segmented to obtain segmented areas, and the target feature vector of each segmented area is calculated. By comparing and analyzing the target feature vector with the historical defect database, temperature field anomaly type probability distribution information including anomaly type and probability distribution is generated.

8. An intelligent monitoring system for forging production, characterized in that: A method for implementing the intelligent monitoring method for forging production according to any one of claims 1 to 7, wherein the intelligent monitoring system for forging production comprises: The acquisition module is used to collect temperature data of wind power flange forgings and perform polar coordinate conversion to obtain polar coordinate temperature field data; An extraction module is used to input the polar coordinate temperature field data into a ring thermal field gradient decomposition network for feature extraction to obtain a temperature field feature vector; An enhancement module, configured to perform phase change critical temperature sensitivity enhancement on the temperature field characteristic vector to obtain a phase change sensitivity enhancement characteristic vector; A reconstruction module is used to perform latent space reconstruction and interpolation on the phase change sensitive enhancement eigenvector based on the constraints of the heat conduction partial differential equation to obtain full-field temperature distribution data; A generation module is used to generate temperature field anomaly scores and anomaly type probability distribution information based on the full-field temperature distribution data.

Images