A Method for Detecting the Uniformity of Heat Treatment in Pump Bodies Based on Infrared Imaging

Abstract

This invention discloses a method for detecting the uniformity of heat treatment in a pump body based on infrared imaging, comprising the following steps: Step 1: Obtaining a three-dimensional structural model of the pump body and calibrating the detection points; Step 2: Applying thermal perturbation to each detection area; Step 3: Acquiring dynamic multi-channel temperature images to obtain a temperature-time series; Step 4: Constructing a non-Fourier thermoelastic inertial response model for each detection point and extracting thermal response feature parameters; Step 5: Constructing a thermal response memory map; Step 6: Inputting the thermal response memory map into an improved TimeSFormer model, which incorporates a thermal potential field tensor guidance mechanism, and combining the spatial location information of the detection points with the structural topology encoding vector to obtain a heat treatment uniformity score; Step 7: Identifying concentrated areas of heat treatment deviation and outputting a detection report. This invention combines infrared imaging with an improved TimeSFormer model to realize an intelligent detection and evaluation method for the uniformity of heat treatment in a pump body.

Description

Technical Field

[0001] This invention relates to the field of heat treatment quality assessment technology, and in particular to a method for detecting the uniformity of heat treatment in pump bodies based on infrared imaging. Background Technology

[0002] With the increasing demands for consistent heat treatment quality in large and complex components, accurately monitoring and identifying deviations in the temperature response behavior of structurally complex pump bodies during heat treatment has become a key research focus in industrial quality control. Currently, commonly used heat treatment quality assessment methods mainly rely on thermocouple point measurements, single-channel infrared imaging, or post-processing statistical analysis methods. However, these methods generally suffer from the following problems in practical applications:

[0003] Existing thermocouple methods have limited sampling points, making it difficult to cover the entire pump surface. Single-channel infrared imaging systems are susceptible to surface emissivity differences and optical interference, resulting in large errors in temperature measurement results and failing to reflect the true heat diffusion process. Most methods only assess steady-state temperature or maximum temperature rise, neglecting dynamic response characteristics during thermal disturbances, such as response delay, thermal hysteresis, and local thermoelastic echo behavior. Furthermore, the lack of spatial analysis methods that combine with three-dimensional structural information makes it difficult to identify uneven heat treatment within complex structural regions. In addition, traditional clustering or classification methods are insufficient in modeling time series and spatial structural information, failing to effectively extract spatiotemporal correlation features and perform inter-regional comparative analysis, thus limiting the accurate location and interpretation of areas with concentrated heat treatment deviations.

[0004] Therefore, how to provide a method for detecting the uniformity of pump body heat treatment based on infrared imaging is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for detecting the uniformity of heat treatment in pump bodies based on infrared imaging. This invention combines infrared multi-band thermal imaging acquisition with an improved TimeSFormer model to construct a thermal response memory spectrum and introduce a thermal potential field tensor guidance mechanism. Combined with topological coding of the detection point structure, it improves the modeling accuracy of the physical nature of the heat diffusion process and the spatial resolution capability for evaluating the uniformity of regional heat treatment. It has the advantages of strong analytical robustness, strong adaptability to complex geometric structures, and intuitive and interpretable detection results.

[0006] A method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Obtain the three-dimensional structural model of the pump body, divide the surface of the pump body into several detection areas according to the structural features, and calibrate the detection points;

[0008] Step 2: Apply thermal disturbance to each detection area by heating with laser pulses;

[0009] Step 3: Use shortwave, medium wave and long wave infrared thermal imagers to dynamically acquire multi-channel temperature images of the pump body surface after thermal disturbance, and obtain the temperature-time series corresponding to each detection point;

[0010] Step 4: Based on the temperature-time series, construct a non-Fourier thermoelastic inertial response model for each detection point and extract thermal response feature parameters;

[0011] Step 5: Construct a temperature-response hysteresis matrix from the thermal response characteristic parameters of each detection point to form a thermal response memory map;

[0012] Step 6: Input the thermal response memory map into the improved TimeSFormer model. The improved TimeSFormer model introduces a thermal potential field tensor guidance mechanism, and combines the spatial location information of the detection points with the structural topology encoding vector to obtain the thermal processing uniformity score of each detection area.

[0013] Step 7: Based on the heat treatment uniformity score and pump body structure information, perform cluster analysis of the detection area to identify areas with concentrated heat treatment deviations and output a detection report.

[0014] Optionally, step one specifically includes:

[0015] A three-dimensional structural model of the pump body is obtained. The three-dimensional structural model of the pump body is a digital model established by an industrial-grade three-dimensional laser scanning device, which has geometric accuracy corresponding to the actual pump body structure.

[0016] The surface of the pump body is divided into regions according to structural features, including ribs, cavities, flanges, flow channels, and abrupt wall thickness boundaries, to generate several detection areas.

[0017] At least one detection point is marked in each detection area. The detection point is set at the geometric center or edge corner of the detection area and corresponds to a unique spatial coordinate identifier in the three-dimensional structural model of the pump body.

[0018] Optionally, step two specifically involves:

[0019] Within each detection area, laser pulse heating is performed using a near-infrared laser with a wavelength of 980 nm to 1550 nm, wherein the pulse width of the near-infrared laser is 100 ms to 500 ms and the pulse frequency is 5 Hz to 20 Hz.

[0020] After being focused by a collimating lens, the laser acts on the surface of the pump body, forming thermal disturbance spots with a diameter of 1 mm to 5 mm.

[0021] The near-infrared laser controls its position using a two-dimensional displacement platform and a scanning galvanometer to locate and heat designated detection points within each detection area.

[0022] Optionally, step three specifically includes:

[0023] A multi-channel infrared imaging system composed of short-wave, mid-wave, and long-wave infrared thermal imagers was used to synchronously acquire temperature images of the pump body surface after thermal disturbance. The working band of the short-wave infrared thermal imager is 0.9 micrometers to 1.7 micrometers, the working band of the mid-wave infrared thermal imager is 3 micrometers to 5 micrometers, and the working band of the long-wave infrared thermal imager is 8 micrometers to 14 micrometers.

[0024] The infrared thermal imager has an image acquisition frame rate of 60 frames per second to 200 frames per second, a thermal sensitivity greater than 40 milliklvin, and a spatial resolution of not less than 640×512 pixels.

[0025] Each infrared channel records thermal images of the same detection area, and a multi-scale gradient orientation histogram algorithm is used for image registration. The temperature-time series of each detection point after heating excitation is extracted from the registered thermal image sequence.

[0026] Optionally, the multi-scale gradient orientation histogram algorithm specifically includes:

[0027] Short-wave, mid-wave, and long-wave infrared thermal images are subjected to grayscale normalization to form a uniform brightness scale;

[0028] A multi-scale image set is constructed based on an image pyramid. At each scale, a sliding window of a set size is used to extract image patches, and gradient orientation histogram features are calculated for the image patches. The gradient is obtained by the Sobel operator, and the orientation histogram is discretely encoded using a preset angle interval.

[0029] Long-wave infrared thermal images are selected as reference images. Cross-band image block matching is performed in short-wave infrared images and mid-wave infrared thermal images using the orientation histogram features. The position of the corresponding image block in the image to be matched is determined by minimizing the L2 norm of the pixel difference between image blocks.

[0030] The image to be matched is subjected to pixel-level position correction, and B-spline interpolation is used to complete image resampling to obtain a registered image that is consistent with the reference image in spatial position, thereby generating a registered thermal image sequence.

[0031] Optionally, step four specifically involves:

[0032] The temperature-time series of the detection points is preprocessed. The preprocessing includes using a sliding mean filter to reduce high-frequency noise, using a signal baseline drift correction method to unify and normalize the initial temperature reference, and aligning the starting points of each sequence according to the laser excitation time.

[0033] During the excitation phase, based on the non-Fourier heat conduction theory, a two-phase hysteresis model is used to fit the temperature change process in the initial stage of excitation. The two-phase hysteresis model introduces heat flow hysteresis time and temperature gradient hysteresis time to describe the relative hysteresis behavior between heat flow and temperature response.

[0034] During the peak response stage, a thermal stress-driven model caused by the temperature gradient is established based on the thermoelastic properties of the material. The thermal stress-driven model couples the instantaneous temperature gradient field with the local thermal expansion coefficient and calculates the stress distribution caused by the non-uniform thermal field through finite element discretization, fitting the main peak time and amplitude in the temperature response.

[0035] During the decay phase, a recovery model expressing the response hysteresis is constructed using a time deconvolution kernel function to clearly distinguish the nonlinear process of thermal diffusivity changing with time.

[0036] The biphase hysteresis model, thermal stress-driven model, and recovery model are continuously spliced ​​together to form a complete non-Fourier thermoelastic inertial response model.

[0037] Thermal response feature parameters are extracted based on the non-Fourier thermoelastic inertial response model. These thermal response feature parameters include:

[0038] Temperature change slope: the average rate of change of the temperature rise segment after the excitation start time;

[0039] Response delay time: The time interval from the start of excitation to the occurrence of the maximum slope change in the temperature curve;

[0040] Thermal hysteresis characteristic: the delay difference between peak response time and excitation termination time;

[0041] Thermoelastic echo morphology: the amplitude and symmetrical distribution characteristics of the rebound temperature fluctuations that occur during the decay phase.

[0042] Optionally, step five specifically includes:

[0043] The thermal response feature parameters extracted from each detection point are arranged according to a preset time node to construct a corresponding temperature-response hysteresis matrix. The rows of the temperature-response hysteresis matrix represent the spatial location of different detection points, the columns represent the time evolution dimension after thermal disturbance, and the matrix elements are the values ​​of thermal response feature parameters at the corresponding time.

[0044] The constructed temperature-response hysteresis matrix is ​​encoded and mapped in two dimensions to generate a thermal response memory map.

[0045] Optionally, step six specifically includes:

[0046] The thermal response memory map is input into the improved TimeSFormer model, which adopts a Transformer architecture with a time-space separation structure, including a spatial encoding module, a temporal encoding module, and a region aggregation module.

[0047] The spatial coding module introduces a thermal potential field tensor guidance mechanism, which specifically includes:

[0048] A three-dimensional thermal potential field tensor is constructed, wherein the three components of the three-dimensional thermal potential field tensor are the derivative of the rate of temperature change with respect to time, the normalized numerical distribution of the local heat capacity of the material, and the thermal hysteresis distance calculated based on the thermal response delay between detection points.

[0049] The three-dimensional thermal potential field tensor is transformed into a guiding vector field by a convolutional encoder during the input stage, and then weighted and fused with the linear transformation results of Query and Key in the spatial encoding module to construct the guiding perception attention weight in the spatial attention mechanism.

[0050] Based on the guided perception attention weights, spatial features are extracted from each frame of the thermal response memory map to generate spatial coding features;

[0051] The temporal encoding module is constructed in the form of a multi-layer Transformer encoder stack. It models the spatial encoding features in the temporal dimension and introduces the structural topology encoding vector of the detection point at each time step. The structural topology encoding vector is obtained by mapping the spatial coordinates of the detection point in the three-dimensional structural model of the pump body through an MLP embedding network. The MLP embedding network includes two hidden layers with ReLU activation functions and an output layer, which are used to encode the three-dimensional spatial coordinates into an embedding representation of a preset dimension and fuse them with the spatial encoding features of the corresponding time step to obtain the temporal encoding features.

[0052] The region aggregation module aggregates the time-encoding features of multiple detection points in each detection region and generates a region-level thermal response representation vector by weighted averaging.

[0053] The region-level thermal response representation vector is input into a fully connected layer, which uses a Sigmoid activation function to output a thermal uniformity score between 0 and 1.

[0054] Optionally, step seven specifically includes:

[0055] The heat treatment uniformity score of each detection area is combined with the geometric center coordinates in the three-dimensional structural model of the pump body to construct a feature vector set containing heat treatment uniformity score and spatial location information.

[0056] A density-based spatial clustering algorithm was used to analyze the feature vector set. The density threshold condition was set to ensure that the difference in heat treatment uniformity score did not exceed 0.05 and the spatial distance did not exceed 10 mm.

[0057] The detection regions that meet the density threshold condition are divided into the same cluster;

[0058] In all clusters, the average heat treatment uniformity score is statistically analyzed, and clusters whose average heat treatment uniformity score is more than 30% lower than the average heat treatment uniformity score of the entire detection area are identified as concentrated areas of heat treatment deviation.

[0059] Output a test report, which includes the spatial distribution of each cluster, statistical results of heat treatment uniformity scores, and location information of concentrated areas of heat treatment deviation.

[0060] The beneficial effects of this invention are:

[0061] This invention addresses the problems of low temperature response coverage, lack of dynamic features, and weak thermal deviation identification capabilities during the heat treatment of complex pump structures by synergistically integrating multi-band infrared imaging and non-Fourier thermoelastic inertial response modeling. It employs near-infrared laser pulses to apply short-duration thermal perturbations to multiple detection points, combined with short-wave, mid-wave, and long-wave infrared thermal imagers to construct a dynamic multi-channel temperature image sequence. Through image registration and temperature time-series extraction, it achieves comprehensive acquisition of high spatial resolution and multi-dimensional temperature response. In the response modeling stage, a non-Fourier thermoelastic inertial response model comprising three stages—excitation, peak response, and decay—is constructed, integrating a two-phase... The system employs a mechanism that leverages delayed heat conduction, temperature gradient-driven stress response, and thermal diffusion deconvolution to precisely extract thermal response characteristic parameters. In the uniformity assessment phase, a thermal response memory map is constructed and input into an improved TimeSFormer model guided by a thermal potential field tensor. This model, through the joint guidance of spatial attention distribution and temporal modeling using a three-dimensional thermal potential field tensor and topological encoding of the detection point structure, significantly improves the physical relevance of thermal diffusion behavior modeling and the accuracy of regional scoring. Finally, density spatial clustering is used to identify spatial deviations in the scoring feature vector set, accurately locating concentrated areas of heat treatment deviations and achieving intelligent detection and output of pump body heat treatment quality. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is an overall flowchart of a pump body heat treatment uniformity detection method based on infrared imaging proposed in this invention.

[0064] Figure 2 This is a flowchart illustrating the construction of a non-Fourier thermoelastic inertial response model and the extraction of thermal response feature parameters for a pump body heat treatment uniformity detection method based on infrared imaging proposed in this invention.

[0065] Figure 3 This is a schematic diagram of the improved TimeSFormer model structure, which introduces a thermal potential field tensor guiding mechanism, in the pump body heat treatment uniformity detection method based on infrared imaging proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figures 1-3 A method for detecting the uniformity of heat treatment in a pump body based on infrared imaging includes the following steps:

[0068] Step 1: Obtain the three-dimensional structural model of the pump body, divide the surface of the pump body into several detection areas according to the structural features, and calibrate the detection points;

[0069] Step 2: Apply thermal disturbance to each detection area by heating with laser pulses;

[0070] Step 3: Use shortwave, medium wave and long wave infrared thermal imagers to dynamically acquire multi-channel temperature images of the pump body surface after thermal disturbance, and obtain the temperature-time series corresponding to each detection point;

[0071] Step 4: Based on the temperature-time series, construct a non-Fourier thermoelastic inertial response model for each detection point and extract thermal response feature parameters;

[0072] Step 5: Construct a temperature-response hysteresis matrix from the thermal response characteristic parameters of each detection point to form a thermal response memory map;

[0073] Step 6: Input the thermal response memory map into the improved TimeSFormer model. The improved TimeSFormer model introduces a thermal potential field tensor guidance mechanism, and combines the spatial location information of the detection points with the structural topology encoding vector to obtain the thermal processing uniformity score of each detection area.

[0074] Step 7: Based on the heat treatment uniformity score and pump body structure information, perform cluster analysis of the detection area to identify areas with concentrated heat treatment deviations and output a detection report.

[0075] In this embodiment, step one specifically includes:

[0076] A three-dimensional structural model of the pump body is obtained. The three-dimensional structural model of the pump body is a digital model established by an industrial-grade three-dimensional laser scanning device, which has geometric accuracy corresponding to the actual pump body structure.

[0077] The surface of the pump body is divided into regions according to structural features, including ribs, cavities, flanges, flow channels, and abrupt wall thickness boundaries, to generate several detection areas.

[0078] At least one detection point is marked in each detection area. The detection point is set at the geometric center or edge corner of the detection area and corresponds to a unique spatial coordinate identifier in the three-dimensional structural model of the pump body.

[0079] In this embodiment, step two specifically includes:

[0080] Within each detection area, laser pulse heating is performed using a near-infrared laser with a wavelength of 980 nm to 1550 nm, wherein the pulse width of the near-infrared laser is 100 ms to 500 ms and the pulse frequency is 5 Hz to 20 Hz.

[0081] After being focused by a collimating lens, the laser acts on the surface of the pump body, forming thermal disturbance spots with a diameter of 1 mm to 5 mm.

[0082] The near-infrared laser controls its position using a two-dimensional displacement platform and a scanning galvanometer to locate and heat designated detection points within each detection area.

[0083] In this embodiment, step three specifically includes:

[0084] A multi-channel infrared imaging system composed of short-wave, mid-wave, and long-wave infrared thermal imagers was used to synchronously acquire temperature images of the pump body surface after thermal disturbance. The working band of the short-wave infrared thermal imager is 0.9 micrometers to 1.7 micrometers, the working band of the mid-wave infrared thermal imager is 3 micrometers to 5 micrometers, and the working band of the long-wave infrared thermal imager is 8 micrometers to 14 micrometers.

[0085] The infrared thermal imager has an image acquisition frame rate of 60 frames per second to 200 frames per second, a thermal sensitivity greater than 40 milliklvin, and a spatial resolution of not less than 640×512 pixels.

[0086] Each infrared channel records thermal images of the same detection area, and a multi-scale gradient orientation histogram algorithm is used for image registration. The temperature-time series of each detection point after heating excitation is extracted from the registered thermal image sequence.

[0087] In this embodiment, the multi-scale gradient orientation histogram algorithm specifically includes:

[0088] Short-wave, mid-wave, and long-wave infrared thermal images are subjected to grayscale normalization to form a uniform brightness scale;

[0089] A multi-scale image set is constructed based on an image pyramid. At each scale, a sliding window of a set size is used to extract image patches, and gradient orientation histogram features are calculated for the image patches. The gradient is obtained by the Sobel operator, and the orientation histogram is discretely encoded using a preset angle interval.

[0090] Long-wave infrared thermal images are selected as reference images. Cross-band image block matching is performed in short-wave infrared images and mid-wave infrared thermal images using the orientation histogram features. The position of the corresponding image block in the image to be matched is determined by minimizing the L2 norm of the pixel difference between image blocks.

[0091] Define the matching cost function as follows: ;

[0092] in, express and The L2 norm of the pixel difference between the two is used to measure similarity; the smaller the value, the higher the degree of matching. Indicates the width of the image block; Indicates the height of the image block; This represents an image patch of the reference image, with a size of [size missing]. The coordinates are The grayscale normalized value at that location; Indicates that the image to be matched contains the following: The image patch with the top-left corner coordinates, relative coordinates The grayscale normalized value at that location;

[0093] Within the search area of ​​the sliding window, select... The smallest top-left corner coordinate is used to determine the position of the corresponding image patch in the image to be matched.

[0094] The image to be matched is subjected to pixel-level position correction, and B-spline interpolation is used to complete image resampling to obtain a registered image that is consistent with the reference image in spatial position, thereby generating a registered thermal image sequence.

[0095] The introduction of a multi-scale gradient orientation histogram algorithm in this invention significantly improves the accuracy and robustness of multi-channel infrared image registration. Traditional image registration methods, such as those based on grayscale or template matching, struggle to achieve high-precision alignment across different bands due to inherent differences in imaging mechanisms, spatial mismatch, and response inconsistencies among short-wave, mid-wave, and long-wave infrared images. However, the multi-scale gradient orientation histogram algorithm employed in this invention effectively captures the structural edge responses of thermally disturbed regions in different bands by extracting local gradient orientation distribution features from a multi-scale image pyramid, achieving stable matching across modal images. Furthermore, by combining the L2 norm matching criterion with B-spline interpolation correction, the accumulation of registration errors in nonlinear structural regions by affine transformation is avoided. This algorithm integration not only improves the registration accuracy of thermal image data but also enhances the reliability of subsequent thermal response feature extraction, thereby improving the accuracy of pump body heat treatment uniformity evaluation. Compared to existing technologies, this invention extends gradient orientation histograms to multi-scale and multi-band domains, realizing the engineering feasibility of cross-band thermal image registration in industrial-grade infrared detection.

[0096] In this embodiment, step four specifically includes:

[0097] The temperature-time series of the detection points is preprocessed. The preprocessing includes using a sliding mean filter to reduce high-frequency noise, using a signal baseline drift correction method to unify and normalize the initial temperature reference, and aligning the starting points of each sequence according to the laser excitation time.

[0098] During the excitation phase, based on the non-Fourier heat conduction theory, a two-phase hysteresis model is used to fit the temperature change process in the initial stage of excitation. The two-phase hysteresis model introduces heat flow hysteresis time and temperature gradient hysteresis time to describe the relative hysteresis behavior between heat flow and temperature response.

[0099] During the peak response stage, a thermal stress-driven model caused by the temperature gradient is established based on the thermoelastic properties of the material. The thermal stress-driven model couples the instantaneous temperature gradient field with the local thermal expansion coefficient and calculates the stress distribution caused by the non-uniform thermal field through finite element discretization, fitting the main peak time and amplitude in the temperature response.

[0100] During the decay phase, a recovery model expressing the response hysteresis is constructed using a time deconvolution kernel function to clearly distinguish the nonlinear process of thermal diffusivity changing with time.

[0101] The time-domain deconvolution kernel function is a time-domain function that describes the dynamic response characteristics of a thermal system. It reflects how the temperature response changes over time after the system receives a unit thermal excitation at a certain moment. In the process of heat conduction inversion, the temperature response received by the system can be regarded as the point-by-point superposition effect of the historical excitation signal and the kernel function over time. The purpose of deconvolution is to deduce the original thermal excitation signal distribution or identify the heat conduction characteristics of the system by combining the known temperature response curve with the time-domain deconvolution kernel function.

[0102] In practical applications, the temporal deconvolution kernel function is used to separate the causal relationship between the thermal disturbance signal and the actual response, thereby more accurately restoring the true temporal characteristics of the thermal input. This function typically exhibits certain hysteresis and diffusion characteristics, and can represent the non-instantaneous response intensity of the system to thermal excitation at different time points, making it an important component in non-Fourier heat conduction modeling.

[0103] The biphase hysteresis model, thermal stress-driven model, and recovery model are continuously spliced ​​together to form a complete non-Fourier thermoelastic inertial response model.

[0104] Thermal response feature parameters are extracted based on the non-Fourier thermoelastic inertial response model. These thermal response feature parameters include:

[0105] Temperature change slope: the average rate of change of the temperature rise segment after the excitation start time;

[0106] Response delay time: The time interval from the start of excitation to the occurrence of the maximum slope change in the temperature curve;

[0107] Thermal hysteresis characteristic: the delay difference between peak response time and excitation termination time;

[0108] Thermoelastic echo morphology: the amplitude and symmetrical distribution characteristics of the rebound temperature fluctuations that occur during the decay phase.

[0109] In this embodiment, step five specifically includes:

[0110] The thermal response feature parameters extracted from each detection point are arranged according to a preset time node to construct a corresponding temperature-response hysteresis matrix. The rows of the temperature-response hysteresis matrix represent the spatial location of different detection points, the columns represent the time evolution dimension after thermal disturbance, and the matrix elements are the values ​​of thermal response feature parameters at the corresponding time.

[0111] The constructed temperature-response hysteresis matrix is ​​encoded and mapped in two dimensions to generate a thermal response memory map.

[0112] In this embodiment, step six specifically includes:

[0113] The thermal response memory map is input into the improved TimeSFormer model, which adopts a Transformer architecture with a time-space separation structure, including a spatial encoding module, a temporal encoding module, and a region aggregation module.

[0114] The spatial coding module introduces a thermal potential field tensor guidance mechanism, which specifically includes:

[0115] A three-dimensional thermal potential field tensor is constructed, wherein the three components of the three-dimensional thermal potential field tensor are the derivative of the rate of temperature change with respect to time, the normalized numerical distribution of the local heat capacity of the material, and the thermal hysteresis distance calculated based on the thermal response delay between detection points.

[0116] In the construction of the thermal potential tensor, the normalized numerical distribution of the local heat capacity of the material is used to characterize the differences in heat storage and transfer capabilities of different structural parts. Specifically, based on the heat capacity (heat that can be stored per unit mass) parameter of the material at each detection point in the three-dimensional structural model of the pump body, combined with its volume and density information, the equivalent heat capacity value of each detection point is calculated. Due to the variations in wall thickness and material type in different detection areas, there are significant differences in the distribution of heat capacity values. Therefore, the heat capacity values ​​of all detection points need to be normalized using the Z-score method, ultimately forming a normalized numerical distribution of local heat capacity covering the entire detection point area of ​​the pump body, which is used as one-dimensional channel in the thermal potential tensor.

[0117] On the other hand, the thermal hysteresis distance calculated based on the thermal response delay between detection points is used to reflect the time delay characteristics of thermal disturbance propagation within the structure. The specific steps are as follows: First, the time difference between the inflection points of temperature change at each pair of detection points under thermal disturbance is calculated. Based on this time difference, combined with the spatial distance between the two points in the three-dimensional structural model, a spatiotemporal joint distribution is constructed to calculate the thermal response propagation delay per unit distance, i.e., the thermal hysteresis distance. This distance is embedded into the spatial attention weight generation of the model as an important component of the thermal potential field tensor, thereby achieving deep modeling of the physical mechanism of structural thermal diffusion.

[0118] The three-dimensional thermal potential field tensor is transformed into a guiding vector field by a convolutional encoder during the input stage, and then weighted and fused with the linear transformation results of Query and Key in the spatial encoding module to construct the guiding perception attention weight in the spatial attention mechanism.

[0119] Based on the guided perception attention weights, spatial features are extracted from each frame of the thermal response memory map to generate spatial coding features;

[0120] The temporal encoding module is constructed in the form of a multi-layer Transformer encoder stack. It models the spatial encoding features in the temporal dimension and introduces the structural topology encoding vector of the detection point at each time step. The structural topology encoding vector is obtained by mapping the spatial coordinates of the detection point in the three-dimensional structural model of the pump body through an MLP embedding network. The MLP embedding network includes two hidden layers with ReLU activation functions and an output layer, which are used to encode the three-dimensional spatial coordinates into an embedding representation of a preset dimension and fuse them with the spatial encoding features of the corresponding time step to obtain the temporal encoding features.

[0121] The region aggregation module aggregates the time-encoding features of multiple detection points in each detection region and generates a region-level thermal response representation vector by weighted averaging.

[0122] The region-level thermal response representation vector is input into a fully connected layer, which uses a Sigmoid activation function to output a thermal uniformity score between 0 and 1.

[0123] Compared to the existing standard TimeSFormer architecture, the improved TimeSFormer model introduces a thermal potential tensor guidance mechanism in the spatial encoding module, enabling physically-aware modeling of non-uniform thermal diffusion behavior in the pump structure. Specifically, this improvement incorporates a three-dimensional thermal potential tensor, integrating the derivative of the rate of temperature change, the local heat capacity distribution of the material, and the thermal hysteresis distance—all closely related to thermal diffusion dynamics—into a spatial attention mechanism, constructing an attention weight generation process with thermophysical prior guidance capabilities. Compared to traditional Transformers that construct attention solely based on the content similarity of the Query and Key, this invention guides the model to focus on physically sensitive regions along the thermal diffusion path through the fusion of physical field information, thereby improving the matching degree and recognition ability of feature representations to thermal conduction behavior.

[0124] In addition, this improved model also introduces structural topology encoding vectors in the time encoding module, which utilizes the relative positional relationship of detection points in the three-dimensional space of the pump body to assist in modeling the temporal correlation of thermal response features and enhance the model's generalization ability in long-term thermal response pattern recognition.

[0125] Through the above improvements, the model not only has the ability to extract spatiotemporal features based on data, but also integrates prior knowledge of pump structure and thermal diffusion physical mechanisms, which significantly improves the accuracy and interpretability of non-uniformity identification in complex heat treatment processes. It has high engineering adaptability and model versatility, and is especially suitable for situations in industrial equipment where local thermal response is difficult to accurately model with pure data models.

[0126] In this embodiment, step seven specifically includes:

[0127] The heat treatment uniformity score of each detection area is combined with the geometric center coordinates in the three-dimensional structural model of the pump body to construct a feature vector set containing heat treatment uniformity score and spatial location information.

[0128] A density-based spatial clustering algorithm was used to analyze the feature vector set. The density threshold condition was set to ensure that the difference in heat treatment uniformity score did not exceed 0.05 and the spatial distance did not exceed 10 mm.

[0129] The detection regions that meet the density threshold condition are divided into the same cluster;

[0130] In all clusters, the average heat treatment uniformity score is statistically analyzed, and clusters whose average heat treatment uniformity score is more than 30% lower than the average heat treatment uniformity score of the entire detection area are identified as concentrated areas of heat treatment deviation.

[0131] Output a test report, which includes the spatial distribution of each cluster, statistical results of heat treatment uniformity scores, and location information of concentrated areas of heat treatment deviation.

[0132] Example 1

[0133] To verify the feasibility of this invention in practice, it was applied to a heat treatment quality inspection scenario in a pump body casting production workshop. This workshop mainly produces high-pressure cast steel pump bodies with complex structures, including multiple internal cavities, stiffeners, thick-walled sections, and slender flow channels. During heat treatment, problems such as uneven temperature zones, thermal stress concentration, and incomplete annealing are prone to occur.

[0134] First, a digital structural model of the pump body was established using a high-precision laser three-dimensional scanning system. The surface of the pump body was divided into 52 detection areas by referring to the geometric boundaries and heat-sensitive areas (cavity corners, reinforcing rib roots, and flow channel outlets). Several detection points were marked in each area, for a total of 317 calibrated detection points.

[0135] Subsequently, after the heat treatment, each detection area was heated by laser pulses (wavelength 1064nm, pulse width 300ms, frequency 10Hz), and a complete dynamic thermal response spectrum was acquired using a multi-channel synchronous imaging system composed of short-wave, mid-wave, and long-wave infrared thermal imagers. The data acquisition frequency was set to 150 frames per second, covering the entire process from the start of excitation to temperature decay stabilization, lasting 15 seconds.

[0136] After preprocessing, the acquired temperature-time series was modeled based on a non-Fourier thermoelastic inertial response model, and four key thermal response characteristic parameters were extracted for each detection point: temperature rise slope, response delay time, thermal hysteresis difference, and thermoelastic echo morphology. To further analyze the heat treatment uniformity of each region, these parameters were combined to construct a temperature-response hysteresis matrix, and a thermal response memory map was generated.

[0137] The thermal response memory map is fed into the improved TimeSFormer model proposed in this invention. This model introduces a thermal potential field tensor guidance mechanism, integrating spatial coordinate encoding and structural topological relationships to improve the accuracy of temporal modeling. During the inference phase, the model outputs a thermal processing uniformity score for each detection region, with values ​​ranging from 0 to 1. Higher values ​​indicate more uniform thermal processing.

[0138] Table 1. Statistical Table of Heat Treatment Uniformity Scores for Partially Tested Areas

[0139] Area code Number of detection points Average rating Lowest rating Standard deviation Structural features A03 6 0.89 0.85 0.013 upper rib A14 5 0.47 0.42 0.023 abrupt changes in wall thickness A19 4 0.52 0.44 0.031 Middle cavity corners A25 5 0.68 0.63 0.021 Flow channel outlet area A31 4 0.39 0.35 0.018 Flange root inner cavity corner

[0140] The data in Table 1 show significant differences in uniformity across different structural regions during heat treatment. Among the regions with higher scores, the A03 testing area exhibited the most stability, with an average score of 0.89 across six testing points, a minimum score of 0.85, and a standard deviation of only 0.013, indicating that its heat treatment process was generally uniform with minimal fluctuations. This region corresponds to the upper stiffening plate of the pump body, which typically has a regular geometry and clear heat conduction paths, making it easier to obtain a stable thermal response.

[0141] In contrast, regions with lower scores, such as A31, showed significantly lower heat treatment scores, with an average of 0.39 and a minimum score of 0.35, and a standard deviation of 0.018. Although the fluctuations were not the most drastic, the overall heat treatment effect was clearly insufficient. This region corresponds to the corner of the inner cavity at the flange root, where the structure is complex and the heat flow path is restricted, potentially leading to cold spot effects or delayed heat accumulation. This invention effectively reveals such weak areas in heat treatment through multi-source infrared spectrum registration and thermal response modeling, providing crucial basis for subsequent process optimization.

[0142] Region A14 also exhibits poor uniformity, with an average score of 0.47 and a standard deviation of 0.023. Corresponding to abrupt changes in wall thickness, this region shows uneven heat capacity distribution and thermal diffusion rate, easily leading to localized overheating or insufficient cooling. The delay and slope features extracted by the non-Fourier thermoelastic inertial response model can effectively identify such response anomalies. Regions A19 (middle cavity corners) and A25 (flow channel outlet region) scored 0.52 and 0.68 respectively. While lower than A03, these scores are improvements over A14 and A31, indicating that the heat treatment quality at the edges and flow junctions is at a moderate level.

[0143] By analyzing the combined indicators of the mean, extreme values ​​and standard deviation of the scores, it is possible to accurately identify local areas with defects in heat treatment, guide the adjustment of heat source parameters and structural optimization, and has good engineering application value.

[0144] To more intuitively identify problematic heat treatment areas, this invention further introduces a structure-guided clustering algorithm. Based on the region's score and its geometric coordinates in a 3D model, a spatial feature vector set of the heat treatment state is constructed. The DBSCAN clustering algorithm is used to analyze all detected regions, and the results identify three clusters with low scores and spatial proximity.

[0145] Table 2 Cluster Analysis of Heat Treatment Deviation

[0146] Cluster number Number of regions Average rating Deviation range Distribution structure characteristics Recommended adjustment measures C1 5 0.43 35.2% Lower corners of the cavity A19~A23 Improve heating uniformity and preheating time C2 4 0.41 37.1% Narrow inner channel A30~A33 Add auxiliary heat source and reduce thermal resistance C3 3 0.44 33.8% Thickness transition zone A13~A15 Reduce temperature gradient and decrease cooling rate

[0147] The average scores of the clusters in Table 2 above are significantly lower than the overall average score (0.66), indicating that these regions are at risk of uneven heat treatment, high residual thermal stress, or insufficient annealing, and require targeted structural optimization in the heat treatment process.

[0148] Cluster C1 covered five regions with an average score of 0.43 and a deviation of 35.2%. This cluster was mainly distributed in the lower corner region of the pump body cavity (A19~A23). Due to its complex geometry and structural inflection points, this region is prone to heat concentration and uneven cooling rates, leading to significant non-uniformity in heat treatment. For this region, it is recommended to improve heating uniformity and appropriately extend the preheating time to optimize the heat conduction path and heat diffusion time, thereby alleviating the low score problem.

[0149] Cluster C2 involved four detection areas, with an average score of 0.41 and the highest deviation at 37.1%, mainly distributed in the narrow channels within the cavity (A30~A33). These structures are inherently narrow in space, with a single and obstructed heat flow channel, easily forming localized thermal resistance and convection bottlenecks. This results in extremely uneven heat distribution and severe thermal hysteresis during heating and cooling, leading to a significantly lower score. Therefore, it is recommended to add auxiliary heat sources or adjust the heating path and angle during heat treatment, while simultaneously reducing the cooling rate and thermal resistance configuration to improve the consistency of heat treatment in this area.

[0150] Cluster C3 comprises three regions with an average score of 0.44 and a deviation of 33.8%, concentrated in the thickness transition region of the stiffeners (A13~A15). This type of structure exhibits significant abrupt changes in heat capacity and geometric transitions, leading to sudden hysteresis fluctuations in the thermal response. During heat treatment, these regions are prone to overheating or cold spots due to uneven stress and unstable heat conduction. It is recommended to adopt a stepped heating control strategy with a low-temperature heating rate and appropriately reduce cooling efficiency for this type of structure to minimize thermal stress accumulation and improve overall thermal stability.

[0151] This embodiment achieves high-precision modeling and intelligent analysis of the unsteady-state temperature response during pump body heat treatment by introducing multi-channel infrared imaging, multi-physics modeling, and an improved TimeSFormer model deep coding architecture. Particularly in typical regions with complex spatial structures and non-uniform heat diffusion, the model's ability to identify local heat treatment deviations is effectively enhanced by constructing a thermal response memory map and combining it with a thermal potential field tensor guidance mechanism. Simultaneously, efficient feature fusion between the detection point level and the region level is achieved using structural topology coding and region aggregation mechanisms, enhancing the model's robustness in representing non-uniform heat treatment patterns. Furthermore, density clustering algorithms are used to identify spatially similar deviation concentration areas with low scores, providing a basis for accurately proposing process optimization suggestions. This invention not only improves the automation and intelligence level of detection but also provides a generalizable technical path for evaluating the heat treatment uniformity of complex structural parts, possessing good engineering practical value and industrialization prospects.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting the uniformity of heat treatment in a pump body based on infrared imaging, characterized in that, Includes the following steps: Step 1: Obtain the three-dimensional structural model of the pump body, divide the surface of the pump body into several detection areas according to the structural features, and calibrate the detection points; Step 2: Apply thermal disturbance to each detection area by heating with laser pulses; Step 3: Use shortwave, medium wave and long wave infrared thermal imagers to dynamically acquire multi-channel temperature images of the pump body surface after thermal disturbance, and obtain the temperature-time series corresponding to each detection point; Step 4: Based on the temperature-time series, construct a non-Fourier thermoelastic inertial response model for each detection point and extract thermal response feature parameters; Step 5: Construct a temperature-response hysteresis matrix from the thermal response characteristic parameters of each detection point to form a thermal response memory map; Step 6: Input the thermal response memory map into the improved TimeSFormer model. The improved TimeSFormer model introduces a thermal potential field tensor guidance mechanism, and combines the spatial location information of the detection points with the structural topology encoding vector to obtain the thermal processing uniformity score of each detection area. Step 7: Based on the heat treatment uniformity score and pump body structure information, perform cluster analysis of the detection area to identify areas with concentrated heat treatment deviations and output a detection report.

2. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 1, characterized in that, Step one specifically involves: A three-dimensional structural model of the pump body is obtained. The three-dimensional structural model of the pump body is a digital model established by an industrial-grade three-dimensional laser scanning device, which has geometric accuracy corresponding to the actual pump body structure. The surface of the pump body is divided into regions according to structural features, including ribs, cavities, flanges, flow channels, and abrupt wall thickness boundaries, to generate several detection areas. At least one detection point is marked in each detection area. The detection point is set at the geometric center or edge corner of the detection area and corresponds to a unique spatial coordinate identifier in the three-dimensional structural model of the pump body.

3. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 1, characterized in that, Step two specifically involves: Within each detection area, laser pulse heating is performed using a near-infrared laser with a wavelength of 980 nm to 1550 nm, wherein the pulse width of the near-infrared laser is 100 ms to 500 ms and the pulse frequency is 5 Hz to 20 Hz. After being focused by a collimating lens, the laser acts on the surface of the pump body, forming thermal disturbance spots with a diameter of 1 mm to 5 mm. The near-infrared laser controls its position using a two-dimensional displacement platform and a scanning galvanometer to locate and heat designated detection points within each detection area.

4. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 1, characterized in that, Step three specifically involves: A multi-channel infrared imaging system composed of short-wave, mid-wave, and long-wave infrared thermal imagers was used to synchronously acquire temperature images of the pump body surface after thermal disturbance. The working band of the short-wave infrared thermal imager is 0.9 micrometers to 1.7 micrometers, the working band of the mid-wave infrared thermal imager is 3 micrometers to 5 micrometers, and the working band of the long-wave infrared thermal imager is 8 micrometers to 14 micrometers. The infrared thermal imager has an image acquisition frame rate of 60 frames per second to 200 frames per second, a thermal sensitivity greater than 40 milliklvin, and a spatial resolution of not less than 640×512 pixels. Each infrared channel records thermal images of the same detection area, and a multi-scale gradient orientation histogram algorithm is used for image registration. The temperature-time series of each detection point after heating excitation is extracted from the registered thermal image sequence.

5. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 4, characterized in that, The multi-scale gradient orientation histogram algorithm specifically includes: Short-wave, mid-wave, and long-wave infrared thermal images are subjected to grayscale normalization to form a uniform brightness scale; A multi-scale image set is constructed based on an image pyramid. At each scale, a sliding window of a set size is used to extract image patches, and gradient orientation histogram features are calculated for the image patches. The gradient is obtained by the Sobel operator, and the orientation histogram is discretely encoded using a preset angle interval. Long-wave infrared thermal images are selected as reference images. Cross-band image block matching is performed in short-wave infrared images and mid-wave infrared thermal images using the orientation histogram features. The position of the corresponding image block in the image to be matched is determined by minimizing the L2 norm of the pixel difference between image blocks. The image to be matched is subjected to pixel-level position correction, and B-spline interpolation is used to complete image resampling to obtain a registered image that is consistent with the reference image in spatial position, thereby generating a registered thermal image sequence.

6. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 1, characterized in that, Step four specifically involves: The temperature-time series of the detection points is preprocessed. The preprocessing includes using a sliding mean filter to reduce high-frequency noise, using a signal baseline drift correction method to unify and normalize the initial temperature reference, and aligning the starting points of each sequence according to the laser excitation time. During the excitation phase, based on the non-Fourier heat conduction theory, a two-phase hysteresis model is used to fit the temperature change process in the initial stage of excitation. The two-phase hysteresis model introduces heat flow hysteresis time and temperature gradient hysteresis time to describe the relative hysteresis behavior between heat flow and temperature response. During the peak response stage, a thermal stress-driven model caused by the temperature gradient is established based on the thermoelastic properties of the material. The thermal stress-driven model couples the instantaneous temperature gradient field with the local thermal expansion coefficient and calculates the stress distribution caused by the non-uniform thermal field through finite element discretization, fitting the main peak time and amplitude in the temperature response. During the decay phase, a recovery model expressing the response hysteresis is constructed using a time deconvolution kernel function to clearly distinguish the nonlinear process of thermal diffusivity changing with time. The biphase hysteresis model, thermal stress-driven model, and recovery model are continuously spliced ​​together to form a complete non-Fourier thermoelastic inertial response model. Thermal response feature parameters are extracted based on the non-Fourier thermoelastic inertial response model. These thermal response feature parameters include: Temperature change slope: the average rate of change of the temperature rise segment after the excitation start time; Response delay time: The time interval from the start of excitation to the occurrence of the maximum slope change in the temperature curve; Thermal hysteresis characteristic: the delay difference between peak response time and excitation termination time; Thermoelastic echo morphology: the amplitude and symmetrical distribution characteristics of the rebound temperature fluctuations that occur during the decay phase.

7. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 1, characterized in that, Step five specifically involves: The thermal response feature parameters extracted from each detection point are arranged according to a preset time node to construct a corresponding temperature-response hysteresis matrix. The rows of the temperature-response hysteresis matrix represent the spatial location of different detection points, the columns represent the time evolution dimension after thermal disturbance, and the matrix elements are the values ​​of thermal response feature parameters at the corresponding time. The constructed temperature-response hysteresis matrix is ​​encoded and mapped in two dimensions to generate a thermal response memory map.

8. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 1, characterized in that, Step six specifically involves: The thermal response memory map is input into the improved TimeSFormer model, which adopts a Transformer architecture with a time-space separation structure, including a spatial encoding module, a temporal encoding module, and a region aggregation module. The spatial coding module introduces a thermal potential field tensor guidance mechanism, which specifically includes: A three-dimensional thermal potential field tensor is constructed, wherein the three components of the three-dimensional thermal potential field tensor are the derivative of the rate of temperature change with respect to time, the normalized numerical distribution of the local heat capacity of the material, and the thermal hysteresis distance calculated based on the thermal response delay between detection points. The three-dimensional thermal potential field tensor is transformed into a guiding vector field by a convolutional encoder during the input stage, and then weighted and fused with the linear transformation results of Query and Key in the spatial encoding module to construct the guiding perception attention weight in the spatial attention mechanism. Based on the guided perception attention weights, spatial features are extracted from each frame of the thermal response memory map to generate spatial coding features; The temporal encoding module is constructed in the form of a multi-layer Transformer encoder stack. It models the spatial encoding features in the temporal dimension and introduces the structural topology encoding vector of the detection point at each time step. The structural topology encoding vector is obtained by mapping the spatial coordinates of the detection point in the three-dimensional structural model of the pump body through an MLP embedding network. The MLP embedding network includes two hidden layers with ReLU activation functions and an output layer, which are used to encode the three-dimensional spatial coordinates into an embedding representation of a preset dimension and fuse them with the spatial encoding features of the corresponding time step to obtain the temporal encoding features. The region aggregation module aggregates the time-encoding features of multiple detection points in each detection region and generates a region-level thermal response representation vector by weighted averaging. The region-level thermal response representation vector is input into a fully connected layer, which uses a Sigmoid activation function to output a thermal uniformity score between 0 and 1.

9. The method for detecting the uniformity of heat treatment of a pump body based on infrared imaging according to claim 1, characterized in that, Step seven specifically involves: The heat treatment uniformity score of each detection area is combined with the geometric center coordinates in the three-dimensional structural model of the pump body to construct a feature vector set containing heat treatment uniformity score and spatial location information. A density-based spatial clustering algorithm was used to analyze the feature vector set. The density threshold condition was set to ensure that the difference in heat treatment uniformity score did not exceed 0.05 and the spatial distance did not exceed 10 mm. The detection regions that meet the density threshold condition are divided into the same cluster; In all clusters, the average heat treatment uniformity score is statistically analyzed, and clusters whose average heat treatment uniformity score is more than 30% lower than the average heat treatment uniformity score of the entire detection area are identified as concentrated areas of heat treatment deviation. Output a test report, which includes the spatial distribution of each cluster, statistical results of heat treatment uniformity scores, and location information of concentrated areas of heat treatment deviation.

Images